Community Scripts

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:47:56
# Model            : Gradient Boosting
# Feature Eng.     : SMA (20,50,200) + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/GBPUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── SMA features (required) ──────────────────────────────────────────────
    for period in [20, 50, 200]:
        sma = close.rolling(period).mean()
        df[f"sma_{period}"] = sma
        df[f"dm_sma_{period}"] = (close - sma) / sma

    # ── SMA crossover signals ────────────────────────────────────────────────
    sma_20 = close.rolling(20).mean()
    sma_50 = close.rolling(50).mean()
    sma_200 = close.rolling(200).mean()

    df["sma_20_50_cross"] = np.where(sma_20 > sma_50, 1.0, -1.0)
    df["sma_20_200_cross"] = np.where(sma_20 > sma_200, 1.0, -1.0)
    df["sma_50_200_cross"] = np.where(sma_50 > sma_200, 1.0, -1.0)

    # ── Price momentum features ──────────────────────────────────────────────
    for lag in [1, 2, 4, 8, 16]:
        df[f"ret_{lag}"] = close.pct_change(lag)

    # ── Volatility: rolling std of returns ──────────────────────────────────
    ret_1 = close.pct_change(1)
    for window in [8, 20, 50]:
        df[f"vol_{window}"] = ret_1.rolling(window).std()

    # ── ATR (Average True Range) ─────────────────────────────────────────────
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)

    for atr_period in [14, 50]:
        atr = tr.rolling(atr_period).mean()
        df[f"atr_{atr_period}"] = atr
        df[f"natr_{atr_period}"] = atr / close

    # ── RSI ──────────────────────────────────────────────────────────────────
    for rsi_period in [14, 28]:
        delta = close.diff()
        gain  = delta.clip(lower=0).rolling(rsi_period).mean()
        loss  = (-delta.clip(upper=0)).rolling(rsi_period).mean()
        rs    = gain / (loss + 1e-10)
        df[f"rsi_{rsi_period}"] = 100 - (100 / (1 + rs))

    # ── MACD ─────────────────────────────────────────────────────────────────
    ema_12 = close.ewm(span=12, adjust=False).mean()
    ema_26 = close.ewm(span=26, adjust=False).mean()
    macd_line   = ema_12 - ema_26
    signal_line = macd_line.ewm(span=9, adjust=False).mean()
    df["macd"]           = macd_line
    df["macd_signal"]    = signal_line
    df["macd_hist"]      = macd_line - signal_line
    df["macd_hist_norm"] = (macd_line - signal_line) / (close + 1e-10)

    # ── Bollinger Bands ───────────────────────────────────────────────────────
    for bb_period in [20, 50]:
        bb_mid  = close.rolling(bb_period).mean()
        bb_std  = close.rolling(bb_period).std()
        bb_upper = bb_mid + 2.0 * bb_std
        bb_lower = bb_mid - 2.0 * bb_std
        bb_width = (bb_upper - bb_lower) / (bb_mid + 1e-10)
        bb_pos   = (close - bb_lower) / (bb_upper - bb_lower + 1e-10)
        df[f"bb_width_{bb_period}"] = bb_width
        df[f"bb_pos_{bb_period}"]   = bb_pos

    # ── Stochastic Oscillator ────────────────────────────────────────────────
    for stoch_period in [14, 28]:
        lowest_low   = low.rolling(stoch_period).min()
        highest_high = high.rolling(stoch_period).max()
        stoch_k = (close - lowest_low) / (highest_high - lowest_low + 1e-10) * 100
        stoch_d = stoch_k.rolling(3).mean()
        df[f"stoch_k_{stoch_period}"] = stoch_k
        df[f"stoch_d_{stoch_period}"] = stoch_d

    # ── Rate of Change (ROC) ──────────────────────────────────────────────────
    for roc_period in [5, 10, 20]:
        df[f"roc_{roc_period}"] = close.pct_change(roc_period)

    # ── Candle body and shadow features ──────────────────────────────────────
    body        = (close - open_).abs()
    candle_range = (high - low).abs()
    df["body_ratio"]       = body / (candle_range + 1e-10)
    df["upper_shadow"]     = (high - pd.concat([close, open_], axis=1).max(axis=1)) / (candle_range + 1e-10)
    df["lower_shadow"]     = (pd.concat([close, open_], axis=1).min(axis=1) - low) / (candle_range + 1e-10)
    df["bullish_candle"]   = np.where(close > open_, 1.0, -1.0)

    # ── Volume-proxy: candle range as volatility proxy ────────────────────────
    df["range_norm"] = candle_range / (close + 1e-10)
    df["range_ma_ratio"] = candle_range / (candle_range.rolling(20).mean() + 1e-10)

    # ── Lag features for return predictors ───────────────────────────────────
    for col_lag in ["rsi_14", "macd_hist", "bb_pos_20"]:
        for lag in [1, 2, 3]:
            df[f"{col_lag}_lag{lag}"] = df[col_lag].shift(lag)

    # ── Distance of close from recent high/low ────────────────────────────────
    for lookback in [10, 20, 50]:
        roll_high = high.rolling(lookback).max()
        roll_low  = low.rolling(lookback).min()
        df[f"dist_high_{lookback}"] = (close - roll_high) / (roll_high + 1e-10)
        df[f"dist_low_{lookback}"]  = (close - roll_low)  / (roll_low  + 1e-10)

    # ── Trend strength: ADX proxy ─────────────────────────────────────────────
    adx_period = 14
    tr_adx  = tr.copy()
    plus_dm  = pd.Series(np.where((high.diff() > 0) & (high.diff() > -low.diff()), high.diff(), 0.0), index=close.index)
    minus_dm = pd.Series(np.where((-low.diff() > 0) & (-low.diff() > high.diff()), -low.diff(), 0.0), index=close.index)
    atr_adx  = tr_adx.rolling(adx_period).mean()
    plus_di  = 100 * plus_dm.rolling(adx_period).mean()  / (atr_adx + 1e-10)
    minus_di = 100 * minus_dm.rolling(adx_period).mean() / (atr_adx + 1e-10)
    dx       = (100 * (plus_di - minus_di).abs() / (plus_di + minus_di + 1e-10))
    df["adx"]       = dx.rolling(adx_period).mean()
    df["plus_di"]   = plus_di
    df["minus_di"]  = minus_di

    # ── Fill NaN from indicator warm-up ──────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "GBP/USD SMA Trend Gradient Boosting Risk-Adj",
        "model_type":       "GradientBoostingClassifier",
        "model_params":     {
            "n_estimators":     400,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.75,
            "min_samples_leaf": 20,
            "max_features":     "sqrt",
            "n_iter_no_change": 30,
            "validation_fraction": 0.1,
            "tol":              1e-4,
            "random_state":     42,
        },
        "signal_threshold": 0.57,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 18],
        "min_atr":          0.0002,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe/Calmar) on GBP/USD 15-min data. "
            "GradientBoostingClassifier chosen for its strong bias-variance tradeoff "
            "on medium-sized tabular datasets without needing GPU. "
            "Hyperparameters: moderate depth=4 prevents overfitting, learning_rate=0.04 "
            "with 400 estimators balances convergence vs generalisation, subsample=0.75 "
            "adds stochasticity to reduce variance, min_samples_leaf=20 enforces statistical "
            "significance at each leaf. Early stopping via n_iter_no_change guards against "
            "overfit on the training fold. Signal threshold 0.57 filters marginal signals "
            "to improve precision. SL=0.5%, TP=1.0% gives 1:2 RR. Session filter 6-18 UTC "
            "covers London+NY overlap — highest GBP/USD liquidity and tighter spreads. "
            "sma_50 trend filter ensures we only trade in the direction of medium-term trend, "
            "reducing whipsaw losses. target_horizon=4 bars (1 hour) gives the model enough "
            "time for moves to develop while staying relevant for intraday trading."
        ),
        "notes": (
            "Features: SMA 20/50/200 with distance metrics (core requirement), RSI 14/28, "
            "MACD, Bollinger Bands 20/50, Stochastic 14/28, ATR 14/50, NATR, ROC, ADX, "
            "candle body/shadow ratios, lagged RSI/MACD/BB features, distance from rolling "
            "high/low, SMA crossover signals, multi-lag return features. "
            "All features are backward-looking only (no lookahead bias). "
            "on_opposite=reverse for fast trend-following entries without missing reversals."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:37:57
# Model            : Gradient Boosting
# Feature Eng.     : EMA (9,21), RSI 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # --- EMA 9 and EMA 21 (required) ---
    ema_9 = close.ewm(span=9, adjust=False).mean()
    ema_21 = close.ewm(span=21, adjust=False).mean()
    df["ema_9"] = ema_9
    df["ema_21"] = ema_21
    df["dm_ema_9"] = (close - ema_9) / ema_9
    df["dm_ema_21"] = (close - ema_21) / ema_21

    # EMA crossover signal: positive when fast > slow
    df["ema_cross"] = ema_9 - ema_21
    # Crossover direction change (sign flip)
    df["ema_cross_signal"] = np.sign(df["ema_cross"])
    df["ema_cross_prev"] = df["ema_cross_signal"].shift(1)
    df["ema_cross_flip"] = (df["ema_cross_signal"] != df["ema_cross_prev"]).astype(float)

    # --- RSI 14 (required) ---
    delta = close.diff()
    gain = delta.clip(lower=0)
    loss = -delta.clip(upper=0)
    avg_gain = gain.ewm(com=13, adjust=False).mean()
    avg_loss = loss.ewm(com=13, adjust=False).mean()
    rs = avg_gain / (avg_loss + 1e-10)
    rsi_14 = 100 - (100 / (1 + rs))
    df["rsi_14"] = rsi_14
    # RSI normalised to [-1, 1] range
    df["rsi_norm"] = (rsi_14 - 50) / 50
    # RSI overbought/oversold flags
    df["rsi_ob"] = np.where(rsi_14 > 70, 1.0, 0.0)
    df["rsi_os"] = np.where(rsi_14 < 30, 1.0, 0.0)

    # --- Additional momentum and volatility features ---

    # EMA 50 for trend context
    ema_50 = close.ewm(span=50, adjust=False).mean()
    df["ema_50"] = ema_50
    df["dm_ema_50"] = (close - ema_50) / ema_50

    # Price momentum: rate of change over multiple horizons
    df["roc_4"] = close.pct_change(4)
    df["roc_8"] = close.pct_change(8)
    df["roc_16"] = close.pct_change(16)

    # ATR (Average True Range) for volatility
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    atr_14 = tr.ewm(span=14, adjust=False).mean()
    df["atr_14"] = atr_14
    # Normalised ATR
    df["natr_14"] = atr_14 / close

    # Bollinger Bands (20-period, 2 std)
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std
    df["bb_pct"] = (close - bb_lower) / (bb_upper - bb_lower + 1e-10)
    df["bb_width"] = (bb_upper - bb_lower) / (bb_mid + 1e-10)
    # BB squeeze: narrow bands signal potential breakout
    df["bb_squeeze"] = np.where(df["bb_width"] < df["bb_width"].rolling(50).mean(), 1.0, 0.0)

    # MACD-like: difference between two EMAs
    ema_12 = close.ewm(span=12, adjust=False).mean()
    ema_26 = close.ewm(span=26, adjust=False).mean()
    macd_line = ema_12 - ema_26
    macd_signal = macd_line.ewm(span=9, adjust=False).mean()
    df["macd_line"] = macd_line / (close + 1e-10)
    df["macd_signal"] = macd_signal / (close + 1e-10)
    df["macd_hist"] = (macd_line - macd_signal) / (close + 1e-10)
    df["macd_cross"] = np.sign(macd_line - macd_signal)

    # Stochastic oscillator (14-period)
    lowest_low = low.rolling(14).min()
    highest_high = high.rolling(14).max()
    stoch_k = 100 * (close - lowest_low) / (highest_high - lowest_low + 1e-10)
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"] = stoch_k
    df["stoch_d"] = stoch_d
    df["stoch_diff"] = stoch_k - stoch_d

    # Volume of price movement (candle body and shadows)
    df["body"] = (close - open_).abs() / (atr_14 + 1e-10)
    df["upper_shadow"] = (high - pd.concat([close, open_], axis=1).max(axis=1)) / (atr_14 + 1e-10)
    df["lower_shadow"] = (pd.concat([close, open_], axis=1).min(axis=1) - low) / (atr_14 + 1e-10)
    df["candle_dir"] = np.sign(close - open_)

    # Rolling volatility (realised vol over 20 bars)
    log_ret = np.log(close / close.shift(1))
    df["realvol_20"] = log_ret.rolling(20).std()

    # Close relative to recent high/low channel (20-bar)
    roll_high_20 = high.rolling(20).max()
    roll_low_20 = low.rolling(20).min()
    df["chan_pct_20"] = (close - roll_low_20) / (roll_high_20 - roll_low_20 + 1e-10)

    # Lagged RSI and EMA cross for temporal context
    df["rsi_14_lag1"] = rsi_14.shift(1)
    df["rsi_14_lag2"] = rsi_14.shift(2)
    df["ema_cross_lag1"] = df["ema_cross"].shift(1)
    df["ema_cross_lag2"] = df["ema_cross"].shift(2)
    df["macd_hist_lag1"] = df["macd_hist"].shift(1)

    # RSI momentum: change in RSI
    df["rsi_delta_1"] = rsi_14.diff(1)
    df["rsi_delta_4"] = rsi_14.diff(4)

    # EMA9 slope (normalised)
    df["ema9_slope"] = ema_9.diff(3) / (ema_9.shift(3) + 1e-10)
    df["ema21_slope"] = ema_21.diff(3) / (ema_21.shift(3) + 1e-10)

    # Interaction: RSI * EMA cross direction
    df["rsi_ema_cross_interact"] = df["rsi_norm"] * df["ema_cross_signal"]

    # Fill NaN from indicator warm-up
    df = df.bfill().ffill()

    return df

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title": "EMA Cross 9/21 + RSI14 Gradient Boost Scalper",
        "model_type": "GradientBoostingClassifier",
        "model_params": {
            "n_estimators": 400,
            "max_depth": 4,
            "learning_rate": 0.04,
            "subsample": 0.8,
            "min_samples_leaf": 20,
            "min_samples_split": 40,
            "max_features": "sqrt",
            "n_iter_no_change": 30,
            "validation_fraction": 0.1,
            "tol": 1e-4,
        },
        "signal_threshold": 0.56,
        "direction": "both",
        "stop_loss": 0.005,
        "take_profit": 0.010,
        "cooldown": 0,
        "max_positions": 1,
        "on_opposite": "reverse",
        "session_filter": [7, 17],
        "min_atr": 0.0002,
        "trend_filter": "sma_50",
        "target_horizon": 4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe/Calmar) using a GradientBoostingClassifier "
            "with EMA 9/21 crossover as the primary signal source and RSI 14 as confirmation. "
            "Deep feature set includes MACD, Bollinger Bands, Stochastic, ATR normalisation, "
            "candle structure, lagged features and RSI/EMA interaction terms. "
            "Gradient boosting chosen for its ability to capture non-linear interactions between "
            "trend, momentum and volatility features without overfitting when regularised via "
            "subsample, max_features, and early stopping. Threshold 0.56 filters marginal signals. "
            "Session filter [7,17] focuses on London/NY overlap for highest EUR/USD liquidity. "
            "SL 0.5% / TP 1.0% gives 1:2 risk-reward aligned with scalper momentum targets. "
            "Reverse on opposite signal to stay in sync with fast EMA crossover momentum."
        ),
        "notes": (
            "EMA 9/21 cross captures short-term momentum shifts typical of active EUR/USD sessions. "
            "RSI 14 filters entries in extreme overbought/oversold conditions. "
            "NATR min_atr filter removes flat/low-vol periods. "
            "Trend filter (SMA 50) ensures longs only above and shorts only below the medium-term trend. "
            "n_iter_no_change=30 provides early stopping to prevent overfitting on the training split. "
            "400 estimators with depth 4 and lr 0.04 balance bias-variance tradeoff for intraday data."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:01:39
# Model            : Gradient Boosting
# Feature Eng.     : BB (20,2.0), ATR 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/USDJPY_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── Bollinger Bands (period=20, std_dev=2.0) ──────────────────────────────
    bb_period = 20
    bb_std    = 2.0

    bb_mid   = close.rolling(bb_period).mean()
    bb_sigma = close.rolling(bb_period).std(ddof=0)
    bb_upper = bb_mid + bb_std * bb_sigma
    bb_lower = bb_mid - bb_std * bb_sigma

    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower

    # bb_width: normalised band width (volatility proxy)
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid

    # bb_pct: position of close within the band [0, 1]
    band_range = bb_upper - bb_lower
    df["bb_pct"] = (close - bb_lower) / band_range

    # Distance from close to mid in units of band width
    df["bb_dist_mid"] = (close - bb_mid) / bb_mid

    # ── ATR (period=14) ───────────────────────────────────────────────────────
    atr_period = 14
    prev_close = close.shift(1)
    tr = pd.concat([
        high - low,
        (high - prev_close).abs(),
        (low  - prev_close).abs()
    ], axis=1).max(axis=1)

    atr  = tr.ewm(span=atr_period, min_periods=atr_period, adjust=False).mean()
    natr = atr / close

    df["atr"]  = atr
    df["natr"] = natr

    # ── Momentum / trend features ─────────────────────────────────────────────
    # Rate of change at multiple horizons
    for n in [1, 4, 8, 16]:
        df[f"roc_{n}"] = close.pct_change(n)

    # RSI (14)
    rsi_period = 14
    delta  = close.diff()
    gain   = delta.clip(lower=0)
    loss   = (-delta).clip(lower=0)
    avg_gain = gain.ewm(span=rsi_period, min_periods=rsi_period, adjust=False).mean()
    avg_loss = loss.ewm(span=rsi_period, min_periods=rsi_period, adjust=False).mean()
    rs  = avg_gain / (avg_loss + 1e-10)
    rsi = 100 - (100 / (1 + rs))
    df["rsi_14"] = rsi

    # RSI derived: distance from 50 (centred, normalised)
    df["rsi_dev"] = (rsi - 50) / 50

    # ── MACD (12, 26, 9) ──────────────────────────────────────────────────────
    ema12 = close.ewm(span=12, adjust=False).mean()
    ema26 = close.ewm(span=26, adjust=False).mean()
    macd_line   = ema12 - ema26
    macd_signal = macd_line.ewm(span=9, adjust=False).mean()
    macd_hist   = macd_line - macd_signal

    df["macd_line"]   = macd_line
    df["macd_signal"] = macd_signal
    df["macd_hist"]   = macd_hist

    # ── Trend (SMA 50) ────────────────────────────────────────────────────────
    sma50 = close.rolling(50).mean()
    df["sma_50"]        = sma50
    df["close_vs_sma50"] = (close - sma50) / sma50   # normalised distance

    # ── Volume / candle structure features ────────────────────────────────────
    body      = (close - open_).abs()
    candle_rng = high - low
    df["body_ratio"]  = body / (candle_rng + 1e-10)   # body as fraction of range
    df["upper_wick"]  = (high  - pd.concat([close, open_], axis=1).max(axis=1)) / (candle_rng + 1e-10)
    df["lower_wick"]  = (pd.concat([close, open_], axis=1).min(axis=1) - low)   / (candle_rng + 1e-10)
    df["candle_dir"]  = np.where(close > open_, 1.0, -1.0)  # bullish / bearish bar

    # ── Lagged bb_pct & rsi (to give the model recent history) ───────────────
    for lag in [1, 2, 3]:
        df[f"bb_pct_lag{lag}"] = df["bb_pct"].shift(lag)
        df[f"rsi_lag{lag}"]    = df["rsi_14"].shift(lag)
        df[f"macd_hist_lag{lag}"] = df["macd_hist"].shift(lag)

    # ── Volatility regime flag ────────────────────────────────────────────────
    natr_ma = natr.rolling(50).mean()
    df["vol_regime"] = np.where(natr > natr_ma, 1.0, 0.0)   # 1 = high-vol regime

    # ── BB squeeze detection ──────────────────────────────────────────────────
    bb_width_ma = df["bb_width"].rolling(50).mean()
    df["bb_squeeze"] = np.where(df["bb_width"] < bb_width_ma, 1.0, 0.0)

    # ── Mean-reversion signal strength ────────────────────────────────────────
    # Positive → oversold (close below lower band), Negative → overbought
    df["mr_signal"] = 0.5 - df["bb_pct"]   # centred: +0.5 at lower band, -0.5 at upper

    # ── Fill NaN from warm-up ─────────────────────────────────────────────────
    df = df.bfill().ffill()

    return df

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "USD/JPY BB Mean-Reversion + ATR Gradient Boost",
        "model_type":       "GradientBoostingClassifier",
        "model_params":     {
            "n_estimators":      500,
            "max_depth":         4,
            "learning_rate":     0.03,
            "subsample":         0.8,
            "min_samples_leaf":  20,
            "max_features":      "sqrt",
            "validation_fraction": 0.1,
            "n_iter_no_change":  30,
            "tol":               1e-4,
            "random_state":      42,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   None,
        "min_atr":          None,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximise Sharpe ratio via a Gradient Boosting classifier trained on "
            "Bollinger Band position (bb_pct), normalised bandwidth (bb_width), "
            "ATR/NATR, RSI, MACD histogram, candle structure, and lagged features. "
            "GBM chosen for its ability to capture non-linear interactions between "
            "volatility (ATR) and mean-reversion (BB) signals. n_iter_no_change "
            "acts as early stopping to prevent overfitting on the 15-min USDJPY series. "
            "SL=0.5% / TP=1.0% gives a 1:2 risk-reward; threshold=0.55 reduces noise trades."
        ),
        "notes": (
            "Bollinger Bands are the primary mean-reversion anchor; ATR/NATR filter "
            "entries to adequate volatility bars. RSI and MACD provide momentum context "
            "to avoid fading strong trends. Lagged features (up to 3 bars) give the model "
            "short-term regime memory without look-ahead. vol_regime and bb_squeeze flags "
            "allow the model to differentiate trending vs. ranging conditions automatically. "
            "No session filter applied — USDJPY is liquid across Asian and European sessions."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:24:20
# Model            : Gradient Boosting
# Feature Eng.     : BB (20,2.0), RSI 14, Stochastic (14,3) + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/AUDUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_period = 20
    bb_std    = 2.0
    bb_mid    = close.rolling(bb_period).mean()
    bb_std_   = close.rolling(bb_period).std(ddof=1)
    bb_upper  = bb_mid + bb_std * bb_std_
    bb_lower  = bb_mid - bb_std * bb_std_
    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid
    df["bb_pct"]   = (close - bb_lower) / (bb_upper - bb_lower)

    # ── RSI (14) ─────────────────────────────────────────────────────────────
    rsi_period = 14
    delta      = close.diff()
    gain       = delta.clip(lower=0)
    loss       = -delta.clip(upper=0)
    avg_gain   = gain.ewm(com=rsi_period - 1, min_periods=rsi_period).mean()
    avg_loss   = loss.ewm(com=rsi_period - 1, min_periods=rsi_period).mean()
    rs         = avg_gain / avg_loss.replace(0, np.nan)
    df["rsi"]  = 100.0 - (100.0 / (1.0 + rs))

    # ── Stochastic Oscillator (K=14, D=3) ────────────────────────────────────
    stoch_k = 14
    stoch_d = 3
    low_min  = low.rolling(stoch_k).min()
    high_max = high.rolling(stoch_k).max()
    k_raw    = 100.0 * (close - low_min) / (high_max - low_min).replace(0, np.nan)
    df["stoch_k"] = k_raw
    df["stoch_d"] = k_raw.rolling(stoch_d).mean()

    # ── ATR (14) ─────────────────────────────────────────────────────────────
    atr_period = 14
    prev_close = close.shift(1)
    tr         = pd.concat([
        high - low,
        (high - prev_close).abs(),
        (low  - prev_close).abs()
    ], axis=1).max(axis=1)
    df["atr"]  = tr.ewm(com=atr_period - 1, min_periods=atr_period).mean()
    df["natr"] = df["atr"] / close

    # ── Trend / Momentum features ─────────────────────────────────────────────
    df["sma_20"]  = close.rolling(20).mean()
    df["sma_50"]  = close.rolling(50).mean()
    df["sma_100"] = close.rolling(100).mean()

    df["price_vs_sma20"]  = (close - df["sma_20"])  / df["sma_20"]
    df["price_vs_sma50"]  = (close - df["sma_50"])  / df["sma_50"]
    df["price_vs_sma100"] = (close - df["sma_100"]) / df["sma_100"]
    df["sma20_vs_sma50"]  = (df["sma_20"] - df["sma_50"]) / df["sma_50"]

    # ── MACD (12, 26, 9) ─────────────────────────────────────────────────────
    ema12        = close.ewm(span=12, min_periods=12).mean()
    ema26        = close.ewm(span=26, min_periods=26).mean()
    macd_line    = ema12 - ema26
    macd_signal  = macd_line.ewm(span=9, min_periods=9).mean()
    df["macd"]         = macd_line
    df["macd_signal"]  = macd_signal
    df["macd_hist"]    = macd_line - macd_signal

    # ── Rate-of-Change features ───────────────────────────────────────────────
    for p in [4, 8, 16]:
        df[f"roc_{p}"] = close.pct_change(p)

    # ── Volatility regime ────────────────────────────────────────────────────
    df["vol_8"]  = close.pct_change().rolling(8).std()
    df["vol_20"] = close.pct_change().rolling(20).std()
    df["vol_ratio"] = df["vol_8"] / df["vol_20"].replace(0, np.nan)

    # ── Candle body / shadow features ────────────────────────────────────────
    df["body"]        = (close - open_).abs()
    df["upper_shadow"] = high - pd.concat([close, open_], axis=1).max(axis=1)
    df["lower_shadow"] = pd.concat([close, open_], axis=1).min(axis=1) - low
    df["body_ratio"]  = df["body"] / (high - low).replace(0, np.nan)

    # ── RSI-derived features ──────────────────────────────────────────────────
    df["rsi_above_50"] = np.where(df["rsi"] > 50, 1, 0)
    df["rsi_overbought"] = np.where(df["rsi"] > 70, 1, 0)
    df["rsi_oversold"]   = np.where(df["rsi"] < 30, 1, 0)
    df["rsi_lag1"] = df["rsi"].shift(1)
    df["rsi_lag4"] = df["rsi"].shift(4)

    # ── Stochastic-derived features ───────────────────────────────────────────
    df["stoch_cross_up"]   = np.where((df["stoch_k"] > df["stoch_d"]) &
                                      (df["stoch_k"].shift(1) <= df["stoch_d"].shift(1)), 1, 0)
    df["stoch_cross_down"] = np.where((df["stoch_k"] < df["stoch_d"]) &
                                      (df["stoch_k"].shift(1) >= df["stoch_d"].shift(1)), 1, 0)
    df["stoch_oversold"]   = np.where(df["stoch_k"] < 20, 1, 0)
    df["stoch_overbought"] = np.where(df["stoch_k"] > 80, 1, 0)

    # ── BB-derived features ───────────────────────────────────────────────────
    df["bb_squeeze"] = np.where(df["bb_width"] < df["bb_width"].rolling(50).quantile(0.20), 1, 0)
    df["above_bb_upper"] = np.where(close > bb_upper, 1, 0)
    df["below_bb_lower"] = np.where(close < bb_lower, 1, 0)
    df["bb_pct_lag1"]    = df["bb_pct"].shift(1)
    df["bb_pct_lag4"]    = df["bb_pct"].shift(4)

    # ── Session hour (UTC) ────────────────────────────────────────────────────
    df["hour_utc"] = df.index.hour if hasattr(df.index, "hour") else 0

    # ── Fill NaN from indicator warm-up ──────────────────────────────────────
    df = df.bfill().ffill()
    return df

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "AUD/USD Stochastic BB Mean-Reversion (GBM)",
        "model_type":       "GradientBoostingClassifier",
        "model_params":     {
            "n_estimators":      400,
            "max_depth":         4,
            "learning_rate":     0.04,
            "subsample":         0.80,
            "min_samples_leaf":  20,
            "max_features":      "sqrt",
            "n_iter_no_change":  30,
            "validation_fraction": 0.10,
            "tol":               1e-4,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   None,
        "min_atr":          None,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe/Calmar) on AUD/USD 15-min. "
            "GradientBoostingClassifier with moderate depth and learning rate chosen "
            "to balance bias-variance. 2:1 reward-to-risk (SL=0.5%, TP=1.0%). "
            "Stochastic crossovers, BB mean-reversion, and RSI regime signals "
            "form the core feature set; MACD, volatility, and candle features add "
            "context. Early stopping (n_iter_no_change=30) prevents overfitting."
        ),
        "notes": (
            "Features: Bollinger Bands (20,2) width/pct, RSI(14) with lag/regime flags, "
            "Stochastic(14,3) K/D with crossover detection, ATR/NATR volatility, MACD "
            "histogram, short/medium SMAs, ROC(4/8/16), volatility ratio, candle body "
            "ratios, and UTC session hour. No session or trend filter to allow full "
            "mean-reversion opportunities across all sessions."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:35:33
# Model            : Gradient Boosting
# Feature Eng.     : RSI 14, MACD (12,26,9) + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/NZDUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):

    # ── RSI 14 ──────────────────────────────────────────────────────────────
    delta = close.diff()
    gain  = delta.clip(lower=0)
    loss  = (-delta).clip(lower=0)
    avg_gain = gain.ewm(alpha=1/14, min_periods=14, adjust=False).mean()
    avg_loss = loss.ewm(alpha=1/14, min_periods=14, adjust=False).mean()
    rs = avg_gain / avg_loss.replace(0, np.nan)
    df["rsi_14"] = 100 - (100 / (1 + rs))

    # RSI derived signals
    df["rsi_ob"]      = np.where(df["rsi_14"] > 70, 1, 0)   # overbought flag
    df["rsi_os"]      = np.where(df["rsi_14"] < 30, 1, 0)   # oversold flag
    df["rsi_mid"]     = df["rsi_14"] - 50                    # centred
    df["rsi_slope"]   = df["rsi_14"].diff(3)                 # momentum of RSI
    df["rsi_accel"]   = df["rsi_slope"].diff(2)              # acceleration

    # RSI regime: above/below 50
    df["rsi_bull"]    = np.where(df["rsi_14"] > 50, 1, -1)

    # ── MACD (12, 26, 9) ────────────────────────────────────────────────────
    ema12 = close.ewm(span=12, adjust=False).mean()
    ema26 = close.ewm(span=26, adjust=False).mean()
    macd_line   = ema12 - ema26
    signal_line = macd_line.ewm(span=9, adjust=False).mean()
    macd_hist   = macd_line - signal_line

    df["macd_line"]   = macd_line
    df["macd_signal"] = signal_line
    df["macd_hist"]   = macd_hist

    # MACD derived
    df["macd_cross"]     = np.where(macd_line > signal_line, 1, -1)
    df["macd_hist_sign"] = np.where(macd_hist > 0, 1, -1)
    df["macd_hist_chg"]  = macd_hist.diff(1)          # histogram change
    df["macd_hist_accel"]= df["macd_hist_chg"].diff(1) # second derivative
    df["macd_zero_cross"]= np.where(macd_line > 0, 1, -1)

    # ── ATR 14 ──────────────────────────────────────────────────────────────
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    atr14 = tr.ewm(alpha=1/14, min_periods=14, adjust=False).mean()
    df["atr_14"]   = atr14
    df["natr_14"]  = atr14 / close          # normalised ATR
    df["atr_ratio"]= atr14 / atr14.rolling(50).mean()  # current vs recent vol

    # ── Volatility regime ───────────────────────────────────────────────────
    df["vol_high"] = np.where(df["natr_14"] > df["natr_14"].rolling(100).median(), 1, 0)

    # ── Price momentum ──────────────────────────────────────────────────────
    df["ret_1"]  = close.pct_change(1)
    df["ret_3"]  = close.pct_change(3)
    df["ret_8"]  = close.pct_change(8)
    df["ret_16"] = close.pct_change(16)

    # Scaled by ATR so the model sees normalised moves
    df["ret_1_atr"]  = df["ret_1"]  / (atr14 / close).replace(0, np.nan)
    df["ret_3_atr"]  = df["ret_3"]  / (atr14 / close).replace(0, np.nan)
    df["ret_8_atr"]  = df["ret_8"]  / (atr14 / close).replace(0, np.nan)

    # ── EMAs & trend structure ───────────────────────────────────────────────
    ema8  = close.ewm(span=8,  adjust=False).mean()
    ema21 = close.ewm(span=21, adjust=False).mean()
    ema50 = close.ewm(span=50, adjust=False).mean()
    ema100= close.ewm(span=100,adjust=False).mean()

    df["ema8_21_spread"] = (ema8 - ema21) / close
    df["ema21_50_spread"]= (ema21 - ema50) / close
    df["ema50_100_spread"]= (ema50 - ema100) / close

    df["price_vs_ema21"] = (close - ema21) / close
    df["price_vs_ema50"] = (close - ema50) / close

    df["trend_align"] = np.where(
        (ema8 > ema21) & (ema21 > ema50), 1,
        np.where((ema8 < ema21) & (ema21 < ema50), -1, 0)
    )

    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_mid  = close.rolling(20).mean()
    bb_std  = close.rolling(20).std(ddof=0)
    bb_up   = bb_mid + 2 * bb_std
    bb_lo   = bb_mid - 2 * bb_std
    bb_bw   = (bb_up - bb_lo) / bb_mid          # bandwidth
    bb_pct  = (close - bb_lo) / (bb_up - bb_lo) # %B

    df["bb_pct"]      = bb_pct
    df["bb_bw"]       = bb_bw
    df["bb_bw_ratio"] = bb_bw / bb_bw.rolling(50).mean()  # squeeze detector
    df["bb_upper_touch"] = np.where(close >= bb_up, 1, 0)
    df["bb_lower_touch"] = np.where(close <= bb_lo, 1, 0)

    # ── Stochastic %K %D (14, 3) ────────────────────────────────────────────
    lo14   = low.rolling(14).min()
    hi14   = high.rolling(14).max()
    stoch_k = 100 * (close - lo14) / (hi14 - lo14).replace(0, np.nan)
    stoch_d = stoch_k.rolling(3).mean()

    df["stoch_k"]      = stoch_k
    df["stoch_d"]      = stoch_d
    df["stoch_kd_diff"]= stoch_k - stoch_d
    df["stoch_ob"]     = np.where(stoch_k > 80, 1, 0)
    df["stoch_os"]     = np.where(stoch_k < 20, 1, 0)

    # ── Candle structure ────────────────────────────────────────────────────
    body      = (close - open_).abs()
    candle_rng= (high - low).replace(0, np.nan)
    df["body_ratio"]   = body / candle_rng        # body vs full range
    df["upper_wick"]   = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_rng
    df["lower_wick"]   = (pd.concat([close, open_], axis=1).min(axis=1) - low) / candle_rng
    df["candle_dir"]   = np.where(close > open_, 1, -1)
    df["candle_dir_3"] = df["candle_dir"].rolling(3).sum()  # short-term bias

    # ── Volume-less momentum oscillator (Williams %R 14) ───────────────────
    df["williams_r"] = -100 * (hi14 - close) / (hi14 - lo14).replace(0, np.nan)

    # ── RSI x MACD composite signal ─────────────────────────────────────────
    df["rsi_macd_bull"] = np.where(
        (df["rsi_14"] > 50) & (macd_hist > 0), 1,
        np.where((df["rsi_14"] < 50) & (macd_hist < 0), -1, 0)
    )

    # ── Divergence proxy: price vs RSI direction (3-bar) ────────────────────
    price_dir3 = np.sign(close.diff(3))
    rsi_dir3   = np.sign(df["rsi_14"].diff(3))
    df["rsi_div"] = np.where(price_dir3 != rsi_dir3, 1, 0)

    # ── Mean-reversion signal: distance from 50-bar mean normalised by ATR ──
    sma50 = close.rolling(50).mean()
    df["zscore_50"] = (close - sma50) / (close.rolling(50).std(ddof=0).replace(0, np.nan))

    df["mean_rev_long"]  = np.where(df["zscore_50"] < -1.5, 1, 0)
    df["mean_rev_short"] = np.where(df["zscore_50"] >  1.5, 1, 0)

    # ── Interaction features ─────────────────────────────────────────────────
    df["rsi_bb_pct"]        = df["rsi_14"] * df["bb_pct"]
    df["macd_hist_rsi_mid"] = df["macd_hist"] * df["rsi_mid"]
    df["stoch_rsi"]         = df["stoch_k"] * df["rsi_14"] / 1e4  # normalised product

    # ── Lag features (avoid lookahead) ──────────────────────────────────────
    for lag in [1, 2, 4, 8]:
        df[f"rsi_lag{lag}"]       = df["rsi_14"].shift(lag)
        df[f"macd_hist_lag{lag}"] = df["macd_hist"].shift(lag)
        df[f"ret_lag{lag}"]       = df["ret_1"].shift(lag)

    # ── Hour-of-day & day-of-week cyclic encoding ───────────────────────────
    if hasattr(df.index, "hour"):
        hour = df.index.hour
        dow  = df.index.dayofweek
        df["hour_sin"] = np.sin(2 * np.pi * hour / 24)
        df["hour_cos"] = np.cos(2 * np.pi * hour / 24)
        df["dow_sin"]  = np.sin(2 * np.pi * dow  / 5)
        df["dow_cos"]  = np.cos(2 * np.pi * dow  / 5)

    # ── Final fill ───────────────────────────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "NZD/USD RSI-MACD Gradient Boost Risk-Adjusted",
        "model_type":       "GradientBoostingClassifier",
        "model_params":     {
            "n_estimators":     600,
            "max_depth":        4,
            "learning_rate":    0.03,
            "subsample":        0.75,
            "max_features":     "sqrt",
            "min_samples_leaf": 20,
            "min_samples_split":40,
            "warm_start":       False,
        },
        "signal_threshold": 0.56,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [21, 21],
        "min_atr":          0.0002,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe/Calmar) using a deep "
            "GradientBoostingClassifier with many slow-learning trees and "
            "aggressive regularisation (min_samples_leaf=20, subsample=0.75). "
            "Feature set deliberately differs from prior RSI+BB+Stoch attempts "
            "by adding: ATR-normalised returns, z-score mean-reversion signals, "
            "RSI divergence proxy, Williams %R, candle structure ratios, cyclic "
            "time encoding, and interaction/lag features to give the model richer "
            "multi-timeframe context. SL=0.5%/TP=1% gives 1:2 RR aligned with "
            "maximising Sharpe."
        ),
        "notes": (
            "Prior PF=1.35 / ret=+18.36% used standard RSI+MACD+BB+Stoch without "
            "ATR normalisation or divergence detection. This version adds z-score "
            "mean-reversion context, candle structure, and temporal encoding to "
            "reduce false positives. session_filter=[21,21] is intentionally "
            "narrow — set to None if you want 24h coverage. min_atr=0.0002 "
            "avoids dead-market signals."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:27:43
# Model            : XGBoost
# Feature Eng.     : SMA (20,50,200) + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/GBPUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── Required SMAs and distance metrics ──────────────────────────────────
    for p in [20, 50, 200]:
        sma = close.rolling(p).mean()
        df[f"sma_{p}"] = sma
        df[f"dm_sma_{p}"] = (close - sma) / sma

    # ── SMA slope (momentum of the moving average itself) ───────────────────
    for p in [20, 50]:
        sma = close.rolling(p).mean()
        df[f"sma_{p}_slope"] = sma.diff(5) / sma.shift(5)

    # ── SMA cross signals ────────────────────────────────────────────────────
    sma20 = close.rolling(20).mean()
    sma50 = close.rolling(50).mean()
    sma200 = close.rolling(200).mean()
    df["sma20_50_cross"] = (sma20 - sma50) / sma50
    df["sma50_200_cross"] = (sma50 - sma200) / sma200
    df["sma20_200_cross"] = (sma20 - sma200) / sma200

    # ── Price momentum over multiple horizons ────────────────────────────────
    for lag in [1, 3, 6, 12, 24, 48]:
        df[f"ret_{lag}"] = close.pct_change(lag)

    # ── Volatility: rolling standard deviation of returns ───────────────────
    ret1 = close.pct_change(1)
    for w in [10, 20, 40]:
        df[f"vol_{w}"] = ret1.rolling(w).std()

    # ── ATR (Average True Range, normalised) ─────────────────────────────────
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    for w in [14, 28]:
        atr = tr.ewm(span=w, adjust=False).mean()
        df[f"natr_{w}"] = atr / close

    # ── Bollinger Bands (20-period, 2σ) ──────────────────────────────────────
    bb_mid  = close.rolling(20).mean()
    bb_std  = close.rolling(20).std()
    bb_up   = bb_mid + 2 * bb_std
    bb_lo   = bb_mid - 2 * bb_std
    bb_width = (bb_up - bb_lo) / bb_mid
    df["bb_pct_b"]  = (close - bb_lo) / (bb_up - bb_lo + 1e-12)
    df["bb_width"]  = bb_width
    df["bb_squeeze"]= np.where(bb_width < bb_width.rolling(50).mean(), 1.0, 0.0)

    # ── Keltner Channel (for squeeze confirmation) ───────────────────────────
    kc_mid  = close.ewm(span=20, adjust=False).mean()
    kc_atr  = tr.ewm(span=20, adjust=False).mean()
    kc_up   = kc_mid + 1.5 * kc_atr
    kc_lo   = kc_mid - 1.5 * kc_atr
    df["kc_pct"] = (close - kc_lo) / (kc_up - kc_lo + 1e-12)

    # ── RSI (Wilder) ─────────────────────────────────────────────────────────
    def wilder_rsi(src, period):
        delta = src.diff(1)
        gain  = delta.clip(lower=0)
        loss  = (-delta).clip(lower=0)
        avg_g = gain.ewm(alpha=1/period, adjust=False).mean()
        avg_l = loss.ewm(alpha=1/period, adjust=False).mean()
        rs    = avg_g / (avg_l + 1e-12)
        return 100 - 100 / (1 + rs)

    rsi14 = wilder_rsi(close, 14)
    rsi6  = wilder_rsi(close, 6)
    rsi28 = wilder_rsi(close, 28)
    df["rsi14"] = rsi14 / 100.0
    df["rsi6"]  = rsi6  / 100.0
    df["rsi28"] = rsi28 / 100.0
    df["rsi14_slope"] = rsi14.diff(3) / 100.0

    # RSI divergence proxy: price new high/low but RSI doesn't confirm
    price_high_12 = close.rolling(12).max()
    price_low_12  = close.rolling(12).min()
    rsi_high_12   = rsi14.rolling(12).max()
    rsi_low_12    = rsi14.rolling(12).min()
    df["rsi_bear_div"] = np.where(
        (close >= price_high_12 * 0.999) & (rsi14 < rsi_high_12 * 0.97), 1.0, 0.0)
    df["rsi_bull_div"] = np.where(
        (close <= price_low_12  * 1.001) & (rsi14 > rsi_low_12  * 1.03), 1.0, 0.0)

    # ── MACD ─────────────────────────────────────────────────────────────────
    ema12  = close.ewm(span=12, adjust=False).mean()
    ema26  = close.ewm(span=26, adjust=False).mean()
    macd   = ema12 - ema26
    signal = macd.ewm(span=9, adjust=False).mean()
    hist   = macd - signal
    df["macd_norm"]    = macd   / close
    df["macd_sig_norm"]= signal / close
    df["macd_hist_norm"]= hist  / close
    df["macd_hist_slope"] = hist.diff(2) / close

    # ── Stochastic Oscillator ─────────────────────────────────────────────────
    for k_period in [14, 5]:
        lo_k = low.rolling(k_period).min()
        hi_k = high.rolling(k_period).max()
        stoch_k = (close - lo_k) / (hi_k - lo_k + 1e-12) * 100
        stoch_d = stoch_k.rolling(3).mean()
        df[f"stoch_k_{k_period}"] = stoch_k / 100.0
        df[f"stoch_d_{k_period}"] = stoch_d / 100.0
        df[f"stoch_kd_{k_period}"] = (stoch_k - stoch_d) / 100.0

    # ── Williams %R ───────────────────────────────────────────────────────────
    hi14 = high.rolling(14).max()
    lo14 = low.rolling(14).min()
    df["williams_r"] = (hi14 - close) / (hi14 - lo14 + 1e-12)

    # ── CCI (Commodity Channel Index) ────────────────────────────────────────
    tp = (high + low + close) / 3.0
    tp_sma = tp.rolling(20).mean()
    tp_mad  = tp.rolling(20).apply(lambda x: np.mean(np.abs(x - x.mean())), raw=True)
    df["cci"] = (tp - tp_sma) / (0.015 * tp_mad + 1e-12) / 100.0

    # ── Volume-like proxy: candle body and wick ratios ────────────────────────
    candle_range = (high - low).replace(0, np.nan)
    df["body_ratio"]  = (close - open_).abs() / candle_range
    df["upper_wick"]  = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_range
    df["lower_wick"]  = (pd.concat([close, open_], axis=1).min(axis=1) - low) / candle_range
    df["bull_candle"]  = np.where(close > open_, 1.0, 0.0)

    # ── Mean reversion signal: z-score of close vs SMA20 ────────────────────
    df["zscore_20"] = (close - sma20) / (close.rolling(20).std() + 1e-12)
    df["zscore_50"] = (close - sma50) / (close.rolling(50).std() + 1e-12)

    # ── Trend strength: ADX proxy ─────────────────────────────────────────────
    plus_dm  = (high.diff(1)).clip(lower=0)
    minus_dm = (-low.diff(1)).clip(lower=0)
    overlap  = pd.concat([plus_dm, minus_dm], axis=1).min(axis=1)
    plus_dm  = plus_dm - overlap
    minus_dm = minus_dm - overlap
    atr14 = tr.ewm(span=14, adjust=False).mean()
    plus_di  = 100 * plus_dm.ewm(span=14, adjust=False).mean()  / (atr14 + 1e-12)
    minus_di = 100 * minus_dm.ewm(span=14, adjust=False).mean() / (atr14 + 1e-12)
    dx = (plus_di - minus_di).abs() / (plus_di + minus_di + 1e-12) * 100
    adx = dx.ewm(span=14, adjust=False).mean()
    df["adx"]      = adx / 100.0
    df["plus_di"]  = plus_di / 100.0
    df["minus_di"] = minus_di / 100.0
    df["di_diff"]  = (plus_di - minus_di) / 100.0

    # ── Regime detection: above/below long-term SMA ──────────────────────────
    df["bull_regime"]  = np.where(close > sma200, 1.0, 0.0)
    df["mid_regime"]   = np.where(close > sma50,  1.0, 0.0)

    # ── Lag features (auto-regressive) ───────────────────────────────────────
    for col, lags in [("rsi14", [1, 2, 4]), ("macd_hist_norm", [1, 2]), ("bb_pct_b", [1, 2])]:
        for lag in lags:
            df[f"{col}_lag{lag}"] = df[col].shift(lag)

    # ── Time-of-day features ─────────────────────────────────────────────────
    if hasattr(df.index, "hour"):
        df["hour_sin"] = np.sin(2 * np.pi * df.index.hour / 24.0)
        df["hour_cos"] = np.cos(2 * np.pi * df.index.hour / 24.0)
        df["dow_sin"]  = np.sin(2 * np.pi * df.index.dayofweek / 5.0)
        df["dow_cos"]  = np.cos(2 * np.pi * df.index.dayofweek / 5.0)

    # ── Fill NaN from warm-up ────────────────────────────────────────────────
    df = df.bfill().ffill()
    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "GBP/USD SMA Trend + Multi-Indicator XGBoost Classifier",
        "model_type":       "XGBClassifier",
        "model_params": {
            "n_estimators":     600,
            "max_depth":        4,
            "learning_rate":    0.03,
            "subsample":        0.75,
            "colsample_bytree": 0.70,
            "min_child_weight": 5,
            "gamma":            0.10,
            "reg_alpha":        0.10,
            "reg_lambda":       1.50,
            "objective":        "binary:logistic",
            "tree_method":      "hist",
            "random_state":     42,
            "n_jobs":           -1,
        },
        "signal_threshold": 0.56,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 18],
        "min_atr":          0.0003,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return on GBP/USD 15-min bars. "
            "Strategy combines required SMA (20/50/200) distance and cross features "
            "with ADX trend strength, RSI divergence, Bollinger squeeze, Keltner, "
            "MACD histogram slope, Stochastic, CCI, Williams %R, and candle-structure "
            "ratios. XGBoost with strong regularisation and subsampling prevents "
            "overfitting on the relatively short 1-year window. "
            "Session filter 06-18 UTC keeps execution in liquid London/NY hours; "
            "0.5% SL and 1.0% TP yield 1:2 R:R; sma_50 trend filter aligns trades "
            "with intermediate momentum to improve win rate and Sharpe."
        ),
        "notes": (
            "Differs from prior RSI/MACD/BB/Stoch attempts by: (1) foregrounding "
            "SMA cross and distance features as primary trend signals; (2) adding "
            "ADX-based regime and DI differential; (3) including RSI divergence "
            "proxy flags; (4) z-score mean-reversion features; (5) candle body/wick "
            "structure ratios as micro-structure proxies; (6) time-of-day cyclical "
            "encoding; (7) heavier regularisation (gamma, alpha, lambda) and higher "
            "min_child_weight to reduce variance on the thin dataset."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:51:31
# Model            : XGBoost
# Feature Eng.     : BB (20,2.0), RSI 14, Stochastic (14,3) + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/AUDUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):

    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_period = 20
    bb_std    = 2.0
    bb_mid    = close.rolling(bb_period).mean()
    bb_std_val = close.rolling(bb_period).std(ddof=0)
    bb_upper  = bb_mid + bb_std * bb_std_val
    bb_lower  = bb_mid - bb_std * bb_std_val

    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid
    df["bb_pct"]   = (close - bb_lower) / (bb_upper - bb_lower)

    # ── RSI (14) ─────────────────────────────────────────────────────────────
    rsi_period = 14
    delta      = close.diff()
    gain       = delta.clip(lower=0)
    loss       = -delta.clip(upper=0)
    avg_gain   = gain.ewm(com=rsi_period - 1, min_periods=rsi_period).mean()
    avg_loss   = loss.ewm(com=rsi_period - 1, min_periods=rsi_period).mean()
    rs         = avg_gain / avg_loss.replace(0, np.nan)
    df["rsi"]  = 100 - (100 / (1 + rs))

    # ── Stochastic Oscillator (K=14, D=3) ────────────────────────────────────
    stoch_k_period = 14
    stoch_d_period = 3
    lowest_low     = low.rolling(stoch_k_period).min()
    highest_high   = high.rolling(stoch_k_period).max()
    denom          = (highest_high - lowest_low).replace(0, np.nan)
    df["stoch_k"]  = 100 * (close - lowest_low) / denom
    df["stoch_d"]  = df["stoch_k"].rolling(stoch_d_period).mean()
    df["stoch_kd_diff"] = df["stoch_k"] - df["stoch_d"]

    # ── Additional derived features ──────────────────────────────────────────

    # RSI overbought / oversold zone flags
    df["rsi_ob"]   = np.where(df["rsi"] > 70, 1, 0)
    df["rsi_os"]   = np.where(df["rsi"] < 30, 1, 0)
    df["rsi_mid"]  = df["rsi"] - 50.0

    # Stochastic overbought / oversold zone flags
    df["stoch_ob"] = np.where(df["stoch_k"] > 80, 1, 0)
    df["stoch_os"] = np.where(df["stoch_k"] < 20, 1, 0)

    # BB position regime: price relative to bands
    df["price_above_bb_upper"] = np.where(close > bb_upper, 1, 0)
    df["price_below_bb_lower"] = np.where(close < bb_lower, 1, 0)
    df["price_vs_bb_mid"]      = close - bb_mid

    # ATR-based volatility (14-bar)
    atr_period = 14
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    df["atr14"]  = tr.ewm(com=atr_period - 1, min_periods=atr_period).mean()
    df["natr14"] = df["atr14"] / close

    # SMA trend context
    df["sma_20"]  = close.rolling(20).mean()
    df["sma_50"]  = close.rolling(50).mean()
    df["sma_200"] = close.rolling(200).mean()
    df["price_vs_sma20"]  = (close - df["sma_20"]) / df["sma_20"]
    df["price_vs_sma50"]  = (close - df["sma_50"]) / df["sma_50"]
    df["sma20_vs_sma50"]  = (df["sma_20"] - df["sma_50"]) / df["sma_50"]

    # Momentum: rate of change
    df["roc_5"]  = close.pct_change(5)
    df["roc_10"] = close.pct_change(10)
    df["roc_20"] = close.pct_change(20)

    # MACD-style (EMA 12 - EMA 26)
    ema12        = close.ewm(span=12, min_periods=12).mean()
    ema26        = close.ewm(span=26, min_periods=26).mean()
    macd_line    = ema12 - ema26
    macd_signal  = macd_line.ewm(span=9, min_periods=9).mean()
    df["macd"]        = macd_line
    df["macd_signal"] = macd_signal
    df["macd_hist"]   = macd_line - macd_signal

    # Candle body / wick ratios
    body          = (close - open_).abs()
    candle_range  = (high - low).replace(0, np.nan)
    df["body_ratio"]  = body / candle_range
    df["upper_wick"]  = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_range
    df["lower_wick"]  = (pd.concat([close, open_], axis=1).min(axis=1) - low) / candle_range
    df["bullish_bar"]  = np.where(close > open_, 1, 0)

    # Lagged RSI / Stoch features (1 and 2 bars back)
    df["rsi_lag1"]     = df["rsi"].shift(1)
    df["rsi_lag2"]     = df["rsi"].shift(2)
    df["stoch_k_lag1"] = df["stoch_k"].shift(1)
    df["bb_pct_lag1"]  = df["bb_pct"].shift(1)

    # RSI slope
    df["rsi_slope"] = df["rsi"] - df["rsi"].shift(3)

    # Stoch K crossing D (momentum signal)
    df["stoch_cross_up"]   = np.where((df["stoch_k"] > df["stoch_d"]) &
                                       (df["stoch_k"].shift(1) <= df["stoch_d"].shift(1)), 1, 0)
    df["stoch_cross_down"] = np.where((df["stoch_k"] < df["stoch_d"]) &
                                       (df["stoch_k"].shift(1) >= df["stoch_d"].shift(1)), 1, 0)

    # Volume (if present)
    if "volume" in df.columns:
        vol_ma = df["volume"].rolling(20).mean()
        df["vol_ratio"] = df["volume"] / vol_ma.replace(0, np.nan)
    else:
        df["vol_ratio"] = 1.0

    # ── Fill NaN from warm-up ────────────────────────────────────────────────
    df = df.bfill().ffill()

    return df

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "AUD/USD Stoch+BB+RSI Mean-Reversion XGBoost",
        "model_type":       "XGBClassifier",
        "model_params": {
            "n_estimators":       400,
            "max_depth":          4,
            "learning_rate":      0.04,
            "subsample":          0.75,
            "colsample_bytree":   0.70,
            "min_child_weight":   5,
            "gamma":              0.15,
            "reg_alpha":          0.10,
            "reg_lambda":         1.50,
            "objective":          "binary:logistic",
            "random_state":       42,
            "n_jobs":             -1,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [0, 23],
        "min_atr":          None,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe / Calmar). "
            "XGBoost chosen for its ability to capture non-linear interactions "
            "between Stochastic, Bollinger Bands, and RSI regimes. "
            "Shallow trees (max_depth=4) + high regularisation (reg_lambda=1.5, gamma=0.15) "
            "prevent overfitting on 15-min FX data. "
            "2:1 TP:SL ratio (1.0% / 0.5%) improves expectancy per trade. "
            "Reverse on opposite signal minimises flat time and captures regime flips."
        ),
        "notes": (
            "Features include BB width/pct, RSI(14) with overbought/oversold flags, "
            "Stochastic K/D crossovers, MACD histogram, ATR volatility, SMA trend context, "
            "candle body ratios, lagged indicators, and momentum ROC. "
            "signal_threshold=0.55 balances precision vs recall on directional calls. "
            "session_filter covers full 24h to capture Asia + London + NY sessions for AUD/USD."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-25 02:27:36
# Model            : XGBoost
# Feature Eng.     : go long when EMA(9) crosses above EMA(21), exit when it crosses back below + Auto-add features: ON
# Signal / Entry   : —
# Optimization     : —
# Risk Mgmt        : —
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

# ── Inlined strategy_utils ──
"""
strategy_utils.py — Standard utility functions for generated strategies.
Claude imports these instead of writing boilerplate from scratch.
This ensures consistent behavior across all generated strategies.
"""
import numpy as np
import pandas as pd
from sklearn.preprocessing import LabelEncoder


# Max backtest window per timeframe. A finer timeframe over a longer window
# blows up the results dict / parquet load / Modal train time (the 2026-05-12
# OOM was a 1-min × multi-year sweep) — and a 1-min strategy gains nothing from
# 2 years of 1-min bars. Enforced HERE because every training path (UI / API /
# Modal) funnels through run_strategy → load_ohlc. Env-overridable so a future
# "max plan" / dedicated-server tier can lift it.
_TF_MAX_DAYS = {
    "1min":  30,
    "5min":  90,
    "15min": 365,
    "1h":    730,
}


def _fetch_ohlc_from_internal(symbol: str, tf: str, start: str, end: str):
    """Phase 3.2: fetch parquet bytes from Server A's /internal/ohlc endpoint
    instead of reading a local file. Used inside Modal containers / Mac worker
    pool (Phase 3.4) so every train sees the same source of truth as the chart.

    Returns: pd.DataFrame (parquet decoded), or raises on any failure so the
    caller can fall back / surface a clear error in the job.
    """
    import hashlib as _hashlib, hmac as _hmac, io as _io, os as _os
    import urllib.request as _ur, urllib.parse as _urp

    base = (_os.environ.get("QM_INTERNAL_OHLC_BASE") or "").rstrip("/")
    secret = (_os.environ.get("INTERNAL_WS_SECRET") or "").strip()
    if not base:
        raise RuntimeError("QM_INTERNAL_OHLC_BASE not set")
    if not secret:
        raise RuntimeError("INTERNAL_WS_SECRET not set")

    msg = f"{symbol}|{tf}|{start}|{end}".encode("utf-8")
    sig = _hmac.new(secret.encode("utf-8"), msg, _hashlib.sha256).hexdigest()

    qs = _urp.urlencode({
        "symbol": symbol, "tf": tf,
        "start": start, "end": end, "sig": sig,
    })
    url = f"{base}/internal/ohlc?{qs}"

    req = _ur.Request(url, headers={"User-Agent": "qm-worker/1.0"})
    with _ur.urlopen(req, timeout=30) as resp:
        if resp.status != 200:
            raise RuntimeError(f"/internal/ohlc returned {resp.status}")
        payload = resp.read()
    print(f"[load_ohlc:internal] {symbol} {tf} fetched {len(payload)} bytes", flush=True)
    return pd.read_parquet(_io.BytesIO(payload))


def _parse_symbol_tf_from_path(data_path: str):
    """Pull SYMBOL + TF out of a path like .../EURUSD_1min.parquet."""
    import os as _os, re as _re
    base = _os.path.basename(str(data_path))
    m = _re.match(r"^([A-Z]{6})_(\d+min|\d+h)\.parquet$", base)
    if not m:
        return None, None
    return m.group(1), m.group(2)


def load_ohlc(data_path, start_date="", end_date=""):
    """Load OHLC parquet, sort index, filter dates. Always returns consistent format.

    The lower bound is clamped per timeframe (see _TF_MAX_DAYS) — a request for
    more history than the cap silently starts later.

    Phase 3.2: when env QM_USE_INTERNAL_OHLC=="1", fetch over HTTP from
    Server A's /internal/ohlc endpoint instead of pd.read_parquet on a local
    file (which on Modal is a stale Volume snapshot). The endpoint applies the
    same day-cap, so the local cap-check below is a defensive no-op in that
    path. Flag defaults to "0" → unchanged behavior.

    Returns: (df, close, open_, high, low)
    """
    import os as _os, re as _re

    _use_internal = _os.environ.get("QM_USE_INTERNAL_OHLC", "0") == "1"
    if _use_internal:
        _sym, _tf = _parse_symbol_tf_from_path(data_path)
        if not _sym or not _tf:
            raise RuntimeError(
                f"QM_USE_INTERNAL_OHLC=1 but DATA_PATH basename does not match "
                f"SYMBOL_TF.parquet: {data_path}"
            )
        df = _fetch_ohlc_from_internal(_sym, _tf, start_date or "", end_date or "")
    else:
        df = pd.read_parquet(data_path)
    df.index = pd.to_datetime(df.index)
    df = df.sort_index()

    # Per-timeframe window cap (timeframe inferred from the parquet filename).
    _m = _re.search(r"_(\d+min|\d+h)\.parquet$", _os.path.basename(str(data_path)))
    _tf = _m.group(1) if _m else None
    _max_days = _TF_MAX_DAYS.get(_tf)
    if _max_days and _max_days > 0 and len(df):
        _env_override = _os.environ.get(f"QM_MAX_DAYS_{_tf.upper()}")
        if _env_override and _env_override.isdigit():
            _max_days = int(_env_override)
        try:
            _eff_end = pd.Timestamp(end_date) if end_date else df.index.max()
            _eff_end = min(_eff_end, df.index.max())
            _floor = _eff_end - pd.Timedelta(days=_max_days)
            _req_start = pd.Timestamp(start_date) if start_date else df.index.min()
            if _req_start < _floor:
                print(f"[load_ohlc] {_tf} backtest window capped to {_max_days}d: "
                      f"start {_req_start.date()} -> {_floor.date()}", flush=True)
                start_date = _floor
        except Exception as _e:
            print(f"[load_ohlc] window-cap check skipped ({_e})", flush=True)

    if start_date:
        df = df[df.index >= start_date]
    if end_date:
        df = df[df.index <= end_date]

    return df, df["close"], df["open"], df["high"], df["low"]


def make_target(close, horizon=4):
    """Create target: direction N bars ahead. Default 4 bars = 1 hour on 15-min data.

    Returns: target (pd.Series of -1, 0, 1)
    """
    return np.sign(close.shift(-horizon) - close)


def split_data(df, target, feature_cols, train_split=0.7, validation_date=""):
    """Train/test split. Handles both ratio and date-based splits.
    Drops NaN from target before splitting. Encodes labels to [0,1,2].

    Returns: dict with keys:
        X_train, X_test, y_train, y_test,
        y_train_enc, y_test_enc, enc,
        close_train, close_test,
        split_idx, split_dt, n_train, n_test
    """
    # Drop NaN from target
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]

    close = df["close"]

    # Build feature matrix
    X = df[feature_cols].copy()
    X = X.bfill().ffill()
    X = X.replace([np.inf, -np.inf], np.nan).fillna(0.0)

    # Split
    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)

    split_idx = max(1, min(split_idx, len(df) - 1))

    X_train = X.iloc[:split_idx]
    X_test = X.iloc[split_idx:]
    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]

    split_dt = str(df.index[split_idx])

    # Label encoding — always fit on [-1, 0, 1]
    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    return {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }


def compute_overlays(close, df_index):
    """Compute BB and MA overlays on full dataset. Always consistent.

    Returns: (bb_dict, ma_dict)
    """
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std

    ma50 = close.rolling(50).mean()
    ma100 = close.rolling(100).mean()
    ma200 = close.rolling(200).mean()

    def _safe(s):
        s = s.reindex(df_index).bfill().ffill()
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in s.values]

    bb = {"upper": _safe(bb_upper), "mid": _safe(bb_mid), "lower": _safe(bb_lower)}
    ma = {"ma50": _safe(ma50), "ma100": _safe(ma100), "ma200": _safe(ma200)}

    return bb, ma


def run_backtest(signal, close, capital=10000, cost=2e-5):
    """Run backtest with transaction costs.
    Uses price-based trade returns (same as webapp _compute_trades).
    Signal 0 = hold (keep current position), not close.

    Returns: dict with equity, trade_returns, long_returns, short_returns, bar_returns
    """
    sig_arr = signal.values
    price_arr = close.values
    idx = signal.index
    n = len(price_arr)

    # Trade returns — price-based (matches webapp _compute_trades exactly)
    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    last_dir = None
    entry_price = None
    entry_bar = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != last_dir:
            # Direction change — close previous trade, open new
            if last_dir is not None and entry_price is not None and entry_price != 0:
                ret = float(last_dir * (c - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if last_dir == 1:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                trade_log.append({
                    "type": "Buy" if last_dir == 1 else "Sell",
                    "entry_time": str(idx[entry_bar]),
                    "exit_time": str(idx[i]),
                    "entry_price": round(entry_price, 5),
                    "exit_price": round(c, 5),
                    "pnl": round(last_dir * (c - entry_price), 5),
                    "pnl_pct": round(ret * 100, 3),
                    "exit_reason": "signal",
                })
            entry_price = c
            entry_bar = i
            last_dir = s

    # Close last open trade
    if last_dir is not None and entry_price is not None and n > 0 and entry_price != 0:
        c = price_arr[-1]
        ret = float(last_dir * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if last_dir == 1:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        trade_log.append({
            "type": "Buy" if last_dir == 1 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[-1]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(c, 5),
            "pnl": round(last_dir * (c - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": "end",
        })

    # Equity curve from trade returns
    cumret = 1.0
    equity_vals = np.full(n, float(capital))
    trade_idx = 0
    in_trade = False
    t_entry_price = None
    t_dir = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != t_dir:
            if t_dir is not None and t_entry_price is not None and t_entry_price != 0:
                t_ret = t_dir * (c - t_entry_price) / t_entry_price - cost
                cumret *= (1 + t_ret)
            t_entry_price = c
            t_dir = s
        equity_vals[i] = capital * cumret

    # Bar returns for Sharpe
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if price_arr[i - 1] != 0 and last_dir is not None:
            bar_returns[i] = sig_arr[i - 1] * (price_arr[i] - price_arr[i - 1]) / price_arr[i - 1] if sig_arr[i - 1] != 0 else 0.0

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


def compute_trade_stats(trades, capital=10000):
    """Single source of truth for trade statistics.
    Every display path reads from this — no recomputation anywhere.
    All values are rounded and JSON-safe (no inf/nan).
    """
    if not trades:
        return {"n": 0, "wins": 0, "losses": 0, "wr": 0, "avg": 0,
                "best": 0, "worst": 0, "ret": 0, "np": 0, "mdd": 0,
                "pf": 0, "rr": 0, "expect": 0}
    w = [r for r in trades if r > 0]
    l = [r for r in trades if r < 0]
    cumret = 1.0
    for r in trades:
        cumret *= (1 + r)
    net_p = capital * (cumret - 1)
    # Max drawdown
    eq = np.cumprod([1.0] + [1 + r for r in trades])
    peak = np.maximum.accumulate(eq)
    mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    # Profit Factor
    gross_w = sum(w) if w else 0
    gross_l = abs(sum(l)) if l else 0
    pf = gross_w / gross_l if gross_l > 0 else (9999.0 if gross_w > 0 else 0)
    # Risk:Reward
    avg_w = float(np.mean(w)) if w else 0
    avg_l = abs(float(np.mean(l))) if l else 0
    rr = avg_w / avg_l if avg_l > 0 else (9999.0 if avg_w > 0 else 0)
    # Expectancy
    expect = net_p / len(trades)
    return {
        "n": len(trades), "wins": len(w), "losses": len(l),
        "wr": round(len(w) / len(trades), 4),
        "avg": round(float(np.mean(trades)), 6),
        "best": round(max(w), 6) if w else 0,
        "worst": round(min(l), 6) if l else 0,
        "ret": round(cumret - 1, 6),
        "np": round(net_p, 2),
        "mdd": round(mdd, 6),
        "pf": round(pf, 2),
        "rr": round(rr, 2),
        "expect": round(expect, 2),
    }


def compute_metrics(bt_result, close_test, capital=10000):
    """Compute all standard metrics from backtest result.
    Uses trade-level compounding (same as webapp _trade_stats) for accuracy.

    Returns: dict with total_ret, bh_ret, sharpe_strat, sharpe_bh, mdd, n_trades
    """
    equity = bt_result["equity"]
    trade_returns = bt_result["trade_returns"]

    # Total return — trade-level compounding (matches webapp)
    if trade_returns:
        cumret = 1.0
        for r in trade_returns:
            cumret *= (1 + r)
        total_ret = cumret - 1
    else:
        total_ret = 0.0

    # Buy and hold
    bh_equity = capital * (close_test / close_test.iloc[0])
    bh_ret = (bh_equity.iloc[-1] - capital) / capital if capital != 0 else 0.0

    # Sharpe ratio — trade-level (matches webapp: sqrt(252*26) annualization)
    if len(trade_returns) >= 2 and float(np.std(trade_returns)) > 0:
        sharpe_strat = float(np.mean(trade_returns) / np.std(trade_returns) * np.sqrt(252 * 26))
    else:
        sharpe_strat = 0.0

    bh_rets = bh_equity.pct_change().dropna()
    if len(bh_rets) > 1 and bh_rets.std() != 0:
        sharpe_bh = float((bh_rets.mean() / bh_rets.std()) * np.sqrt(252 * 24 * 4))
    else:
        sharpe_bh = 0.0

    # Max drawdown — trade-level (matches webapp)
    if trade_returns:
        eq = np.cumprod([1.0] + [1 + r for r in trade_returns])
        peak = np.maximum.accumulate(eq)
        mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    else:
        mdd = 0.0

    return {
        "total_ret": float(total_ret),
        "bh_ret": float(bh_ret),
        "sharpe_strat": float(sharpe_strat) if not np.isnan(sharpe_strat) else 0.0,
        "sharpe_bh": float(sharpe_bh) if not np.isnan(sharpe_bh) else 0.0,
        "mdd": float(mdd),
        "n_trades": len(trade_returns),
    }


# Diagnostics line/histogram series (equity / drawdown / rolling_acc / conf_hist)
# only feed the small Diagnostics charts — they're never used by the price chart
# or scroll-back. On a 1-min model trained over the (2.2-capped) window these are
# still ~30k points each; downsample to a visually-identical resolution before the
# dict leaves the trainer so it doesn't carry that into Server-A RAM / Postgres.
_RESULTS_SERIES_MAX = 5000


def _downsample_idx(n, cap=_RESULTS_SERIES_MAX):
    """Evenly-spaced index list spanning [0, n-1] (first+last always kept), or
    None when no downsampling is needed (n <= cap)."""
    if n <= cap:
        return None
    return np.unique(np.linspace(0, n - 1, cap).astype(int)).tolist()


def _take(arr, idx):
    """Subset a list by an index list (idx may be None → return arr unchanged)."""
    if idx is None or not isinstance(arr, list):
        return arr
    return [arr[i] for i in idx]


# trade_log / train_trade_log are lists of per-trade dicts (display-only — the
# Trade Log tab). They scale with TRADE count, not bar count, so the bar-window
# cap (Phase 2.2) doesn't bound them — a degenerate near-every-bar model can put
# 10k+ trade dicts in the blob (>3 MB). Cap each (independently — a small-N model
# keeps every trade) to the most-recent N, recording `*_total` + `*_truncated`
# so the true count is still reported. Real strategies have far fewer than
# _TRADE_LOG_MAX trades, so this only ever bites pathological models.
_TRADE_LOG_MAX = 5000


def _cap_trade_log(tl):
    """Return (capped_list, original_len, was_truncated)."""
    if not isinstance(tl, list) or len(tl) <= _TRADE_LOG_MAX:
        return tl, (len(tl) if isinstance(tl, list) else 0), False
    return tl[-_TRADE_LOG_MAX:], len(tl), True


def build_return_dict(split_result, bt_result, metrics, model, feature_cols,
                      signal_full, p_pos_test, p_neg_test, custom_figs=None,
                      bt_train_result=None, pre_stats=None):
    """Assemble the complete return dict. Handles ALL serialization.
    Never returns Timestamps, numpy arrays, or non-JSON types.

    Returns: JSON-safe dict with all required keys
    """
    df = split_result["df"]
    close = split_result["close"]
    close_test = split_result["close_test"]
    X_test = split_result["X_test"]
    y_test = split_result["y_test"]
    equity = bt_result["equity"]
    bar_returns = bt_result["bar_returns"]

    # OHLC
    ohlc_dates = [str(x) for x in df.index.tolist()]

    def _safe_list(arr):
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in arr]

    # Overlays
    bb, ma = compute_overlays(close, df.index)

    # Buy and hold equity
    capital = equity.iloc[0] if len(equity) > 0 else 10000
    bh_equity = capital * (close_test / close_test.iloc[0])

    # Confusion matrix
    from sklearn.metrics import confusion_matrix
    pred_test = model.predict(X_test)
    y_test_arr = np.asarray(y_test)
    cm = confusion_matrix(y_test_arr, pred_test, labels=[-1, 0, 1])

    # Rolling accuracy
    sig_arr = signal_full.reindex(close_test.index).values
    correct = pd.Series((pred_test == y_test_arr).astype(float), index=X_test.index)
    active_test = pd.Series(sig_arr != 0, index=close_test.index) if len(sig_arr) == len(close_test) else pd.Series(True, index=close_test.index)
    correct_active = correct.where(active_test, other=np.nan)
    rolling_acc = correct_active.rolling(30, min_periods=1).mean()

    # Feature importance
    importances = model.feature_importances_
    fi_pairs = sorted(zip(feature_cols, importances), key=lambda x: x[1])[-15:]

    # Drawdown
    rolling_max = equity.cummax()
    drawdown = (equity - rolling_max) / rolling_max.replace(0, np.nan)
    drawdown = drawdown.fillna(0.0)

    # ── Downsample the Diagnostics-only series (see _downsample_idx) ──────────
    _eq_dates = [str(x) for x in close_test.index.tolist()]
    _eq_strat = _safe_list(equity.values)
    _eq_bh    = _safe_list(bh_equity.values)
    _eq_idx   = _downsample_idx(len(_eq_dates))
    _eq_dates, _eq_strat, _eq_bh = _take(_eq_dates, _eq_idx), _take(_eq_strat, _eq_idx), _take(_eq_bh, _eq_idx)

    _ra_dates = [str(x) for x in rolling_acc.index.tolist()]
    _ra_vals  = [float(x) if (not np.isnan(x) and not np.isinf(x)) else None for x in rolling_acc.values]
    _ra_idx   = _downsample_idx(len(_ra_dates))
    _ra_dates, _ra_vals = _take(_ra_dates, _ra_idx), _take(_ra_vals, _ra_idx)

    _dd_dates = [str(x) for x in drawdown.index.tolist()]
    _dd_vals  = _safe_list(drawdown.values)
    _dd_idx   = _downsample_idx(len(_dd_dates))
    _dd_dates, _dd_vals = _take(_dd_dates, _dd_idx), _take(_dd_vals, _dd_idx)

    _cp_pos = [float(x) for x in (p_pos_test.tolist() if hasattr(p_pos_test, 'tolist') else list(p_pos_test))]
    _cp_neg = [float(x) for x in (p_neg_test.tolist() if hasattr(p_neg_test, 'tolist') else list(p_neg_test))]
    _cp_pos = _take(_cp_pos, _downsample_idx(len(_cp_pos)))
    _cp_neg = _take(_cp_neg, _downsample_idx(len(_cp_neg)))

    # ── Trade logs — display-only (Trade Log tab); cap to most-recent N with a
    #    `_total` field so the true count is still reported (see _cap_trade_log).
    #    NB: ret_dist arrays are left FULL — a downstream path in callbacks.py
    #    recomputes n_trades/win-rate from len(ret_dist), so a sample would skew
    #    the displayed counts; they're small anyway and gzip handles them.
    _tl_test, _tl_test_n, _tl_test_tr = _cap_trade_log(bt_result.get("trade_log", []))
    _tl_tr,   _tl_tr_n,   _tl_tr_tr   = _cap_trade_log(bt_train_result.get("trade_log", []) if bt_train_result else [])

    return {
        "ohlc": {
            "dates": ohlc_dates,
            "open": _safe_list(df["open"].values),
            "high": _safe_list(df["high"].values),
            "low": _safe_list(df["low"].values),
            "close": _safe_list(df["close"].values),
        },
        "signals": {
            "dates": [str(x) for x in signal_full.index.tolist()],
            "values": [float(x) for x in signal_full.values],
        },
        "bb": bb,
        "ma": ma,
        "equity": {
            "dates": _eq_dates,
            "strategy": _eq_strat,
            "bh": _eq_bh,
        },
        "feature_importance": {
            "names": [p[0] for p in fi_pairs],
            "values": [float(p[1]) for p in fi_pairs],
        },
        "conf_matrix": cm.tolist(),
        "conf_hist": {
            "p_pos": _cp_pos,
            "p_neg": _cp_neg,
        },
        "rolling_acc": {
            "dates": _ra_dates,
            "values": _ra_vals,
        },
        "drawdown": {
            "dates": _dd_dates,
            "values": _dd_vals,
        },
        "ret_dist": [float(x) for x in bt_result["trade_returns"]],
        "ret_dist_long": [float(x) for x in bt_result["long_returns"]],
        "ret_dist_short": [float(x) for x in bt_result["short_returns"]],
        "train_ret_dist": [float(x) for x in bt_train_result["trade_returns"]] if bt_train_result else [],
        "train_ret_dist_long": [float(x) for x in bt_train_result["long_returns"]] if bt_train_result else [],
        "train_ret_dist_short": [float(x) for x in bt_train_result["short_returns"]] if bt_train_result else [],
        "trade_log": _tl_test,
        "train_trade_log": _tl_tr,
        "trade_log_total": _tl_test_n,
        "train_trade_log_total": _tl_tr_n,
        "trade_log_truncated": _tl_test_tr,
        "train_trade_log_truncated": _tl_tr_tr,
        **(pre_stats or {}),
        "metrics": metrics,
        "split_dt": split_result["split_dt"],
        "split_idx": int(split_result["split_idx"]),
        "n_train": int(split_result["n_train"]),
        "n_test": int(split_result["n_test"]),
        "feature_cols": list(feature_cols),
        "custom_figs": custom_figs or [],
    }


# ════════════════════════════════════════════════════════════════════════════
# STRATEGY FRAMEWORK v2 — Config-driven architecture
# Claude writes feature_engineering() + strategy_config(). Framework does rest.
# ════════════════════════════════════════════════════════════════════════════

import importlib

_MODEL_REGISTRY = {
    "XGBClassifier":              ("xgboost", "XGBClassifier"),
    "RandomForestClassifier":     ("sklearn.ensemble", "RandomForestClassifier"),
    "GradientBoostingClassifier": ("sklearn.ensemble", "GradientBoostingClassifier"),
    "LogisticRegression":         ("sklearn.linear_model", "LogisticRegression"),
    "ExtraTreesClassifier":       ("sklearn.ensemble", "ExtraTreesClassifier"),
    "AdaBoostClassifier":         ("sklearn.ensemble", "AdaBoostClassifier"),
}


def _build_model_from_config(config, X_train, y_train_enc):
    """Build, fit, and wrap a model from strategy_config dict."""
    model_type = config.get("model_type", "RandomForestClassifier")
    model_params = dict(config.get("model_params", {}))

    if model_type not in _MODEL_REGISTRY:
        raise ValueError(f"Unknown model_type '{model_type}'. Valid: {list(_MODEL_REGISTRY.keys())}")

    module_path, class_name = _MODEL_REGISTRY[model_type]
    mod = importlib.import_module(module_path)
    cls = getattr(mod, class_name)

    # XGBoost defaults
    if class_name == "XGBClassifier":
        model_params.setdefault("use_label_encoder", False)
        model_params.setdefault("eval_metric", "mlogloss")
        model_params.setdefault("tree_method", "hist")
        # Determinism > speed (2026-05-25). XGBoost hist with n_jobs=-1 is
        # NON-reproducible even with random_state set — the parallel histogram
        # gradient-sum order varies across threads, so the SAME code + data
        # gives a slightly different model (and backtest) every run. Forcing
        # single-thread makes training bit-reproducible so: (a) a user who
        # copies a strategy and reruns it gets identical numbers, (b) the
        # community "Live" score matches a redeploy, (c) "same code, different
        # result" support reports go away. Cost: single-threaded XGB (a few
        # seconds slower on large windows; hist is fast so it's minor). FORCED
        # (not setdefault) so the guarantee can't be silently broken by a
        # strategy passing n_jobs. Exact reproducibility holds within the
        # platform (pinned versions / same Modal image); a user's own machine
        # with different xgboost/numpy/CPU can still differ in low-order bits.
        model_params["n_jobs"] = 1

    # Common defaults
    model_params.setdefault("random_state", 42)

    from model_wrapper import ModelWrapper

    clf = cls(**model_params)
    clf.fit(X_train, y_train_enc)

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    return ModelWrapper(clf, original_classes=enc.classes_, n_features=X_train.shape[1])


def _generate_signals(model, X, threshold):
    """Framework-owned signal generation. Deterministic threshold logic."""
    proba = model.predict_proba(X)
    classes = list(model.classes_)

    idx_pos = classes.index(1) if 1 in classes else None
    idx_neg = classes.index(-1) if -1 in classes else None

    p_pos = proba[:, idx_pos] if idx_pos is not None else np.zeros(len(X))
    p_neg = proba[:, idx_neg] if idx_neg is not None else np.zeros(len(X))

    signal_vals = np.zeros(len(X))
    signal_vals = np.where(p_pos >= threshold, 1.0, signal_vals)
    signal_vals = np.where(p_neg >= threshold, -1.0, signal_vals)
    # Both exceed: pick stronger
    both = (p_pos >= threshold) & (p_neg >= threshold)
    signal_vals[both] = np.where(p_pos[both] >= p_neg[both], 1.0, -1.0)

    return pd.Series(signal_vals, index=X.index), p_pos, p_neg


# ── Filter functions (all no-ops when config value is None) ──────────────

def _apply_direction_filter(signal, direction):
    """Zero out signals that don't match allowed direction."""
    if direction is None or direction == "both":
        return signal
    s = signal.copy()
    if direction == "long":
        s[s < 0] = 0.0
    elif direction == "short":
        s[s > 0] = 0.0
    return s


def _apply_session_filter(signal, index, session_hours):
    """Zero out signals outside session hours [start, end] UTC."""
    if session_hours is None:
        return signal
    s = signal.copy()
    start_h, end_h = session_hours[0], session_hours[1]
    hours = index.hour
    if start_h <= end_h:
        mask = (hours >= start_h) & (hours < end_h)
    else:  # wrap around midnight, e.g. [22, 6]
        mask = (hours >= start_h) | (hours < end_h)
    s[~mask] = 0.0
    return s


def _apply_atr_filter(signal, close, high, low, min_atr):
    """Zero out signals when NATR(14) is below threshold."""
    if min_atr is None:
        return signal
    hl = high - low
    hc = (high - close.shift(1)).abs()
    lc = (low - close.shift(1)).abs()
    tr = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr14 = tr.ewm(com=13, adjust=False).mean()
    natr = atr14 / close.replace(0, np.nan)
    s = signal.copy()
    s[natr < min_atr] = 0.0
    return s


def _apply_trend_filter(signal, close, trend_filter):
    """Only allow signals aligned with trend. e.g. 'sma_50': longs above SMA, shorts below."""
    if trend_filter is None:
        return signal
    # Parse: "sma_50" → SMA with period 50
    parts = trend_filter.lower().replace("-", "_").split("_")
    if len(parts) >= 2 and parts[0] in ("sma", "ema"):
        period = int(parts[1])
    else:
        return signal  # unknown filter, skip
    if parts[0] == "sma":
        trend_line = close.rolling(period).mean()
    else:
        trend_line = close.ewm(span=period, adjust=False).mean()
    s = signal.copy()
    # Longs only above trend, shorts only below
    s[(s > 0) & (close < trend_line)] = 0.0
    s[(s < 0) & (close > trend_line)] = 0.0
    return s


# ── run_backtest_v2: framework-owned SL/TP/cooldown/position management ──

def run_backtest_v2(signal, close, high, low, config, capital=10000, cost=2e-5):
    """Backtest with SL/TP/cooldown/direction handling built into the engine.
    Unlike run_backtest (v1), this function handles position exits internally.

    Returns: same dict shape as run_backtest()
    """
    stop_loss = config.get("stop_loss")
    take_profit = config.get("take_profit")
    cooldown = config.get("cooldown", 0)
    on_opposite = config.get("on_opposite", "reverse")

    sig_arr = signal.values
    close_arr = close.values
    high_arr = high.values
    low_arr = low.values
    idx = signal.index
    n = len(close_arr)

    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    equity_vals = np.full(n, float(capital))
    cumret = 1.0

    position = 0.0       # current direction: 1.0, -1.0, or 0.0 (flat)
    entry_price = None
    entry_bar = None      # index into arrays for entry time
    cooldown_remaining = 0

    def _log_trade(exit_bar, exit_px, ret, reason):
        trade_log.append({
            "type": "Buy" if position == 1.0 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[exit_bar]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(exit_px, 5),
            "pnl": round(position * (exit_px - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": reason,
        })

    for i in range(n):
        c = close_arr[i]
        h = high_arr[i]
        lo = low_arr[i]
        s = sig_arr[i]

        # 1. Check SL/TP if in trade
        if position != 0.0 and entry_price is not None:
            hit_sl = False
            hit_tp = False
            exit_price = None

            if position == 1.0:  # long
                if stop_loss is not None and lo <= entry_price * (1 - stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 - stop_loss)
                elif take_profit is not None and h >= entry_price * (1 + take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 + take_profit)
            else:  # short
                if stop_loss is not None and h >= entry_price * (1 + stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 + stop_loss)
                elif take_profit is not None and lo <= entry_price * (1 - take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 - take_profit)

            if hit_sl or hit_tp:
                ret = float(position * (exit_price - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if position == 1.0:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                _log_trade(i, exit_price, ret, "SL" if hit_sl else "TP")
                cumret *= (1 + ret)
                position = 0.0
                entry_price = None
                entry_bar = None
                cooldown_remaining = cooldown
                equity_vals[i] = capital * cumret
                continue

        # 2. Cooldown
        if cooldown_remaining > 0:
            cooldown_remaining -= 1
            equity_vals[i] = capital * cumret
            continue

        # 3. Signal processing
        if s != 0.0:
            if position == 0.0:
                # Open new trade
                position = s
                entry_price = c
                entry_bar = i
            elif s != position:
                # Opposite signal
                if on_opposite == "reverse":
                    # Close current + open opposite
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "signal")
                    cumret *= (1 + ret)
                    position = s
                    entry_price = c
                    entry_bar = i
                else:  # close_only
                    # Close current, go flat
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "close_only")
                    cumret *= (1 + ret)
                    position = 0.0
                    entry_price = None
                    entry_bar = None
                    cooldown_remaining = cooldown

        equity_vals[i] = capital * cumret

    # Close last open trade at final close
    if position != 0.0 and entry_price is not None and n > 0 and entry_price != 0:
        c = close_arr[-1]
        ret = float(position * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if position == 1.0:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        _log_trade(n - 1, c, ret, "end")
        cumret *= (1 + ret)
        equity_vals[-1] = capital * cumret

    # Bar returns for Sharpe (approximate)
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if close_arr[i - 1] != 0 and sig_arr[i - 1] != 0:
            bar_returns[i] = sig_arr[i - 1] * (close_arr[i] - close_arr[i - 1]) / close_arr[i - 1]

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


# ── run_strategy: the v2 orchestrator ────────────────────────────────────

def run_strategy(feature_fn, config_fn, data_path, start_date="", end_date="",
                 validation_date="", train_split=0.7, register_model_fn=None):
    """Config-driven strategy execution. Claude writes feature_fn + config_fn,
    framework does everything else.

    Returns: results dict (same format as webapp expects)
    """
    config = config_fn()

    # Auto-correct SL/TP if Claude passed percentage instead of decimal
    for _key in ("stop_loss", "take_profit"):
        _val = config.get(_key)
        if _val is not None and _val > 0.1:  # >10% is almost certainly a percentage
            config[_key] = _val / 100.0
            print(f"[strategy] Auto-corrected {_key}: {_val} -> {config[_key]} (was percentage, converted to decimal)")

    # 1. Load data
    df, close, open_, high, low = load_ohlc(data_path, start_date, end_date)

    # 2. Feature engineering (Claude's function)
    df = feature_fn(df, close, open_, high, low)
    close = df["close"]
    open_ = df["open"]
    high = df["high"]
    low = df["low"]

    # 3. Warm-up detection: drop rows where features have NaN BEFORE any fill
    feature_cols = [c for c in df.columns if c not in ("open", "high", "low", "close")]
    raw_nans = df[feature_cols].isna().any(axis=1)
    valid_rows = ~raw_nans
    if valid_rows.any():
        first_valid = valid_rows.idxmax()
        if raw_nans.loc[:first_valid].any():
            df = df.loc[first_valid:].copy()
            close = df["close"]
            open_ = df["open"]
            high = df["high"]
            low = df["low"]

    # 4. Target
    horizon = config.get("target_horizon", 4)
    target = make_target(close, horizon=horizon)

    # 5. Split (ffill only within each partition — no bfill leak)
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]
    close = df["close"]
    high = df["high"]
    low = df["low"]

    X = df[feature_cols].copy()
    X = X.replace([np.inf, -np.inf], np.nan)

    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)
    split_idx = max(1, min(split_idx, len(df) - 1))

    # ffill within train and test separately (no leak)
    X_train = X.iloc[:split_idx].ffill().fillna(0.0)
    X_test = X.iloc[split_idx:].ffill().fillna(0.0)
    X = pd.concat([X_train, X_test])

    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]
    high_test = high.iloc[split_idx:]
    low_test = low.iloc[split_idx:]

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    split_dt = str(df.index[split_idx])

    sp = {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }

    # 6. Build model from config
    model = _build_model_from_config(config, X_train, y_train_enc)

    # 7. Generate signals
    threshold = config.get("signal_threshold", 0.55)
    signal_train, p_pos_train, p_neg_train = _generate_signals(model, X_train, threshold)
    signal_test, p_pos_test, p_neg_test = _generate_signals(model, X_test, threshold)

    # 8. Apply filters (order: direction → session → ATR → trend)
    direction = config.get("direction", "both")
    signal_test = _apply_direction_filter(signal_test, direction)
    signal_train = _apply_direction_filter(signal_train, direction)

    session_filter = config.get("session_filter")
    signal_test = _apply_session_filter(signal_test, signal_test.index, session_filter)
    signal_train = _apply_session_filter(signal_train, signal_train.index, session_filter)

    min_atr = config.get("min_atr")
    if min_atr is not None:
        signal_test = _apply_atr_filter(signal_test, close_test, high_test, low_test, min_atr)

    trend_filter = config.get("trend_filter")
    if trend_filter is not None:
        signal_test = _apply_trend_filter(signal_test, close_test, trend_filter)

    signal_full = pd.concat([signal_train, signal_test])

    # 9. Backtest with SL/TP/cooldown (test + train)
    high_train = high.iloc[:split_idx]
    low_train = low.iloc[:split_idx]
    has_risk = (config.get("stop_loss") is not None or
                config.get("take_profit") is not None or
                config.get("cooldown", 0) > 0 or
                config.get("on_opposite", "reverse") != "reverse")

    if has_risk:
        bt = run_backtest_v2(signal_test, close_test, high_test, low_test, config, capital=10000)
        bt_train = run_backtest_v2(signal_train, close_train, high_train, low_train, config, capital=10000)
    else:
        bt = run_backtest(signal_test, close_test, capital=10000)
        bt_train = run_backtest(signal_train, close_train, capital=10000)

    # 10. Metrics
    metrics = compute_metrics(bt, close_test, capital=10000)

    # 11. Pre-compute all trade stats (single source of truth)
    pre_stats = {
        "train_stats": compute_trade_stats(bt_train.get("trade_returns", []), capital=10000),
        "test_stats":  compute_trade_stats(bt.get("trade_returns", []), capital=10000),
        "long_stats":  compute_trade_stats(bt.get("long_returns", []), capital=10000),
        "short_stats": compute_trade_stats(bt.get("short_returns", []), capital=10000),
    }

    # 12. Register model
    if register_model_fn is not None:
        register_model_fn(model)

    # 13. Build return dict
    return build_return_dict(sp, bt, metrics, model, feature_cols,
                             signal_full, p_pos_test, p_neg_test, custom_figs=[],
                             bt_train_result=bt_train, pre_stats=pre_stats)

# ── End strategy_utils ──


DATA_PATH = '/root/Desktop/QuantifyMe/data/ohlc/GBPUSD_15min.parquet'
START_DATE = '2026-04-15'
END_DATE = '2026-05-25'
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── EMA crossover core signals ──────────────────────────────────────────
    ema9  = close.ewm(span=9,  adjust=False).mean()
    ema21 = close.ewm(span=21, adjust=False).mean()
    ema50 = close.ewm(span=50, adjust=False).mean()
    ema200 = close.ewm(span=200, adjust=False).mean()

    df["ema9"]  = ema9
    df["ema21"] = ema21
    df["ema50"] = ema50
    df["ema200"] = ema200

    # Raw spread and normalised spread
    df["ema_diff"]       = ema9 - ema21
    df["ema_diff_norm"]  = (ema9 - ema21) / close

    # Cross signal: +1 when ema9 > ema21, -1 otherwise
    df["ema_cross_sign"] = np.where(ema9 > ema21, 1.0, -1.0)

    # Momentum of the spread (rate of change of spread)
    df["ema_diff_roc1"]  = df["ema_diff"].diff(1)
    df["ema_diff_roc3"]  = df["ema_diff"].diff(3)

    # Distance of price from ema50 and ema200 (normalised)
    df["dist_ema50"]  = (close - ema50)  / close
    df["dist_ema200"] = (close - ema200) / close

    # ── RSI (14) ────────────────────────────────────────────────────────────
    delta  = close.diff()
    gain   = delta.clip(lower=0.0)
    loss   = (-delta).clip(lower=0.0)
    avg_g  = gain.ewm(com=13, adjust=False).mean()
    avg_l  = loss.ewm(com=13, adjust=False).mean()
    rs     = avg_g / avg_l.replace(0.0, np.nan)
    rsi14  = 100.0 - 100.0 / (1.0 + rs)
    df["rsi14"] = rsi14

    # RSI normalised and centred
    df["rsi14_norm"] = (rsi14 - 50.0) / 50.0

    # ── MACD ────────────────────────────────────────────────────────────────
    ema12      = close.ewm(span=12, adjust=False).mean()
    ema26      = close.ewm(span=26, adjust=False).mean()
    macd_line  = ema12 - ema26
    signal_ln  = macd_line.ewm(span=9, adjust=False).mean()
    macd_hist  = macd_line - signal_ln

    df["macd_line"]    = macd_line / close
    df["macd_signal"]  = signal_ln / close
    df["macd_hist"]    = macd_hist / close
    df["macd_cross"]   = np.where(macd_line > signal_ln, 1.0, -1.0)

    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_mid   = close.rolling(20).mean()
    bb_std   = close.rolling(20).std(ddof=0)
    bb_upper = bb_mid + 2.0 * bb_std
    bb_lower = bb_mid - 2.0 * bb_std
    bb_width = (bb_upper - bb_lower) / bb_mid.replace(0.0, np.nan)
    bb_pct   = (close - bb_lower) / (bb_upper - bb_lower).replace(0.0, np.nan)

    df["bb_width"] = bb_width
    df["bb_pct"]   = bb_pct

    # ── ATR (14) ─────────────────────────────────────────────────────────────
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    atr14 = tr.ewm(com=13, adjust=False).mean()
    df["atr14"]      = atr14
    df["natr14"]     = atr14 / close          # normalised ATR (volatility proxy)

    # ── Stochastic %K / %D (14, 3) ──────────────────────────────────────────
    low14  = low.rolling(14).min()
    high14 = high.rolling(14).max()
    stoch_k = 100.0 * (close - low14) / (high14 - low14).replace(0.0, np.nan)
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"]      = stoch_k / 100.0
    df["stoch_d"]      = stoch_d / 100.0
    df["stoch_diff"]   = (stoch_k - stoch_d) / 100.0

    # ── Rate of Change ───────────────────────────────────────────────────────
    df["roc1"]  = close.pct_change(1)
    df["roc4"]  = close.pct_change(4)
    df["roc8"]  = close.pct_change(8)
    df["roc16"] = close.pct_change(16)

    # ── Candle features ──────────────────────────────────────────────────────
    body       = (close - open_).abs()
    candle_rng = (high - low).replace(0.0, np.nan)
    df["body_ratio"]   = body / candle_rng
    df["upper_wick"]   = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_rng
    df["lower_wick"]   = (pd.concat([close, open_], axis=1).min(axis=1) - low) / candle_rng
    df["candle_dir"]   = np.where(close >= open_, 1.0, -1.0)

    # ── Volume-like proxy: range relative to rolling average ────────────────
    df["range_ratio"]  = candle_rng / candle_rng.rolling(20).mean()

    # ── Lagged EMA diff features ─────────────────────────────────────────────
    for lag in [1, 2, 3, 4]:
        df[f"ema_diff_lag{lag}"] = df["ema_diff_norm"].shift(lag)

    # ── Lagged RSI ───────────────────────────────────────────────────────────
    for lag in [1, 2, 4]:
        df[f"rsi14_lag{lag}"] = df["rsi14_norm"].shift(lag)

    # ── Rolling volatility (std of returns) ──────────────────────────────────
    ret = close.pct_change()
    df["vol_8"]  = ret.rolling(8).std()
    df["vol_16"] = ret.rolling(16).std()
    df["vol_32"] = ret.rolling(32).std()

    # ── Trend strength: ADX-like (simplified) ────────────────────────────────
    plus_dm  = (high.diff()).clip(lower=0.0)
    minus_dm = (-low.diff()).clip(lower=0.0)
    overlap  = pd.concat([plus_dm, minus_dm], axis=1).min(axis=1)
    plus_dm  = plus_dm - overlap
    minus_dm = minus_dm - overlap
    smooth_tr  = tr.ewm(com=13, adjust=False).mean()
    plus_di    = 100.0 * plus_dm.ewm(com=13, adjust=False).mean()  / smooth_tr.replace(0.0, np.nan)
    minus_di   = 100.0 * minus_dm.ewm(com=13, adjust=False).mean() / smooth_tr.replace(0.0, np.nan)
    di_sum     = (plus_di + minus_di).replace(0.0, np.nan)
    adx        = ((plus_di - minus_di).abs() / di_sum * 100.0).ewm(com=13, adjust=False).mean()
    df["adx"]       = adx / 100.0
    df["plus_di"]   = plus_di / 100.0
    df["minus_di"]  = minus_di / 100.0

    # ── Session hour (UTC) ───────────────────────────────────────────────────
    if hasattr(df.index, "hour"):
        df["hour_sin"] = np.sin(2.0 * np.pi * df.index.hour / 24.0)
        df["hour_cos"] = np.cos(2.0 * np.pi * df.index.hour / 24.0)
    else:
        df["hour_sin"] = 0.0
        df["hour_cos"] = 1.0

    # ── Fill any NaN from warm-up periods ────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EMA 9/21 Crossover + MACD Momentum (XGBoost)",
        "model_type":       "XGBClassifier",
        "model_params":     {
            "n_estimators":     400,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.80,
            "colsample_bytree": 0.75,
            "min_child_weight": 3,
            "gamma":            0.10,
            "reg_alpha":        0.05,
            "reg_lambda":       1.50,
            "objective":        "binary:logistic",
            "tree_method":      "hist",
            "random_state":     42,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.0030,
        "take_profit":      0.0060,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 20],
        "min_atr":          None,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective":        (
            "Maximize Sharpe ratio on EUR/USD 15-min data. "
            "Core signal: EMA(9) vs EMA(21) crossover enriched with MACD, RSI, "
            "Bollinger %B, Stochastic, ATR, ADX, candle structure and rolling "
            "volatility. XGBoost with moderate depth (4) and strong regularisation "
            "(gamma, alpha, lambda) prevents overfitting on ~6 weeks of intraday data. "
            "A 0.55 probability threshold filters low-confidence signals. "
            "A 2:1 TP:SL ratio (30 bp SL / 60 bp TP) improves the reward-risk "
            "balance. Session filter [6,20] UTC keeps the model away from the thin "
            "Asian pre-open. trend_filter sma_50 aligns entries with the prevailing "
            "short-term trend to reduce chop. Cooldown=0 and reverse-on-opposite "
            "allow continuous participation in trending EMA crossover moves."
        ),
        "notes":            (
            "round-trip cost 2e-5 is accounted for by the framework. "
            "target_horizon=4 bars (1 hour ahead) suits EMA crossover which "
            "generates medium-frequency signals rather than tick-level scalps. "
            "All features are normalised or expressed as ratios to minimise "
            "scale sensitivity for the logistic-objective XGBoost."
        ),
    }

# ── Framework v2: auto-generated wrapper ──
def train_and_backtest():
    _vd = VALIDATION_DATE if 'VALIDATION_DATE' in globals() else ''
    _ts = TRAIN_SPLIT if 'TRAIN_SPLIT' in globals() else 0.7
    return run_strategy(
        feature_engineering, strategy_config,
        DATA_PATH, START_DATE, END_DATE,
        _vd, _ts,
        register_model_fn=register_model
    )

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:32:18
# Model            : XGBoost
# Feature Eng.     : SMA (20,50,200), BB (20,2.0), RSI 14, MACD (12,26,9), ATR 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# AUDUSD 15-min XGBoost Strategy
# SMA + RSI + MACD + Bollinger Bands + ATR Feature Set
# Optimized for Risk-Adjusted Return
# ============================================================

# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/AUDUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):

    # ── SMA 20, 50, 200 + distance from close ──────────────────────────────
    for period in [20, 50, 200]:
        sma = close.rolling(period).mean()
        df[f"sma_{period}"] = sma
        df[f"dm_sma_{period}"] = (close - sma) / sma

    # ── Bollinger Bands (20, 2.0) ───────────────────────────────────────────
    bb_mid   = close.rolling(20).mean()
    bb_std   = close.rolling(20).std(ddof=0)
    bb_upper = bb_mid + 2.0 * bb_std
    bb_lower = bb_mid - 2.0 * bb_std
    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid
    bb_range       = bb_upper - bb_lower
    df["bb_pct"]   = (close - bb_lower) / bb_range

    # ── RSI 14 ──────────────────────────────────────────────────────────────
    delta  = close.diff()
    gain   = delta.clip(lower=0)
    loss   = (-delta).clip(lower=0)
    avg_gain = gain.ewm(com=13, min_periods=14).mean()
    avg_loss = loss.ewm(com=13, min_periods=14).mean()
    rs         = avg_gain / avg_loss.replace(0, np.nan)
    df["rsi_14"] = 100.0 - (100.0 / (1.0 + rs))

    # ── MACD (12, 26, 9) ────────────────────────────────────────────────────
    ema_12     = close.ewm(span=12, adjust=False).mean()
    ema_26     = close.ewm(span=26, adjust=False).mean()
    macd_line  = ema_12 - ema_26
    signal_line = macd_line.ewm(span=9, adjust=False).mean()
    df["macd_line"]   = macd_line
    df["macd_signal"] = signal_line
    df["macd_hist"]   = macd_line - signal_line

    # ── ATR 14 + Normalised ATR ─────────────────────────────────────────────
    prev_close = close.shift(1)
    tr = pd.concat([
        high - low,
        (high - prev_close).abs(),
        (low  - prev_close).abs()
    ], axis=1).max(axis=1)
    atr = tr.ewm(com=13, min_periods=14).mean()
    df["atr_14"] = atr
    df["natr"]   = atr / close

    # ── Price momentum / rate-of-change ────────────────────────────────────
    for n in [1, 4, 8, 16]:
        df[f"roc_{n}"] = close.pct_change(n)

    # ── Candle body & wick features ─────────────────────────────────────────
    body        = (close - open_).abs()
    candle_range = (high - low).replace(0, np.nan)
    df["body_ratio"]  = body / candle_range
    df["upper_wick"]  = (high  - pd.concat([close, open_], axis=1).max(axis=1)) / candle_range
    df["lower_wick"]  = (pd.concat([close, open_], axis=1).min(axis=1) - low)   / candle_range
    df["candle_dir"]  = np.where(close >= open_, 1.0, -1.0)

    # ── Volume (if present) ─────────────────────────────────────────────────
    if "volume" in df.columns:
        vol_ma = df["volume"].rolling(20).mean()
        df["vol_ratio"] = df["volume"] / vol_ma.replace(0, np.nan)

    # ── Lagged RSI & MACD histogram ─────────────────────────────────────────
    for lag in [1, 2, 3]:
        df[f"rsi_14_lag{lag}"]   = df["rsi_14"].shift(lag)
        df[f"macd_hist_lag{lag}"] = df["macd_hist"].shift(lag)

    # ── RSI overbought / oversold zones ─────────────────────────────────────
    df["rsi_ob"]     = np.where(df["rsi_14"] > 70, 1.0, 0.0)
    df["rsi_os"]     = np.where(df["rsi_14"] < 30, 1.0, 0.0)
    df["rsi_mid_up"] = np.where((df["rsi_14"] > 50) & (df["rsi_14"] <= 70), 1.0, 0.0)
    df["rsi_mid_dn"] = np.where((df["rsi_14"] >= 30) & (df["rsi_14"] < 50), 1.0, 0.0)

    # ── MACD cross signals ───────────────────────────────────────────────────
    df["macd_cross_up"] = np.where(
        (df["macd_line"] > df["macd_signal"]) &
        (df["macd_line"].shift(1) <= df["macd_signal"].shift(1)),
        1.0, 0.0
    )
    df["macd_cross_dn"] = np.where(
        (df["macd_line"] < df["macd_signal"]) &
        (df["macd_line"].shift(1) >= df["macd_signal"].shift(1)),
        1.0, 0.0
    )

    # ── Price position relative to SMA alignment ────────────────────────────
    df["trend_aligned_bull"] = np.where(
        (close > df["sma_20"]) & (df["sma_20"] > df["sma_50"]) & (df["sma_50"] > df["sma_200"]),
        1.0, 0.0
    )
    df["trend_aligned_bear"] = np.where(
        (close < df["sma_20"]) & (df["sma_20"] < df["sma_50"]) & (df["sma_50"] < df["sma_200"]),
        1.0, 0.0
    )

    # ── Bollinger Band squeeze (low volatility) ──────────────────────────────
    bb_width_ma = df["bb_width"].rolling(20).mean()
    df["bb_squeeze"] = np.where(df["bb_width"] < bb_width_ma, 1.0, 0.0)

    # ── Rolling close statistics ─────────────────────────────────────────────
    df["close_zscore_20"] = (close - close.rolling(20).mean()) / close.rolling(20).std(ddof=0)
    df["close_zscore_50"] = (close - close.rolling(50).mean()) / close.rolling(50).std(ddof=0)

    # ── Volatility regime ────────────────────────────────────────────────────
    natr_ma = df["natr"].rolling(20).mean()
    df["vol_regime_high"] = np.where(df["natr"] > natr_ma, 1.0, 0.0)

    # ── Fill NaN from indicator warm-up ─────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "AUD/USD XGBoost SMA+RSI+MACD+BB Momentum",
        "model_type":       "XGBClassifier",
        "model_params":     {
            "n_estimators":       500,
            "max_depth":          4,
            "learning_rate":      0.04,
            "subsample":          0.75,
            "colsample_bytree":   0.70,
            "min_child_weight":   5,
            "gamma":              0.1,
            "reg_alpha":          0.1,
            "reg_lambda":         1.5,
            "scale_pos_weight":   1.0,
            "objective":          "binary:logistic",
            "tree_method":        "hist",
            "random_state":       42,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [0, 23],
        "min_atr":          None,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective":        (
            "Maximize risk-adjusted return (Sharpe/Calmar) on AUD/USD 15-min. "
            "XGBoost with depth-4 trees and conservative regularization (reg_lambda=1.5, "
            "min_child_weight=5) to reduce overfitting on FX data. "
            "2:1 RR (SL=0.5%, TP=1.0%) ensures positive expectancy with ~40%+ win rate. "
            "Subsample + colsample add stochastic diversity. 500 estimators with lr=0.04 "
            "balances bias-variance. Threshold 0.55 filters marginal signals."
        ),
        "notes":            (
            "Features: SMA(20/50/200) with distances, Bollinger Bands width+pct, RSI-14 "
            "with zone flags, MACD histogram + crosses, ATR-14 + NATR, momentum ROC(1/4/8/16), "
            "candle body/wick ratios, trend alignment flags, BB squeeze, z-scores, vol regime. "
            "Reverse on opposite signal to capture trend reversals. Session filter disabled "
            "to capture AUD/USD Asian + London + NY sessions. Target horizon = 4 bars (1 hour)."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:21:53
# Model            : Gradient Boosting
# Feature Eng.     : EMA (9,21), RSI 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/AUDUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # --- EMA 9 and EMA 21 (required) ---
    ema_9 = close.ewm(span=9, adjust=False).mean()
    ema_21 = close.ewm(span=21, adjust=False).mean()

    df["ema_9"] = ema_9
    df["ema_21"] = ema_21
    df["dm_ema_9"] = (close - ema_9) / ema_9
    df["dm_ema_21"] = (close - ema_21) / ema_21

    # EMA crossover signal and spread
    df["ema_cross"] = ema_9 - ema_21
    df["ema_cross_prev"] = df["ema_cross"].shift(1)
    df["ema_cross_sign"] = np.sign(df["ema_cross"])
    df["ema_cross_change"] = df["ema_cross_sign"] - np.sign(df["ema_cross_prev"])

    # --- RSI 14 (required) ---
    delta = close.diff()
    gain = delta.clip(lower=0)
    loss = -delta.clip(upper=0)
    avg_gain = gain.ewm(com=13, adjust=False).mean()
    avg_loss = loss.ewm(com=13, adjust=False).mean()
    rs = avg_gain / avg_loss.replace(0, np.nan)
    rsi_14 = 100 - (100 / (1 + rs))
    df["rsi_14"] = rsi_14

    # RSI derived features
    df["rsi_14_norm"] = (rsi_14 - 50) / 50
    df["rsi_overbought"] = np.where(rsi_14 > 70, 1, 0)
    df["rsi_oversold"] = np.where(rsi_14 < 30, 1, 0)
    df["rsi_mid_cross"] = np.where(rsi_14 > 50, 1, -1)

    # --- Additional EMAs for context ---
    ema_50 = close.ewm(span=50, adjust=False).mean()
    ema_200 = close.ewm(span=200, adjust=False).mean()
    df["ema_50"] = ema_50
    df["ema_200"] = ema_200
    df["dm_ema_50"] = (close - ema_50) / ema_50
    df["dm_ema_200"] = (close - ema_200) / ema_200
    df["ema_50_200_spread"] = (ema_50 - ema_200) / ema_200

    # --- ATR (14 periods) ---
    tr1 = high - low
    tr2 = (high - close.shift(1)).abs()
    tr3 = (low - close.shift(1)).abs()
    true_range = pd.concat([tr1, tr2, tr3], axis=1).max(axis=1)
    atr_14 = true_range.ewm(com=13, adjust=False).mean()
    df["atr_14"] = atr_14
    df["natr_14"] = atr_14 / close

    # --- Bollinger Bands (20, 2) ---
    sma_20 = close.rolling(20).mean()
    std_20 = close.rolling(20).std()
    bb_upper = sma_20 + 2 * std_20
    bb_lower = sma_20 - 2 * std_20
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower
    df["bb_pct_b"] = (close - bb_lower) / (bb_upper - bb_lower).replace(0, np.nan)
    df["bb_width"] = (bb_upper - bb_lower) / sma_20

    # --- MACD (12, 26, 9) ---
    ema_12 = close.ewm(span=12, adjust=False).mean()
    ema_26 = close.ewm(span=26, adjust=False).mean()
    macd_line = ema_12 - ema_26
    macd_signal = macd_line.ewm(span=9, adjust=False).mean()
    macd_hist = macd_line - macd_signal
    df["macd_line"] = macd_line
    df["macd_signal_line"] = macd_signal
    df["macd_hist"] = macd_hist
    df["macd_hist_sign"] = np.sign(macd_hist)
    df["macd_hist_change"] = np.sign(macd_hist) - np.sign(macd_hist.shift(1))

    # --- Momentum & Rate of Change ---
    df["mom_5"] = close.pct_change(5)
    df["mom_10"] = close.pct_change(10)
    df["mom_20"] = close.pct_change(20)
    df["roc_3"] = close.pct_change(3)

    # --- Candlestick features ---
    df["body"] = (close - open_) / atr_14
    df["upper_wick"] = (high - pd.concat([close, open_], axis=1).max(axis=1)) / atr_14
    df["lower_wick"] = (pd.concat([close, open_], axis=1).min(axis=1) - low) / atr_14
    df["bar_range"] = (high - low) / atr_14

    # --- Volume-proxy: price range relative momentum ---
    df["high_low_ratio"] = (high - low) / close

    # --- Stochastic Oscillator (14, 3) ---
    lowest_low = low.rolling(14).min()
    highest_high = high.rolling(14).max()
    stoch_k = 100 * (close - lowest_low) / (highest_high - lowest_low).replace(0, np.nan)
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"] = stoch_k
    df["stoch_d"] = stoch_d
    df["stoch_kd_diff"] = stoch_k - stoch_d

    # --- Rolling volatility ---
    df["vol_10"] = close.pct_change().rolling(10).std()
    df["vol_20"] = close.pct_change().rolling(20).std()
    df["vol_ratio"] = df["vol_10"] / df["vol_20"].replace(0, np.nan)

    # --- Lagged RSI and EMA cross (for sequential signal detection) ---
    df["rsi_14_lag1"] = rsi_14.shift(1)
    df["rsi_14_lag2"] = rsi_14.shift(2)
    df["ema_cross_lag1"] = df["ema_cross"].shift(1)
    df["ema_cross_lag2"] = df["ema_cross"].shift(2)

    # --- Trend alignment: both EMAs agree ---
    df["trend_aligned_bull"] = np.where((ema_9 > ema_21) & (ema_21 > ema_50), 1, 0)
    df["trend_aligned_bear"] = np.where((ema_9 < ema_21) & (ema_21 < ema_50), 1, 0)

    # --- RSI momentum divergence proxy ---
    price_chg_5 = close.pct_change(5)
    rsi_chg_5 = rsi_14.diff(5)
    df["rsi_price_div"] = np.where(
        (price_chg_5 > 0) & (rsi_chg_5 < 0), -1,
        np.where((price_chg_5 < 0) & (rsi_chg_5 > 0), 1, 0)
    )

    # --- SMA 50 distance (for trend_filter compatibility) ---
    df["sma_50"] = close.rolling(50).mean()

    df = df.bfill().ffill()
    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title": "AUD/USD EMA Cross RSI Gradient Boost Scalper",
        "model_type": "GradientBoostingClassifier",
        "model_params": {
            "n_estimators": 400,
            "max_depth": 4,
            "learning_rate": 0.04,
            "subsample": 0.8,
            "max_features": "sqrt",
            "min_samples_leaf": 20,
            "min_samples_split": 40,
            "validation_fraction": 0.1,
            "n_iter_no_change": 30,
            "tol": 1e-4,
        },
        "signal_threshold": 0.55,
        "direction": "both",
        "stop_loss": 0.005,
        "take_profit": 0.010,
        "cooldown": 0,
        "max_positions": 1,
        "on_opposite": "reverse",
        "session_filter": [0, 23],
        "min_atr": None,
        "trend_filter": None,
        "target_horizon": 4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe) on AUD/USD 15-min bars. "
            "GradientBoostingClassifier with shrinkage (lr=0.04), moderate depth (4), "
            "subsampling (0.8) and sqrt feature fraction controls overfitting on a noisy FX "
            "series. Early stopping (n_iter_no_change=30) prevents over-training. "
            "SL=0.5%/TP=1.0% gives 1:2 RR. Threshold=0.55 filters low-confidence signals. "
            "EMA 9/21 crossover with RSI 14 confirmation is the primary signal logic, "
            "reinforced by MACD, Bollinger Bands, Stochastic, and multi-period momentum."
        ),
        "notes": (
            "Features: EMA 9, 21, 50, 200 distances; RSI 14 with overbought/oversold flags; "
            "MACD histogram; Bollinger %B and width; Stochastic K/D; ATR-normalized candle "
            "body/wicks; 5/10/20-bar momentum; rolling volatility ratio; trend alignment flags; "
            "RSI-price divergence proxy. Target horizon 4 bars (1 hour ahead). "
            "All features are lagged or rolling — no lookahead bias."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:40:19
# Model            : XGBoost
# Feature Eng.     : EMA (9,21), RSI 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/AUDUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── EMA 9 and EMA 21 ──────────────────────────────────────────────────
    ema_9  = close.ewm(span=9,  adjust=False).mean()
    ema_21 = close.ewm(span=21, adjust=False).mean()

    df["ema_9"]    = ema_9
    df["ema_21"]   = ema_21
    df["dm_ema_9"]  = (close - ema_9)  / ema_9
    df["dm_ema_21"] = (close - ema_21) / ema_21

    # EMA crossover signal: positive when fast > slow
    df["ema_cross"]       = ema_9 - ema_21
    df["ema_cross_prev"]  = df["ema_cross"].shift(1)
    # Binary: did a cross just occur?
    df["ema_cross_up"]    = np.where((df["ema_cross"] > 0) & (df["ema_cross_prev"] <= 0), 1, 0)
    df["ema_cross_down"]  = np.where((df["ema_cross"] < 0) & (df["ema_cross_prev"] >= 0), 1, 0)
    # Trend direction encoded as -1 / 1
    df["ema_trend"]       = np.where(ema_9 > ema_21, 1, -1)

    # ── RSI 14 ────────────────────────────────────────────────────────────
    delta  = close.diff()
    gain   = delta.clip(lower=0)
    loss   = (-delta).clip(lower=0)
    avg_gain = gain.ewm(com=13, adjust=False).mean()
    avg_loss = loss.ewm(com=13, adjust=False).mean()
    rs     = avg_gain / avg_loss.replace(0, np.nan)
    rsi_14 = 100 - (100 / (1 + rs))
    df["rsi_14"] = rsi_14

    # RSI regime flags
    df["rsi_oversold"]   = np.where(rsi_14 < 30, 1, 0)
    df["rsi_overbought"] = np.where(rsi_14 > 70, 1, 0)
    df["rsi_mid"]        = rsi_14 - 50          # centred

    # RSI momentum (1-bar change in RSI)
    df["rsi_delta"]      = rsi_14.diff(1)
    df["rsi_delta2"]     = rsi_14.diff(3)

    # ── Additional momentum / volatility features ─────────────────────────
    # ATR-like normalised true range
    prev_close = close.shift(1)
    tr = pd.concat([
        high - low,
        (high - prev_close).abs(),
        (low  - prev_close).abs()
    ], axis=1).max(axis=1)
    atr_14 = tr.ewm(span=14, adjust=False).mean()
    df["atr_14"]  = atr_14
    df["natr_14"] = atr_14 / close          # normalised ATR

    # Rate-of-change over various horizons
    for n in [4, 8, 16]:
        df[f"roc_{n}"] = close.pct_change(n)

    # Bollinger Band width and %B (using 20-period SMA)
    sma_20   = close.rolling(20).mean()
    std_20   = close.rolling(20).std()
    bb_upper = sma_20 + 2 * std_20
    bb_lower = sma_20 - 2 * std_20
    df["bb_width"] = (bb_upper - bb_lower) / sma_20
    df["bb_pct"]   = (close - bb_lower) / (bb_upper - bb_lower).replace(0, np.nan)

    # Candle body and wick features
    df["body"]        = (close - open_) / close
    df["upper_wick"]  = (high - pd.concat([close, open_], axis=1).max(axis=1)) / close
    df["lower_wick"]  = (pd.concat([close, open_], axis=1).min(axis=1) - low)  / close

    # Volume-normalised momentum proxy: price range relative to ATR
    df["range_vs_atr"] = (high - low) / atr_14.replace(0, np.nan)

    # Lagged EMA cross signal
    df["ema_cross_lag1"] = df["ema_cross"].shift(1)
    df["ema_cross_lag2"] = df["ema_cross"].shift(2)

    # Combined signal: RSI and EMA cross alignment
    df["rsi_ema_bull"] = np.where((rsi_14 > 50) & (ema_9 > ema_21), 1, 0)
    df["rsi_ema_bear"] = np.where((rsi_14 < 50) & (ema_9 < ema_21), 1, 0)

    # Hour-of-day (cyclical encoding) — no lookahead
    hour = df.index.hour
    df["hour_sin"] = np.sin(2 * np.pi * hour / 24)
    df["hour_cos"] = np.cos(2 * np.pi * hour / 24)

    # Day-of-week (cyclical encoding)
    dow = df.index.dayofweek
    df["dow_sin"] = np.sin(2 * np.pi * dow / 5)
    df["dow_cos"] = np.cos(2 * np.pi * dow / 5)

    # Fill NaN from warm-up periods
    df = df.bfill().ffill()

    return df

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "AUD/USD EMA Cross (9/21) + RSI14 XGBoost Scalper",
        "model_type":       "XGBClassifier",
        "model_params":     {
            "n_estimators":     400,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.8,
            "colsample_bytree": 0.75,
            "min_child_weight": 3,
            "gamma":            0.1,
            "reg_alpha":        0.05,
            "reg_lambda":       1.5,
            "objective":        "binary:logistic",
            "tree_method":      "hist",
            "random_state":     42,
        },
        "signal_threshold": 0.54,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [0, 23],
        "min_atr":          None,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective":        (
            "Maximise risk-adjusted return on AUD/USD 15-min bars. "
            "XGBoost chosen for its ability to capture non-linear interactions between "
            "the EMA-cross regime, RSI momentum, volatility (NATR/BB width), and time-of-day. "
            "Shallow trees (max_depth=4) with strong regularisation (reg_lambda=1.5, gamma=0.1) "
            "reduce overfitting on the limited 1-year window. "
            "2:1 R:R (SL=0.5%, TP=1.0%) improves Sharpe; reverse on opposite signal captures "
            "trend momentum without missing transitions."
        ),
        "notes":            (
            "Features: EMA-9/21 cross and distances, RSI-14 with regime flags and delta, "
            "ATR-14, NATR, Bollinger Band width/%B, 4/8/16-bar ROC, candle anatomy, "
            "time cyclical encodings. Threshold 0.54 filters marginal signals to raise precision. "
            "No session filter applied — AUD/USD has meaningful moves across Asian and London sessions."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-25 02:29:29
# Model            : XGBoost
# Feature Eng.     : buy when price closes below the lower Bollinger Band(20,2) and RSI(14) < 35, exit at the middle band + Auto-add features: ON
# Signal / Entry   : —
# Optimization     : —
# Risk Mgmt        : —
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# Bollinger Band Mean-Reversion + RSI Filter (XGBoost, Sharpe)
# ============================================================

# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

# ── Inlined strategy_utils ──
"""
strategy_utils.py — Standard utility functions for generated strategies.
Claude imports these instead of writing boilerplate from scratch.
This ensures consistent behavior across all generated strategies.
"""
import numpy as np
import pandas as pd
from sklearn.preprocessing import LabelEncoder


# Max backtest window per timeframe. A finer timeframe over a longer window
# blows up the results dict / parquet load / Modal train time (the 2026-05-12
# OOM was a 1-min × multi-year sweep) — and a 1-min strategy gains nothing from
# 2 years of 1-min bars. Enforced HERE because every training path (UI / API /
# Modal) funnels through run_strategy → load_ohlc. Env-overridable so a future
# "max plan" / dedicated-server tier can lift it.
_TF_MAX_DAYS = {
    "1min":  30,
    "5min":  90,
    "15min": 365,
    "1h":    730,
}


def _fetch_ohlc_from_internal(symbol: str, tf: str, start: str, end: str):
    """Phase 3.2: fetch parquet bytes from Server A's /internal/ohlc endpoint
    instead of reading a local file. Used inside Modal containers / Mac worker
    pool (Phase 3.4) so every train sees the same source of truth as the chart.

    Returns: pd.DataFrame (parquet decoded), or raises on any failure so the
    caller can fall back / surface a clear error in the job.
    """
    import hashlib as _hashlib, hmac as _hmac, io as _io, os as _os
    import urllib.request as _ur, urllib.parse as _urp

    base = (_os.environ.get("QM_INTERNAL_OHLC_BASE") or "").rstrip("/")
    secret = (_os.environ.get("INTERNAL_WS_SECRET") or "").strip()
    if not base:
        raise RuntimeError("QM_INTERNAL_OHLC_BASE not set")
    if not secret:
        raise RuntimeError("INTERNAL_WS_SECRET not set")

    msg = f"{symbol}|{tf}|{start}|{end}".encode("utf-8")
    sig = _hmac.new(secret.encode("utf-8"), msg, _hashlib.sha256).hexdigest()

    qs = _urp.urlencode({
        "symbol": symbol, "tf": tf,
        "start": start, "end": end, "sig": sig,
    })
    url = f"{base}/internal/ohlc?{qs}"

    req = _ur.Request(url, headers={"User-Agent": "qm-worker/1.0"})
    with _ur.urlopen(req, timeout=30) as resp:
        if resp.status != 200:
            raise RuntimeError(f"/internal/ohlc returned {resp.status}")
        payload = resp.read()
    print(f"[load_ohlc:internal] {symbol} {tf} fetched {len(payload)} bytes", flush=True)
    return pd.read_parquet(_io.BytesIO(payload))


def _parse_symbol_tf_from_path(data_path: str):
    """Pull SYMBOL + TF out of a path like .../EURUSD_1min.parquet."""
    import os as _os, re as _re
    base = _os.path.basename(str(data_path))
    m = _re.match(r"^([A-Z]{6})_(\d+min|\d+h)\.parquet$", base)
    if not m:
        return None, None
    return m.group(1), m.group(2)


def load_ohlc(data_path, start_date="", end_date=""):
    """Load OHLC parquet, sort index, filter dates. Always returns consistent format.

    The lower bound is clamped per timeframe (see _TF_MAX_DAYS) — a request for
    more history than the cap silently starts later.

    Phase 3.2: when env QM_USE_INTERNAL_OHLC=="1", fetch over HTTP from
    Server A's /internal/ohlc endpoint instead of pd.read_parquet on a local
    file (which on Modal is a stale Volume snapshot). The endpoint applies the
    same day-cap, so the local cap-check below is a defensive no-op in that
    path. Flag defaults to "0" → unchanged behavior.

    Returns: (df, close, open_, high, low)
    """
    import os as _os, re as _re

    _use_internal = _os.environ.get("QM_USE_INTERNAL_OHLC", "0") == "1"
    if _use_internal:
        _sym, _tf = _parse_symbol_tf_from_path(data_path)
        if not _sym or not _tf:
            raise RuntimeError(
                f"QM_USE_INTERNAL_OHLC=1 but DATA_PATH basename does not match "
                f"SYMBOL_TF.parquet: {data_path}"
            )
        df = _fetch_ohlc_from_internal(_sym, _tf, start_date or "", end_date or "")
    else:
        df = pd.read_parquet(data_path)
    df.index = pd.to_datetime(df.index)
    df = df.sort_index()

    # Per-timeframe window cap (timeframe inferred from the parquet filename).
    _m = _re.search(r"_(\d+min|\d+h)\.parquet$", _os.path.basename(str(data_path)))
    _tf = _m.group(1) if _m else None
    _max_days = _TF_MAX_DAYS.get(_tf)
    if _max_days and _max_days > 0 and len(df):
        _env_override = _os.environ.get(f"QM_MAX_DAYS_{_tf.upper()}")
        if _env_override and _env_override.isdigit():
            _max_days = int(_env_override)
        try:
            _eff_end = pd.Timestamp(end_date) if end_date else df.index.max()
            _eff_end = min(_eff_end, df.index.max())
            _floor = _eff_end - pd.Timedelta(days=_max_days)
            _req_start = pd.Timestamp(start_date) if start_date else df.index.min()
            if _req_start < _floor:
                print(f"[load_ohlc] {_tf} backtest window capped to {_max_days}d: "
                      f"start {_req_start.date()} -> {_floor.date()}", flush=True)
                start_date = _floor
        except Exception as _e:
            print(f"[load_ohlc] window-cap check skipped ({_e})", flush=True)

    if start_date:
        df = df[df.index >= start_date]
    if end_date:
        df = df[df.index <= end_date]

    return df, df["close"], df["open"], df["high"], df["low"]


def make_target(close, horizon=4):
    """Create target: direction N bars ahead. Default 4 bars = 1 hour on 15-min data.

    Returns: target (pd.Series of -1, 0, 1)
    """
    return np.sign(close.shift(-horizon) - close)


def split_data(df, target, feature_cols, train_split=0.7, validation_date=""):
    """Train/test split. Handles both ratio and date-based splits.
    Drops NaN from target before splitting. Encodes labels to [0,1,2].

    Returns: dict with keys:
        X_train, X_test, y_train, y_test,
        y_train_enc, y_test_enc, enc,
        close_train, close_test,
        split_idx, split_dt, n_train, n_test
    """
    # Drop NaN from target
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]

    close = df["close"]

    # Build feature matrix
    X = df[feature_cols].copy()
    X = X.bfill().ffill()
    X = X.replace([np.inf, -np.inf], np.nan).fillna(0.0)

    # Split
    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)

    split_idx = max(1, min(split_idx, len(df) - 1))

    X_train = X.iloc[:split_idx]
    X_test = X.iloc[split_idx:]
    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]

    split_dt = str(df.index[split_idx])

    # Label encoding — always fit on [-1, 0, 1]
    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    return {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }


def compute_overlays(close, df_index):
    """Compute BB and MA overlays on full dataset. Always consistent.

    Returns: (bb_dict, ma_dict)
    """
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std

    ma50 = close.rolling(50).mean()
    ma100 = close.rolling(100).mean()
    ma200 = close.rolling(200).mean()

    def _safe(s):
        s = s.reindex(df_index).bfill().ffill()
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in s.values]

    bb = {"upper": _safe(bb_upper), "mid": _safe(bb_mid), "lower": _safe(bb_lower)}
    ma = {"ma50": _safe(ma50), "ma100": _safe(ma100), "ma200": _safe(ma200)}

    return bb, ma


def run_backtest(signal, close, capital=10000, cost=2e-5):
    """Run backtest with transaction costs.
    Uses price-based trade returns (same as webapp _compute_trades).
    Signal 0 = hold (keep current position), not close.

    Returns: dict with equity, trade_returns, long_returns, short_returns, bar_returns
    """
    sig_arr = signal.values
    price_arr = close.values
    idx = signal.index
    n = len(price_arr)

    # Trade returns — price-based (matches webapp _compute_trades exactly)
    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    last_dir = None
    entry_price = None
    entry_bar = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != last_dir:
            # Direction change — close previous trade, open new
            if last_dir is not None and entry_price is not None and entry_price != 0:
                ret = float(last_dir * (c - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if last_dir == 1:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                trade_log.append({
                    "type": "Buy" if last_dir == 1 else "Sell",
                    "entry_time": str(idx[entry_bar]),
                    "exit_time": str(idx[i]),
                    "entry_price": round(entry_price, 5),
                    "exit_price": round(c, 5),
                    "pnl": round(last_dir * (c - entry_price), 5),
                    "pnl_pct": round(ret * 100, 3),
                    "exit_reason": "signal",
                })
            entry_price = c
            entry_bar = i
            last_dir = s

    # Close last open trade
    if last_dir is not None and entry_price is not None and n > 0 and entry_price != 0:
        c = price_arr[-1]
        ret = float(last_dir * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if last_dir == 1:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        trade_log.append({
            "type": "Buy" if last_dir == 1 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[-1]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(c, 5),
            "pnl": round(last_dir * (c - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": "end",
        })

    # Equity curve from trade returns
    cumret = 1.0
    equity_vals = np.full(n, float(capital))
    trade_idx = 0
    in_trade = False
    t_entry_price = None
    t_dir = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != t_dir:
            if t_dir is not None and t_entry_price is not None and t_entry_price != 0:
                t_ret = t_dir * (c - t_entry_price) / t_entry_price - cost
                cumret *= (1 + t_ret)
            t_entry_price = c
            t_dir = s
        equity_vals[i] = capital * cumret

    # Bar returns for Sharpe
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if price_arr[i - 1] != 0 and last_dir is not None:
            bar_returns[i] = sig_arr[i - 1] * (price_arr[i] - price_arr[i - 1]) / price_arr[i - 1] if sig_arr[i - 1] != 0 else 0.0

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


def compute_trade_stats(trades, capital=10000):
    """Single source of truth for trade statistics.
    Every display path reads from this — no recomputation anywhere.
    All values are rounded and JSON-safe (no inf/nan).
    """
    if not trades:
        return {"n": 0, "wins": 0, "losses": 0, "wr": 0, "avg": 0,
                "best": 0, "worst": 0, "ret": 0, "np": 0, "mdd": 0,
                "pf": 0, "rr": 0, "expect": 0}
    w = [r for r in trades if r > 0]
    l = [r for r in trades if r < 0]
    cumret = 1.0
    for r in trades:
        cumret *= (1 + r)
    net_p = capital * (cumret - 1)
    # Max drawdown
    eq = np.cumprod([1.0] + [1 + r for r in trades])
    peak = np.maximum.accumulate(eq)
    mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    # Profit Factor
    gross_w = sum(w) if w else 0
    gross_l = abs(sum(l)) if l else 0
    pf = gross_w / gross_l if gross_l > 0 else (9999.0 if gross_w > 0 else 0)
    # Risk:Reward
    avg_w = float(np.mean(w)) if w else 0
    avg_l = abs(float(np.mean(l))) if l else 0
    rr = avg_w / avg_l if avg_l > 0 else (9999.0 if avg_w > 0 else 0)
    # Expectancy
    expect = net_p / len(trades)
    return {
        "n": len(trades), "wins": len(w), "losses": len(l),
        "wr": round(len(w) / len(trades), 4),
        "avg": round(float(np.mean(trades)), 6),
        "best": round(max(w), 6) if w else 0,
        "worst": round(min(l), 6) if l else 0,
        "ret": round(cumret - 1, 6),
        "np": round(net_p, 2),
        "mdd": round(mdd, 6),
        "pf": round(pf, 2),
        "rr": round(rr, 2),
        "expect": round(expect, 2),
    }


def compute_metrics(bt_result, close_test, capital=10000):
    """Compute all standard metrics from backtest result.
    Uses trade-level compounding (same as webapp _trade_stats) for accuracy.

    Returns: dict with total_ret, bh_ret, sharpe_strat, sharpe_bh, mdd, n_trades
    """
    equity = bt_result["equity"]
    trade_returns = bt_result["trade_returns"]

    # Total return — trade-level compounding (matches webapp)
    if trade_returns:
        cumret = 1.0
        for r in trade_returns:
            cumret *= (1 + r)
        total_ret = cumret - 1
    else:
        total_ret = 0.0

    # Buy and hold
    bh_equity = capital * (close_test / close_test.iloc[0])
    bh_ret = (bh_equity.iloc[-1] - capital) / capital if capital != 0 else 0.0

    # Sharpe ratio — trade-level (matches webapp: sqrt(252*26) annualization)
    if len(trade_returns) >= 2 and float(np.std(trade_returns)) > 0:
        sharpe_strat = float(np.mean(trade_returns) / np.std(trade_returns) * np.sqrt(252 * 26))
    else:
        sharpe_strat = 0.0

    bh_rets = bh_equity.pct_change().dropna()
    if len(bh_rets) > 1 and bh_rets.std() != 0:
        sharpe_bh = float((bh_rets.mean() / bh_rets.std()) * np.sqrt(252 * 24 * 4))
    else:
        sharpe_bh = 0.0

    # Max drawdown — trade-level (matches webapp)
    if trade_returns:
        eq = np.cumprod([1.0] + [1 + r for r in trade_returns])
        peak = np.maximum.accumulate(eq)
        mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    else:
        mdd = 0.0

    return {
        "total_ret": float(total_ret),
        "bh_ret": float(bh_ret),
        "sharpe_strat": float(sharpe_strat) if not np.isnan(sharpe_strat) else 0.0,
        "sharpe_bh": float(sharpe_bh) if not np.isnan(sharpe_bh) else 0.0,
        "mdd": float(mdd),
        "n_trades": len(trade_returns),
    }


# Diagnostics line/histogram series (equity / drawdown / rolling_acc / conf_hist)
# only feed the small Diagnostics charts — they're never used by the price chart
# or scroll-back. On a 1-min model trained over the (2.2-capped) window these are
# still ~30k points each; downsample to a visually-identical resolution before the
# dict leaves the trainer so it doesn't carry that into Server-A RAM / Postgres.
_RESULTS_SERIES_MAX = 5000


def _downsample_idx(n, cap=_RESULTS_SERIES_MAX):
    """Evenly-spaced index list spanning [0, n-1] (first+last always kept), or
    None when no downsampling is needed (n <= cap)."""
    if n <= cap:
        return None
    return np.unique(np.linspace(0, n - 1, cap).astype(int)).tolist()


def _take(arr, idx):
    """Subset a list by an index list (idx may be None → return arr unchanged)."""
    if idx is None or not isinstance(arr, list):
        return arr
    return [arr[i] for i in idx]


# trade_log / train_trade_log are lists of per-trade dicts (display-only — the
# Trade Log tab). They scale with TRADE count, not bar count, so the bar-window
# cap (Phase 2.2) doesn't bound them — a degenerate near-every-bar model can put
# 10k+ trade dicts in the blob (>3 MB). Cap each (independently — a small-N model
# keeps every trade) to the most-recent N, recording `*_total` + `*_truncated`
# so the true count is still reported. Real strategies have far fewer than
# _TRADE_LOG_MAX trades, so this only ever bites pathological models.
_TRADE_LOG_MAX = 5000


def _cap_trade_log(tl):
    """Return (capped_list, original_len, was_truncated)."""
    if not isinstance(tl, list) or len(tl) <= _TRADE_LOG_MAX:
        return tl, (len(tl) if isinstance(tl, list) else 0), False
    return tl[-_TRADE_LOG_MAX:], len(tl), True


def build_return_dict(split_result, bt_result, metrics, model, feature_cols,
                      signal_full, p_pos_test, p_neg_test, custom_figs=None,
                      bt_train_result=None, pre_stats=None):
    """Assemble the complete return dict. Handles ALL serialization.
    Never returns Timestamps, numpy arrays, or non-JSON types.

    Returns: JSON-safe dict with all required keys
    """
    df = split_result["df"]
    close = split_result["close"]
    close_test = split_result["close_test"]
    X_test = split_result["X_test"]
    y_test = split_result["y_test"]
    equity = bt_result["equity"]
    bar_returns = bt_result["bar_returns"]

    # OHLC
    ohlc_dates = [str(x) for x in df.index.tolist()]

    def _safe_list(arr):
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in arr]

    # Overlays
    bb, ma = compute_overlays(close, df.index)

    # Buy and hold equity
    capital = equity.iloc[0] if len(equity) > 0 else 10000
    bh_equity = capital * (close_test / close_test.iloc[0])

    # Confusion matrix
    from sklearn.metrics import confusion_matrix
    pred_test = model.predict(X_test)
    y_test_arr = np.asarray(y_test)
    cm = confusion_matrix(y_test_arr, pred_test, labels=[-1, 0, 1])

    # Rolling accuracy
    sig_arr = signal_full.reindex(close_test.index).values
    correct = pd.Series((pred_test == y_test_arr).astype(float), index=X_test.index)
    active_test = pd.Series(sig_arr != 0, index=close_test.index) if len(sig_arr) == len(close_test) else pd.Series(True, index=close_test.index)
    correct_active = correct.where(active_test, other=np.nan)
    rolling_acc = correct_active.rolling(30, min_periods=1).mean()

    # Feature importance
    importances = model.feature_importances_
    fi_pairs = sorted(zip(feature_cols, importances), key=lambda x: x[1])[-15:]

    # Drawdown
    rolling_max = equity.cummax()
    drawdown = (equity - rolling_max) / rolling_max.replace(0, np.nan)
    drawdown = drawdown.fillna(0.0)

    # ── Downsample the Diagnostics-only series (see _downsample_idx) ──────────
    _eq_dates = [str(x) for x in close_test.index.tolist()]
    _eq_strat = _safe_list(equity.values)
    _eq_bh    = _safe_list(bh_equity.values)
    _eq_idx   = _downsample_idx(len(_eq_dates))
    _eq_dates, _eq_strat, _eq_bh = _take(_eq_dates, _eq_idx), _take(_eq_strat, _eq_idx), _take(_eq_bh, _eq_idx)

    _ra_dates = [str(x) for x in rolling_acc.index.tolist()]
    _ra_vals  = [float(x) if (not np.isnan(x) and not np.isinf(x)) else None for x in rolling_acc.values]
    _ra_idx   = _downsample_idx(len(_ra_dates))
    _ra_dates, _ra_vals = _take(_ra_dates, _ra_idx), _take(_ra_vals, _ra_idx)

    _dd_dates = [str(x) for x in drawdown.index.tolist()]
    _dd_vals  = _safe_list(drawdown.values)
    _dd_idx   = _downsample_idx(len(_dd_dates))
    _dd_dates, _dd_vals = _take(_dd_dates, _dd_idx), _take(_dd_vals, _dd_idx)

    _cp_pos = [float(x) for x in (p_pos_test.tolist() if hasattr(p_pos_test, 'tolist') else list(p_pos_test))]
    _cp_neg = [float(x) for x in (p_neg_test.tolist() if hasattr(p_neg_test, 'tolist') else list(p_neg_test))]
    _cp_pos = _take(_cp_pos, _downsample_idx(len(_cp_pos)))
    _cp_neg = _take(_cp_neg, _downsample_idx(len(_cp_neg)))

    # ── Trade logs — display-only (Trade Log tab); cap to most-recent N with a
    #    `_total` field so the true count is still reported (see _cap_trade_log).
    #    NB: ret_dist arrays are left FULL — a downstream path in callbacks.py
    #    recomputes n_trades/win-rate from len(ret_dist), so a sample would skew
    #    the displayed counts; they're small anyway and gzip handles them.
    _tl_test, _tl_test_n, _tl_test_tr = _cap_trade_log(bt_result.get("trade_log", []))
    _tl_tr,   _tl_tr_n,   _tl_tr_tr   = _cap_trade_log(bt_train_result.get("trade_log", []) if bt_train_result else [])

    return {
        "ohlc": {
            "dates": ohlc_dates,
            "open": _safe_list(df["open"].values),
            "high": _safe_list(df["high"].values),
            "low": _safe_list(df["low"].values),
            "close": _safe_list(df["close"].values),
        },
        "signals": {
            "dates": [str(x) for x in signal_full.index.tolist()],
            "values": [float(x) for x in signal_full.values],
        },
        "bb": bb,
        "ma": ma,
        "equity": {
            "dates": _eq_dates,
            "strategy": _eq_strat,
            "bh": _eq_bh,
        },
        "feature_importance": {
            "names": [p[0] for p in fi_pairs],
            "values": [float(p[1]) for p in fi_pairs],
        },
        "conf_matrix": cm.tolist(),
        "conf_hist": {
            "p_pos": _cp_pos,
            "p_neg": _cp_neg,
        },
        "rolling_acc": {
            "dates": _ra_dates,
            "values": _ra_vals,
        },
        "drawdown": {
            "dates": _dd_dates,
            "values": _dd_vals,
        },
        "ret_dist": [float(x) for x in bt_result["trade_returns"]],
        "ret_dist_long": [float(x) for x in bt_result["long_returns"]],
        "ret_dist_short": [float(x) for x in bt_result["short_returns"]],
        "train_ret_dist": [float(x) for x in bt_train_result["trade_returns"]] if bt_train_result else [],
        "train_ret_dist_long": [float(x) for x in bt_train_result["long_returns"]] if bt_train_result else [],
        "train_ret_dist_short": [float(x) for x in bt_train_result["short_returns"]] if bt_train_result else [],
        "trade_log": _tl_test,
        "train_trade_log": _tl_tr,
        "trade_log_total": _tl_test_n,
        "train_trade_log_total": _tl_tr_n,
        "trade_log_truncated": _tl_test_tr,
        "train_trade_log_truncated": _tl_tr_tr,
        **(pre_stats or {}),
        "metrics": metrics,
        "split_dt": split_result["split_dt"],
        "split_idx": int(split_result["split_idx"]),
        "n_train": int(split_result["n_train"]),
        "n_test": int(split_result["n_test"]),
        "feature_cols": list(feature_cols),
        "custom_figs": custom_figs or [],
    }


# ════════════════════════════════════════════════════════════════════════════
# STRATEGY FRAMEWORK v2 — Config-driven architecture
# Claude writes feature_engineering() + strategy_config(). Framework does rest.
# ════════════════════════════════════════════════════════════════════════════

import importlib

_MODEL_REGISTRY = {
    "XGBClassifier":              ("xgboost", "XGBClassifier"),
    "RandomForestClassifier":     ("sklearn.ensemble", "RandomForestClassifier"),
    "GradientBoostingClassifier": ("sklearn.ensemble", "GradientBoostingClassifier"),
    "LogisticRegression":         ("sklearn.linear_model", "LogisticRegression"),
    "ExtraTreesClassifier":       ("sklearn.ensemble", "ExtraTreesClassifier"),
    "AdaBoostClassifier":         ("sklearn.ensemble", "AdaBoostClassifier"),
}


def _build_model_from_config(config, X_train, y_train_enc):
    """Build, fit, and wrap a model from strategy_config dict."""
    model_type = config.get("model_type", "RandomForestClassifier")
    model_params = dict(config.get("model_params", {}))

    if model_type not in _MODEL_REGISTRY:
        raise ValueError(f"Unknown model_type '{model_type}'. Valid: {list(_MODEL_REGISTRY.keys())}")

    module_path, class_name = _MODEL_REGISTRY[model_type]
    mod = importlib.import_module(module_path)
    cls = getattr(mod, class_name)

    # XGBoost defaults
    if class_name == "XGBClassifier":
        model_params.setdefault("use_label_encoder", False)
        model_params.setdefault("eval_metric", "mlogloss")
        model_params.setdefault("tree_method", "hist")
        # Determinism > speed (2026-05-25). XGBoost hist with n_jobs=-1 is
        # NON-reproducible even with random_state set — the parallel histogram
        # gradient-sum order varies across threads, so the SAME code + data
        # gives a slightly different model (and backtest) every run. Forcing
        # single-thread makes training bit-reproducible so: (a) a user who
        # copies a strategy and reruns it gets identical numbers, (b) the
        # community "Live" score matches a redeploy, (c) "same code, different
        # result" support reports go away. Cost: single-threaded XGB (a few
        # seconds slower on large windows; hist is fast so it's minor). FORCED
        # (not setdefault) so the guarantee can't be silently broken by a
        # strategy passing n_jobs. Exact reproducibility holds within the
        # platform (pinned versions / same Modal image); a user's own machine
        # with different xgboost/numpy/CPU can still differ in low-order bits.
        model_params["n_jobs"] = 1

    # Common defaults
    model_params.setdefault("random_state", 42)

    from model_wrapper import ModelWrapper

    clf = cls(**model_params)
    clf.fit(X_train, y_train_enc)

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    return ModelWrapper(clf, original_classes=enc.classes_, n_features=X_train.shape[1])


def _generate_signals(model, X, threshold):
    """Framework-owned signal generation. Deterministic threshold logic."""
    proba = model.predict_proba(X)
    classes = list(model.classes_)

    idx_pos = classes.index(1) if 1 in classes else None
    idx_neg = classes.index(-1) if -1 in classes else None

    p_pos = proba[:, idx_pos] if idx_pos is not None else np.zeros(len(X))
    p_neg = proba[:, idx_neg] if idx_neg is not None else np.zeros(len(X))

    signal_vals = np.zeros(len(X))
    signal_vals = np.where(p_pos >= threshold, 1.0, signal_vals)
    signal_vals = np.where(p_neg >= threshold, -1.0, signal_vals)
    # Both exceed: pick stronger
    both = (p_pos >= threshold) & (p_neg >= threshold)
    signal_vals[both] = np.where(p_pos[both] >= p_neg[both], 1.0, -1.0)

    return pd.Series(signal_vals, index=X.index), p_pos, p_neg


# ── Filter functions (all no-ops when config value is None) ──────────────

def _apply_direction_filter(signal, direction):
    """Zero out signals that don't match allowed direction."""
    if direction is None or direction == "both":
        return signal
    s = signal.copy()
    if direction == "long":
        s[s < 0] = 0.0
    elif direction == "short":
        s[s > 0] = 0.0
    return s


def _apply_session_filter(signal, index, session_hours):
    """Zero out signals outside session hours [start, end] UTC."""
    if session_hours is None:
        return signal
    s = signal.copy()
    start_h, end_h = session_hours[0], session_hours[1]
    hours = index.hour
    if start_h <= end_h:
        mask = (hours >= start_h) & (hours < end_h)
    else:  # wrap around midnight, e.g. [22, 6]
        mask = (hours >= start_h) | (hours < end_h)
    s[~mask] = 0.0
    return s


def _apply_atr_filter(signal, close, high, low, min_atr):
    """Zero out signals when NATR(14) is below threshold."""
    if min_atr is None:
        return signal
    hl = high - low
    hc = (high - close.shift(1)).abs()
    lc = (low - close.shift(1)).abs()
    tr = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr14 = tr.ewm(com=13, adjust=False).mean()
    natr = atr14 / close.replace(0, np.nan)
    s = signal.copy()
    s[natr < min_atr] = 0.0
    return s


def _apply_trend_filter(signal, close, trend_filter):
    """Only allow signals aligned with trend. e.g. 'sma_50': longs above SMA, shorts below."""
    if trend_filter is None:
        return signal
    # Parse: "sma_50" → SMA with period 50
    parts = trend_filter.lower().replace("-", "_").split("_")
    if len(parts) >= 2 and parts[0] in ("sma", "ema"):
        period = int(parts[1])
    else:
        return signal  # unknown filter, skip
    if parts[0] == "sma":
        trend_line = close.rolling(period).mean()
    else:
        trend_line = close.ewm(span=period, adjust=False).mean()
    s = signal.copy()
    # Longs only above trend, shorts only below
    s[(s > 0) & (close < trend_line)] = 0.0
    s[(s < 0) & (close > trend_line)] = 0.0
    return s


# ── run_backtest_v2: framework-owned SL/TP/cooldown/position management ──

def run_backtest_v2(signal, close, high, low, config, capital=10000, cost=2e-5):
    """Backtest with SL/TP/cooldown/direction handling built into the engine.
    Unlike run_backtest (v1), this function handles position exits internally.

    Returns: same dict shape as run_backtest()
    """
    stop_loss = config.get("stop_loss")
    take_profit = config.get("take_profit")
    cooldown = config.get("cooldown", 0)
    on_opposite = config.get("on_opposite", "reverse")

    sig_arr = signal.values
    close_arr = close.values
    high_arr = high.values
    low_arr = low.values
    idx = signal.index
    n = len(close_arr)

    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    equity_vals = np.full(n, float(capital))
    cumret = 1.0

    position = 0.0       # current direction: 1.0, -1.0, or 0.0 (flat)
    entry_price = None
    entry_bar = None      # index into arrays for entry time
    cooldown_remaining = 0

    def _log_trade(exit_bar, exit_px, ret, reason):
        trade_log.append({
            "type": "Buy" if position == 1.0 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[exit_bar]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(exit_px, 5),
            "pnl": round(position * (exit_px - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": reason,
        })

    for i in range(n):
        c = close_arr[i]
        h = high_arr[i]
        lo = low_arr[i]
        s = sig_arr[i]

        # 1. Check SL/TP if in trade
        if position != 0.0 and entry_price is not None:
            hit_sl = False
            hit_tp = False
            exit_price = None

            if position == 1.0:  # long
                if stop_loss is not None and lo <= entry_price * (1 - stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 - stop_loss)
                elif take_profit is not None and h >= entry_price * (1 + take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 + take_profit)
            else:  # short
                if stop_loss is not None and h >= entry_price * (1 + stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 + stop_loss)
                elif take_profit is not None and lo <= entry_price * (1 - take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 - take_profit)

            if hit_sl or hit_tp:
                ret = float(position * (exit_price - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if position == 1.0:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                _log_trade(i, exit_price, ret, "SL" if hit_sl else "TP")
                cumret *= (1 + ret)
                position = 0.0
                entry_price = None
                entry_bar = None
                cooldown_remaining = cooldown
                equity_vals[i] = capital * cumret
                continue

        # 2. Cooldown
        if cooldown_remaining > 0:
            cooldown_remaining -= 1
            equity_vals[i] = capital * cumret
            continue

        # 3. Signal processing
        if s != 0.0:
            if position == 0.0:
                # Open new trade
                position = s
                entry_price = c
                entry_bar = i
            elif s != position:
                # Opposite signal
                if on_opposite == "reverse":
                    # Close current + open opposite
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "signal")
                    cumret *= (1 + ret)
                    position = s
                    entry_price = c
                    entry_bar = i
                else:  # close_only
                    # Close current, go flat
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "close_only")
                    cumret *= (1 + ret)
                    position = 0.0
                    entry_price = None
                    entry_bar = None
                    cooldown_remaining = cooldown

        equity_vals[i] = capital * cumret

    # Close last open trade at final close
    if position != 0.0 and entry_price is not None and n > 0 and entry_price != 0:
        c = close_arr[-1]
        ret = float(position * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if position == 1.0:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        _log_trade(n - 1, c, ret, "end")
        cumret *= (1 + ret)
        equity_vals[-1] = capital * cumret

    # Bar returns for Sharpe (approximate)
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if close_arr[i - 1] != 0 and sig_arr[i - 1] != 0:
            bar_returns[i] = sig_arr[i - 1] * (close_arr[i] - close_arr[i - 1]) / close_arr[i - 1]

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


# ── run_strategy: the v2 orchestrator ────────────────────────────────────

def run_strategy(feature_fn, config_fn, data_path, start_date="", end_date="",
                 validation_date="", train_split=0.7, register_model_fn=None):
    """Config-driven strategy execution. Claude writes feature_fn + config_fn,
    framework does everything else.

    Returns: results dict (same format as webapp expects)
    """
    config = config_fn()

    # Auto-correct SL/TP if Claude passed percentage instead of decimal
    for _key in ("stop_loss", "take_profit"):
        _val = config.get(_key)
        if _val is not None and _val > 0.1:  # >10% is almost certainly a percentage
            config[_key] = _val / 100.0
            print(f"[strategy] Auto-corrected {_key}: {_val} -> {config[_key]} (was percentage, converted to decimal)")

    # 1. Load data
    df, close, open_, high, low = load_ohlc(data_path, start_date, end_date)

    # 2. Feature engineering (Claude's function)
    df = feature_fn(df, close, open_, high, low)
    close = df["close"]
    open_ = df["open"]
    high = df["high"]
    low = df["low"]

    # 3. Warm-up detection: drop rows where features have NaN BEFORE any fill
    feature_cols = [c for c in df.columns if c not in ("open", "high", "low", "close")]
    raw_nans = df[feature_cols].isna().any(axis=1)
    valid_rows = ~raw_nans
    if valid_rows.any():
        first_valid = valid_rows.idxmax()
        if raw_nans.loc[:first_valid].any():
            df = df.loc[first_valid:].copy()
            close = df["close"]
            open_ = df["open"]
            high = df["high"]
            low = df["low"]

    # 4. Target
    horizon = config.get("target_horizon", 4)
    target = make_target(close, horizon=horizon)

    # 5. Split (ffill only within each partition — no bfill leak)
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]
    close = df["close"]
    high = df["high"]
    low = df["low"]

    X = df[feature_cols].copy()
    X = X.replace([np.inf, -np.inf], np.nan)

    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)
    split_idx = max(1, min(split_idx, len(df) - 1))

    # ffill within train and test separately (no leak)
    X_train = X.iloc[:split_idx].ffill().fillna(0.0)
    X_test = X.iloc[split_idx:].ffill().fillna(0.0)
    X = pd.concat([X_train, X_test])

    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]
    high_test = high.iloc[split_idx:]
    low_test = low.iloc[split_idx:]

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    split_dt = str(df.index[split_idx])

    sp = {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }

    # 6. Build model from config
    model = _build_model_from_config(config, X_train, y_train_enc)

    # 7. Generate signals
    threshold = config.get("signal_threshold", 0.55)
    signal_train, p_pos_train, p_neg_train = _generate_signals(model, X_train, threshold)
    signal_test, p_pos_test, p_neg_test = _generate_signals(model, X_test, threshold)

    # 8. Apply filters (order: direction → session → ATR → trend)
    direction = config.get("direction", "both")
    signal_test = _apply_direction_filter(signal_test, direction)
    signal_train = _apply_direction_filter(signal_train, direction)

    session_filter = config.get("session_filter")
    signal_test = _apply_session_filter(signal_test, signal_test.index, session_filter)
    signal_train = _apply_session_filter(signal_train, signal_train.index, session_filter)

    min_atr = config.get("min_atr")
    if min_atr is not None:
        signal_test = _apply_atr_filter(signal_test, close_test, high_test, low_test, min_atr)

    trend_filter = config.get("trend_filter")
    if trend_filter is not None:
        signal_test = _apply_trend_filter(signal_test, close_test, trend_filter)

    signal_full = pd.concat([signal_train, signal_test])

    # 9. Backtest with SL/TP/cooldown (test + train)
    high_train = high.iloc[:split_idx]
    low_train = low.iloc[:split_idx]
    has_risk = (config.get("stop_loss") is not None or
                config.get("take_profit") is not None or
                config.get("cooldown", 0) > 0 or
                config.get("on_opposite", "reverse") != "reverse")

    if has_risk:
        bt = run_backtest_v2(signal_test, close_test, high_test, low_test, config, capital=10000)
        bt_train = run_backtest_v2(signal_train, close_train, high_train, low_train, config, capital=10000)
    else:
        bt = run_backtest(signal_test, close_test, capital=10000)
        bt_train = run_backtest(signal_train, close_train, capital=10000)

    # 10. Metrics
    metrics = compute_metrics(bt, close_test, capital=10000)

    # 11. Pre-compute all trade stats (single source of truth)
    pre_stats = {
        "train_stats": compute_trade_stats(bt_train.get("trade_returns", []), capital=10000),
        "test_stats":  compute_trade_stats(bt.get("trade_returns", []), capital=10000),
        "long_stats":  compute_trade_stats(bt.get("long_returns", []), capital=10000),
        "short_stats": compute_trade_stats(bt.get("short_returns", []), capital=10000),
    }

    # 12. Register model
    if register_model_fn is not None:
        register_model_fn(model)

    # 13. Build return dict
    return build_return_dict(sp, bt, metrics, model, feature_cols,
                             signal_full, p_pos_test, p_neg_test, custom_figs=[],
                             bt_train_result=bt_train, pre_stats=pre_stats)

# ── End strategy_utils ──


DATA_PATH = '/root/Desktop/QuantifyMe/data/ohlc/AUDUSD_15min.parquet'
START_DATE = '2026-04-15'
END_DATE = '2026-05-25'
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):

    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_period = 20
    bb_std    = 2.0
    bb_mid    = close.rolling(bb_period).mean()
    bb_sigma  = close.rolling(bb_period).std(ddof=0)
    bb_upper  = bb_mid + bb_std * bb_sigma
    bb_lower  = bb_mid - bb_std * bb_sigma

    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower

    # %B — position of close within the band (0 = lower, 1 = upper)
    bb_range       = bb_upper - bb_lower
    df["bb_pct_b"] = np.where(bb_range > 0, (close - bb_lower) / bb_range, 0.5)

    # Bandwidth — normalised band width (regime filter)
    df["bb_bandwidth"] = np.where(bb_mid > 0, bb_range / bb_mid, 0.0)

    # Distance from each band (signed, normalised by sigma)
    df["dist_lower"] = np.where(bb_sigma > 0, (close - bb_lower) / bb_sigma, 0.0)
    df["dist_upper"] = np.where(bb_sigma > 0, (bb_upper - close) / bb_sigma, 0.0)
    df["dist_mid"]   = np.where(bb_sigma > 0, (close - bb_mid)   / bb_sigma, 0.0)

    # Below lower band flag
    df["below_lower"] = np.where(close < bb_lower, 1, 0)
    # Above upper band flag
    df["above_upper"] = np.where(close > bb_upper, 1, 0)

    # ── RSI (14) ─────────────────────────────────────────────────────────────
    rsi_period = 14
    delta      = close.diff()
    gain       = delta.clip(lower=0)
    loss       = (-delta).clip(lower=0)
    avg_gain   = gain.ewm(com=rsi_period - 1, min_periods=rsi_period).mean()
    avg_loss   = loss.ewm(com=rsi_period - 1, min_periods=rsi_period).mean()
    rs         = np.where(avg_loss > 0, avg_gain / avg_loss, 100.0)
    rsi        = 100.0 - 100.0 / (1.0 + rs)
    df["rsi"]  = rsi

    # RSI-derived flags and distances
    df["rsi_oversold"]   = np.where(rsi < 35, 1, 0)
    df["rsi_overbought"] = np.where(rsi > 65, 1, 0)
    df["rsi_dist_35"]    = rsi - 35.0   # negative when oversold
    df["rsi_dist_65"]    = rsi - 65.0   # positive when overbought
    df["rsi_norm"]       = (rsi - 50.0) / 50.0  # centred, ±1 range

    # ── Core entry condition features ────────────────────────────────────────
    # Buy setup: close < lower BB AND RSI < 35
    df["long_setup"]  = np.where((close < bb_lower) & (rsi < 35), 1, 0)
    # Sell setup: close > upper BB AND RSI > 65
    df["short_setup"] = np.where((close > bb_upper) & (rsi > 65), 1, 0)

    # ── ATR (14) — volatility context ────────────────────────────────────────
    atr_period = 14
    hl         = high - low
    hc         = (high - close.shift(1)).abs()
    lc         = (low  - close.shift(1)).abs()
    tr         = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr        = tr.ewm(com=atr_period - 1, min_periods=atr_period).mean()
    df["atr"]  = atr
    df["natr"] = np.where(close > 0, atr / close, 0.0)

    # ── Momentum / Rate-of-Change ─────────────────────────────────────────────
    for n in [1, 3, 5, 10]:
        df[f"roc_{n}"] = np.where(
            close.shift(n) > 0,
            (close - close.shift(n)) / close.shift(n),
            0.0
        )

    # ── EMA trend context (fast / slow) ──────────────────────────────────────
    ema_fast = close.ewm(span=9,  min_periods=9).mean()
    ema_slow = close.ewm(span=21, min_periods=21).mean()
    df["ema_fast"] = ema_fast
    df["ema_slow"] = ema_slow
    df["ema_diff"] = np.where(ema_slow > 0, (ema_fast - ema_slow) / ema_slow, 0.0)
    df["ema_bull"] = np.where(ema_fast > ema_slow, 1, 0)

    # SMA-50 trend filter helper (used by framework trend_filter)
    df["sma_50"] = close.rolling(50).mean()

    # ── Candle body & wick features ───────────────────────────────────────────
    body        = (close - open_).abs()
    candle_rng  = (high - low).replace(0, np.nan)
    df["body_ratio"]  = (body / candle_rng).fillna(0.0)
    df["upper_wick"]  = np.where(candle_rng.notna(), (high  - close.clip(lower=open_)) / candle_rng.fillna(1), 0.0)
    df["lower_wick"]  = np.where(candle_rng.notna(), (close.clip(upper=open_) - low)   / candle_rng.fillna(1), 0.0)
    df["bull_candle"]  = np.where(close > open_, 1, 0)

    # ── Volume-like proxy — true range z-score ────────────────────────────────
    tr_mean = tr.rolling(20).mean()
    tr_std  = tr.rolling(20).std(ddof=0).replace(0, np.nan)
    df["tr_zscore"] = ((tr - tr_mean) / tr_std).fillna(0.0)

    # ── Lagged RSI and %B (1, 2, 3 bars back) ────────────────────────────────
    for lag in [1, 2, 3]:
        df[f"rsi_lag{lag}"]    = df["rsi"].shift(lag)
        df[f"bb_pct_b_lag{lag}"] = df["bb_pct_b"].shift(lag)

    # ── RSI slope ────────────────────────────────────────────────────────────
    df["rsi_slope3"] = df["rsi"] - df["rsi"].shift(3)

    # ── Mean-reversion proximity: how far price is from middle band ───────────
    df["pct_to_mid"] = np.where(close > 0, (bb_mid - close) / close, 0.0)

    # ── Fill any NaNs from warm-up ────────────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "BB Mean-Reversion + RSI Oversold/Overbought (XGBoost)",
        "model_type":       "XGBClassifier",
        "model_params": {
            "n_estimators":      500,
            "max_depth":         4,
            "learning_rate":     0.03,
            "subsample":         0.75,
            "colsample_bytree":  0.70,
            "min_child_weight":  5,
            "gamma":             0.1,
            "reg_alpha":         0.05,
            "reg_lambda":        1.5,
            "objective":         "binary:logistic",
            "random_state":      42,
            "n_jobs":            -1,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.0010,
        "take_profit":      0.0020,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [7, 17],
        "min_atr":          0.00005,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximize Sharpe ratio by exploiting Bollinger Band mean-reversion "
            "with RSI confirmation. Entry conditions (close < lower BB, RSI < 35 "
            "for longs; close > upper BB, RSI > 65 for shorts) are encoded as "
            "features together with momentum, ATR volatility, candle structure, "
            "and lagged indicators. XGBoost with strong regularisation "
            "(reg_lambda=1.5, gamma=0.1, min_child_weight=5) and a low learning "
            "rate avoids overfitting on the 6-week window. Session filter "
            "[7,17] UTC targets liquid London/NY overlap, reducing noise. "
            "TP:SL ratio of 2:1 supports positive expected value even at "
            "moderate win rates, pushing Sharpe higher."
        ),
        "notes": (
            "Features: %B position, RSI (raw + flags + slope + lags), "
            "EMA cross, ATR/NATR, ROC(1/3/5/10), candle body/wick ratios, "
            "TR z-score, distance-to-midband, long/short setup flags. "
            "Round-trip cost ~2e-5 is implicitly absorbed by the 10-pip TP target. "
            "Cooldown=0 allows immediate re-entry after mean-reversion completes."
        ),
    }

# ── Framework v2: auto-generated wrapper ──
def train_and_backtest():
    _vd = VALIDATION_DATE if 'VALIDATION_DATE' in globals() else ''
    _ts = TRAIN_SPLIT if 'TRAIN_SPLIT' in globals() else 0.7
    return run_strategy(
        feature_engineering, strategy_config,
        DATA_PATH, START_DATE, END_DATE,
        _vd, _ts,
        register_model_fn=register_model
    )

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:35:43
# Model            : Gradient Boosting
# Feature Eng.     : SMA (20,50,200) + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── SMA core features (required) ──────────────────────────────────────
    for period in [20, 50, 200]:
        sma = close.rolling(period).mean()
        df[f"sma_{period}"] = sma
        df[f"dm_sma_{period}"] = (close - sma) / sma

    # ── SMA slope (momentum of the moving average itself) ──────────────────
    for period in [20, 50, 200]:
        df[f"sma_{period}_slope"] = df[f"sma_{period}"].diff(5) / df[f"sma_{period}"].shift(5)

    # ── SMA crossover signals ──────────────────────────────────────────────
    df["sma_20_50_cross"]  = df["sma_20"] - df["sma_50"]
    df["sma_50_200_cross"] = df["sma_50"] - df["sma_200"]
    df["sma_20_200_cross"] = df["sma_20"] - df["sma_200"]

    # Sign of crossover difference (trend direction)
    df["trend_20_50"]  = np.where(df["sma_20_50_cross"]  > 0, 1, -1)
    df["trend_50_200"] = np.where(df["sma_50_200_cross"] > 0, 1, -1)

    # ── Price momentum ────────────────────────────────────────────────────
    for lag in [1, 4, 8, 16, 32]:
        df[f"return_{lag}"] = close.pct_change(lag)

    # ── Volatility features ───────────────────────────────────────────────
    # True Range
    prev_close = close.shift(1)
    tr = pd.concat([
        high - low,
        (high - prev_close).abs(),
        (low  - prev_close).abs()
    ], axis=1).max(axis=1)

    for atr_period in [14, 50]:
        atr = tr.rolling(atr_period).mean()
        df[f"atr_{atr_period}"] = atr
        df[f"natr_{atr_period}"] = atr / close

    # Rolling realised volatility
    log_ret = np.log(close / close.shift(1))
    for vol_period in [20, 50]:
        df[f"realvol_{vol_period}"] = log_ret.rolling(vol_period).std()

    # ── Bollinger Bands (20, 2σ) ──────────────────────────────────────────
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std
    df["bb_width"]    = (bb_upper - bb_lower) / bb_mid
    df["bb_pct"]      = (close - bb_lower) / (bb_upper - bb_lower + 1e-12)
    df["bb_position"] = (close - bb_mid) / (bb_std + 1e-12)

    # ── RSI (14) ──────────────────────────────────────────────────────────
    delta = close.diff()
    gain  = delta.clip(lower=0).rolling(14).mean()
    loss  = (-delta.clip(upper=0)).rolling(14).mean()
    rs    = gain / (loss + 1e-12)
    df["rsi_14"] = 100 - (100 / (1 + rs))

    # RSI normalised to [-1, 1]
    df["rsi_14_norm"] = (df["rsi_14"] - 50) / 50

    # ── MACD (12, 26, 9) ─────────────────────────────────────────────────
    ema12 = close.ewm(span=12, adjust=False).mean()
    ema26 = close.ewm(span=26, adjust=False).mean()
    macd_line   = ema12 - ema26
    signal_line = macd_line.ewm(span=9, adjust=False).mean()
    df["macd"]          = macd_line
    df["macd_signal"]   = signal_line
    df["macd_hist"]     = macd_line - signal_line
    df["macd_hist_chg"] = df["macd_hist"].diff()

    # Normalise MACD by price
    df["macd_norm"]   = df["macd"]      / close
    df["macd_hist_norm"] = df["macd_hist"] / close

    # ── Stochastic Oscillator (14, 3) ─────────────────────────────────────
    low14  = low.rolling(14).min()
    high14 = high.rolling(14).max()
    stoch_k = 100 * (close - low14) / (high14 - low14 + 1e-12)
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"]    = stoch_k
    df["stoch_d"]    = stoch_d
    df["stoch_diff"] = stoch_k - stoch_d

    # ── Rate-of-change ────────────────────────────────────────────────────
    for roc_period in [5, 10, 20]:
        df[f"roc_{roc_period}"] = (close - close.shift(roc_period)) / (close.shift(roc_period) + 1e-12)

    # ── Candle body and wick features ─────────────────────────────────────
    body      = (close - open_).abs()
    candle_rng = (high - low).replace(0, np.nan)
    df["body_ratio"]       = body / candle_rng
    df["upper_wick_ratio"] = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_rng
    df["lower_wick_ratio"] = (pd.concat([close, open_], axis=1).min(axis=1) - low)  / candle_rng
    df["candle_direction"]  = np.where(close >= open_, 1, -1)

    # ── Time-of-day features (hour) ───────────────────────────────────────
    hour = df.index.hour
    df["hour_sin"] = np.sin(2 * np.pi * hour / 24)
    df["hour_cos"] = np.cos(2 * np.pi * hour / 24)

    # ── Lagged returns as features ────────────────────────────────────────
    for lag in [1, 2, 3, 4]:
        df[f"close_lag_{lag}"] = close.shift(lag)
        df[f"ret_lag_{lag}"]   = log_ret.shift(lag)

    # ── Volume proxy: range-based ─────────────────────────────────────────
    df["range_abs"]  = high - low
    df["range_norm"] = (high - low) / close

    # ── High-Low channel position ─────────────────────────────────────────
    for ch_period in [20, 50]:
        ch_high = high.rolling(ch_period).max()
        ch_low  = low.rolling(ch_period).min()
        df[f"channel_pos_{ch_period}"] = (close - ch_low) / (ch_high - ch_low + 1e-12)

    # ── Fill NaN from warm-up ─────────────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EUR/USD SMA Trend + Multi-Indicator GBM Scalper",
        "model_type":       "GradientBoostingClassifier",
        "model_params":     {
            "n_estimators":     400,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.75,
            "max_features":     "sqrt",
            "min_samples_leaf": 20,
            "min_samples_split": 40,
            "validation_fraction": 0.1,
            "n_iter_no_change": 30,
            "tol":              1e-4,
            "random_state":     42,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 18],
        "min_atr":          None,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective": (
            "Maximise risk-adjusted return (Sharpe / Calmar). "
            "GradientBoostingClassifier chosen for its strong performance on "
            "tabular financial data with moderate feature sets. "
            "Shallow trees (max_depth=4) with high n_estimators and a low "
            "learning_rate reduce overfitting. subsample=0.75 adds stochastic "
            "regularisation. Early stopping via n_iter_no_change prevents "
            "over-training on the validation split. "
            "SL=0.5%, TP=1.0% gives a 1:2 risk-reward ratio, "
            "targeting positive expectancy even at sub-60% accuracy. "
            "Session filter (06-18 UTC) restricts trading to liquid hours "
            "covering London and New York overlap for EUR/USD. "
            "SMA-50 trend filter ensures trades align with the medium-term "
            "trend, reducing counter-trend noise."
        ),
        "notes": (
            "Features include required SMA(20,50,200) distances and crossovers, "
            "RSI-14, MACD histogram, Bollinger Band position, Stochastic, ATR, "
            "realised volatility, rate-of-change, candle structure ratios, "
            "channel position, lagged returns, and cyclical time encoding. "
            "Target horizon of 4 bars (1 hour) on 15-min data balances "
            "signal frequency with meaningful directional moves."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:36:58
# Model            : XGBoost
# Feature Eng.     : BB (20,2.0), ATR 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/USDCAD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_period = 20
    bb_std    = 2.0

    bb_mid   = close.rolling(bb_period).mean()
    bb_sigma = close.rolling(bb_period).std(ddof=0)
    bb_upper = bb_mid + bb_std * bb_sigma
    bb_lower = bb_mid - bb_std * bb_sigma

    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid
    df["bb_pct"]   = (close - bb_lower) / (bb_upper - bb_lower)

    # ── ATR (14) & Normalised ATR ────────────────────────────────────────────
    atr_period = 14
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)

    atr  = tr.ewm(alpha=1.0 / atr_period, min_periods=atr_period, adjust=False).mean()
    natr = atr / close

    df["atr"]  = atr
    df["natr"] = natr

    # ── Price momentum / returns ─────────────────────────────────────────────
    df["ret_1"]  = close.pct_change(1)
    df["ret_4"]  = close.pct_change(4)
    df["ret_8"]  = close.pct_change(8)
    df["ret_16"] = close.pct_change(16)

    # ── Distance from Bollinger mid / bands ──────────────────────────────────
    df["close_minus_mid"]   = (close - bb_mid)   / bb_mid
    df["close_minus_upper"] = (close - bb_upper) / bb_mid
    df["close_minus_lower"] = (close - bb_lower) / bb_mid

    # ── BB squeeze flag: width below rolling median ───────────────────────────
    bb_width_med = df["bb_width"].rolling(50).median()
    df["bb_squeeze"] = np.where(df["bb_width"] < bb_width_med, 1.0, 0.0)

    # ── BB mean-reversion z-score ────────────────────────────────────────────
    df["bb_z"] = (close - bb_mid) / (bb_sigma + 1e-12)

    # ── Candle body / wick features ──────────────────────────────────────────
    body       = (close - open_).abs()
    candle_rng = (high - low).replace(0, np.nan)
    df["body_ratio"]  = body / candle_rng
    df["upper_wick"]  = (high  - pd.concat([close, open_], axis=1).max(axis=1)) / candle_rng
    df["lower_wick"]  = (pd.concat([close, open_], axis=1).min(axis=1) - low)   / candle_rng
    df["bull_candle"]  = np.where(close > open_, 1.0, 0.0)

    # ── RSI (14) built from scratch ──────────────────────────────────────────
    rsi_period = 14
    delta  = close.diff()
    gain   = delta.clip(lower=0)
    loss   = (-delta).clip(lower=0)
    avg_g  = gain.ewm(alpha=1.0 / rsi_period, min_periods=rsi_period, adjust=False).mean()
    avg_l  = loss.ewm(alpha=1.0 / rsi_period, min_periods=rsi_period, adjust=False).mean()
    rs     = avg_g / (avg_l + 1e-12)
    rsi    = 100.0 - (100.0 / (1.0 + rs))
    df["rsi_14"] = rsi

    # RSI deviation from neutral 50
    df["rsi_dev"] = (rsi - 50.0) / 50.0

    # ── MACD (12, 26, 9) ─────────────────────────────────────────────────────
    ema12      = close.ewm(span=12, adjust=False).mean()
    ema26      = close.ewm(span=26, adjust=False).mean()
    macd_line  = ema12 - ema26
    macd_sig   = macd_line.ewm(span=9, adjust=False).mean()
    df["macd"]      = macd_line / close
    df["macd_hist"] = (macd_line - macd_sig) / close

    # ── Rolling volatility (realised over 20 bars) ───────────────────────────
    df["vol_20"] = df["ret_1"].rolling(20).std()

    # ── ATR z-score vs 50-bar rolling mean ───────────────────────────────────
    atr_mean     = atr.rolling(50).mean()
    atr_std      = atr.rolling(50).std(ddof=0)
    df["atr_z"]  = (atr - atr_mean) / (atr_std + 1e-12)

    # ── Volume-of-BB-touches over last 10 bars ───────────────────────────────
    near_upper = (close >= bb_upper * 0.998).astype(float)
    near_lower = (close <= bb_lower * 1.002).astype(float)
    df["touch_upper_10"] = near_upper.rolling(10).sum()
    df["touch_lower_10"] = near_lower.rolling(10).sum()

    # ── SMA 50 (trend filter helper) ─────────────────────────────────────────
    df["sma_50"]       = close.rolling(50).mean()
    df["close_vs_sma"] = (close - df["sma_50"]) / df["sma_50"]

    # ── EMA cross (9 / 21) ───────────────────────────────────────────────────
    ema9  = close.ewm(span=9,  adjust=False).mean()
    ema21 = close.ewm(span=21, adjust=False).mean()
    df["ema_cross"] = (ema9 - ema21) / close

    # ── Bar-of-day / session ─────────────────────────────────────────────────
    if hasattr(df.index, "hour"):
        df["hour_sin"] = np.sin(2 * np.pi * df.index.hour / 24.0)
        df["hour_cos"] = np.cos(2 * np.pi * df.index.hour / 24.0)
    else:
        df["hour_sin"] = 0.0
        df["hour_cos"] = 1.0

    # ── Lag features on bb_pct and rsi ───────────────────────────────────────
    for lag in [1, 2, 4]:
        df[f"bb_pct_lag{lag}"]  = df["bb_pct"].shift(lag)
        df[f"rsi_14_lag{lag}"]  = df["rsi_14"].shift(lag)
        df[f"macd_hist_lag{lag}"] = df["macd_hist"].shift(lag)

    # ── Fill NaN from warm-up ─────────────────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "USD/CAD BB Mean-Reversion + ATR XGBoost",
        "model_type":       "XGBClassifier",
        "model_params": {
            "n_estimators":      400,
            "max_depth":         4,
            "learning_rate":     0.04,
            "subsample":         0.75,
            "colsample_bytree":  0.70,
            "min_child_weight":  3,
            "gamma":             0.15,
            "reg_alpha":         0.10,
            "reg_lambda":        1.50,
            "objective":         "binary:logistic",
            "random_state":      42,
            "n_jobs":            -1,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [7, 20],
        "min_atr":          0.0002,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximise risk-adjusted return (Sharpe/Calmar) on USD/CAD 15-min "
            "using Bollinger Band mean-reversion signals augmented by ATR, RSI, "
            "MACD, and EMA-cross features fed into a regularised XGBoost classifier. "
            "SL=0.5% / TP=1.0% gives a 1:2 RR floor. Conservative depth (4) and "
            "strong L1/L2 regularisation prevent overfitting on a single year of data."
        ),
        "notes": (
            "BB squeeze flag and bb_z capture regime; atr_z filters noisy bars. "
            "Session filter 07-20 UTC covers London + NY overlap for tighter spreads. "
            "min_atr=0.0002 avoids dead-market whipsaws. Lag features on bb_pct and "
            "rsi_14 give the model short-term momentum context without look-ahead."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:44:26
# Model            : XGBoost
# Feature Eng.     : BB (20,2.0), ATR 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/AUDUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_period = 20
    bb_std    = 2.0

    bb_mid   = close.rolling(bb_period).mean()
    bb_sigma = close.rolling(bb_period).std(ddof=0)
    bb_upper = bb_mid + bb_std * bb_sigma
    bb_lower = bb_mid - bb_std * bb_sigma

    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid
    # guard against zero range
    bb_range = bb_upper - bb_lower
    df["bb_pct"]   = np.where(bb_range != 0, (close - bb_lower) / bb_range, 0.5)

    # ── ATR (14) & Normalised ATR ────────────────────────────────────────────
    atr_period = 14
    prev_close = close.shift(1)
    tr = pd.concat([
        high - low,
        (high - prev_close).abs(),
        (low  - prev_close).abs()
    ], axis=1).max(axis=1)

    atr  = tr.ewm(span=atr_period, min_periods=atr_period, adjust=False).mean()
    df["atr"]  = atr
    df["natr"] = np.where(close != 0, atr / close, 0.0)

    # ── RSI (14) ─────────────────────────────────────────────────────────────
    rsi_period = 14
    delta  = close.diff()
    gain   = delta.clip(lower=0)
    loss   = (-delta).clip(lower=0)
    avg_gain = gain.ewm(span=rsi_period, min_periods=rsi_period, adjust=False).mean()
    avg_loss = loss.ewm(span=rsi_period, min_periods=rsi_period, adjust=False).mean()
    rs  = avg_gain / avg_loss.replace(0, np.nan)
    rsi = 100 - (100 / (1 + rs))
    df["rsi_14"] = rsi

    # ── MACD (12, 26, 9) ─────────────────────────────────────────────────────
    ema12       = close.ewm(span=12, adjust=False).mean()
    ema26       = close.ewm(span=26, adjust=False).mean()
    macd_line   = ema12 - ema26
    signal_line = macd_line.ewm(span=9, adjust=False).mean()
    df["macd"]        = macd_line
    df["macd_signal"] = signal_line
    df["macd_hist"]   = macd_line - signal_line

    # ── EMA trend features ───────────────────────────────────────────────────
    ema_20  = close.ewm(span=20, adjust=False).mean()
    ema_50  = close.ewm(span=50, adjust=False).mean()
    ema_200 = close.ewm(span=200, adjust=False).mean()
    df["ema_20"]        = ema_20
    df["ema_50"]        = ema_50
    df["ema_200"]       = ema_200
    df["close_vs_ema20"]  = (close - ema_20) / ema_20
    df["close_vs_ema50"]  = (close - ema_50) / ema_50
    df["ema20_vs_ema50"]  = (ema_20 - ema_50) / ema_50
    df["ema50_vs_ema200"] = (ema_50 - ema_200) / ema_200

    # ── Price momentum / rate-of-change ──────────────────────────────────────
    for n in [1, 4, 8, 16]:
        df[f"roc_{n}"] = close.pct_change(n)

    # ── Rolling volatility ────────────────────────────────────────────────────
    df["vol_10"]  = close.pct_change().rolling(10).std()
    df["vol_20"]  = close.pct_change().rolling(20).std()
    df["vol_ratio"] = np.where(df["vol_20"] != 0,
                               df["vol_10"] / df["vol_20"], 1.0)

    # ── Stochastic %K / %D (14, 3) ───────────────────────────────────────────
    low_14  = low.rolling(14).min()
    high_14 = high.rolling(14).max()
    stoch_range = high_14 - low_14
    stoch_k = np.where(stoch_range != 0,
                       100 * (close - low_14) / stoch_range, 50.0)
    df["stoch_k"] = stoch_k
    df["stoch_d"] = pd.Series(stoch_k, index=close.index).rolling(3).mean()

    # ── Candle body / wick features ──────────────────────────────────────────
    body      = (close - open_).abs()
    total_rng = (high - low).replace(0, np.nan)
    df["body_ratio"]  = body / total_rng
    df["upper_wick"]  = (high - pd.concat([close, open_], axis=1).max(axis=1)) / total_rng
    df["lower_wick"]  = (pd.concat([close, open_], axis=1).min(axis=1) - low) / total_rng
    df["candle_dir"]  = np.sign(close - open_)

    # ── BB interaction features ───────────────────────────────────────────────
    df["price_above_bb_upper"] = np.where(close > bb_upper, 1, 0)
    df["price_below_bb_lower"] = np.where(close < bb_lower, 1, 0)
    df["bb_squeeze"]           = np.where(df["bb_width"] < df["bb_width"].rolling(50).quantile(0.20), 1, 0)

    # ── RSI regime bins (replacing pd.cut) ───────────────────────────────────
    df["rsi_oversold"]  = np.where(rsi < 30, 1, 0)
    df["rsi_overbought"] = np.where(rsi > 70, 1, 0)
    df["rsi_neutral"]   = np.where((rsi >= 30) & (rsi <= 70), 1, 0)

    # ── Volume proxy (if volume column exists) ───────────────────────────────
    if "volume" in df.columns:
        vol_ma = df["volume"].rolling(20).mean()
        df["volume_ratio"] = np.where(vol_ma != 0,
                                      df["volume"] / vol_ma, 1.0)

    # ── Lagged features (1-bar lag to avoid lookahead) ───────────────────────
    for feat in ["bb_pct", "rsi_14", "macd_hist", "natr", "stoch_k"]:
        df[f"{feat}_lag1"] = df[feat].shift(1)
        df[f"{feat}_lag2"] = df[feat].shift(2)

    # ── Fill NaNs from warm-up ────────────────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "AUD/USD Bollinger + ATR Mean-Rev (XGBoost)",
        "model_type":       "XGBClassifier",
        "model_params": {
            "n_estimators":       400,
            "max_depth":          4,
            "learning_rate":      0.04,
            "subsample":          0.75,
            "colsample_bytree":   0.70,
            "min_child_weight":   3,
            "gamma":              0.10,
            "reg_alpha":          0.10,
            "reg_lambda":         1.50,
            "objective":          "binary:logistic",
            "random_state":       42,
            "n_jobs":             -1,
        },
        "signal_threshold": 0.56,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [0, 23],
        "min_atr":          None,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe). "
            "XGBoost with moderate depth and heavy regularisation "
            "(gamma, alpha, lambda) prevents overfit on AUD/USD 15-min data. "
            "Bollinger Bands capture mean-reversion; ATR normalises volatility; "
            "RSI + MACD confirm momentum; 2:1 TP:SL ratio supports positive expectancy."
        ),
        "notes": (
            "Features: BB (20,2) width/pct, ATR-14/NATR, RSI-14, MACD histogram, "
            "EMA 20/50/200 spreads, Stochastic %K/%D, candle-body ratios, "
            "ROC at multiple horizons, volatility ratio, BB squeeze flag, "
            "lagged versions of key features. "
            "Threshold 0.56 filters marginal signals, improving precision. "
            "target_horizon=4 (1 hour) balances signal frequency vs. noise."
        ),
    }