analysis/quant_model.py

"""
Quantitative Gap Prediction Model v7 — 3-Model Ensemble
═══════════════════════════════════════════════════════
Production model for REAL MONEY trading on XM Broker (GOLD symbol).

Ensemble: GradientBoosting + LogisticRegression + RandomForest
Decision: Average of 3 model probabilities (most robust from backtesting)

Lot management (anti-martingale):
  - Base lot: 0.01 per $20 capital
  - Below 0.20 lot: loss reduces by 0.01
  - At/above 0.20 lot: loss reduces by 0.02
  - After win: restore to calculated level
  - 15% risk cap per trade
  - Hard cap at 1.00 lot

Features: 25 (20 raw + 5 interaction features)
"""

from __future__ import annotations

import asyncio
import logging
import math
from pathlib import Path
from typing import Any

import numpy as np
import pandas as pd

logger = logging.getLogger("gap_system.analysis.quant_model")

try:
    from sklearn.linear_model import LogisticRegression
    from sklearn.ensemble import GradientBoostingClassifier, RandomForestClassifier
    from sklearn.preprocessing import StandardScaler
    from sklearn.model_selection import TimeSeriesSplit
    from sklearn.metrics import accuracy_score, log_loss, brier_score_loss
    _sklearn_available = True
except ImportError:
    _sklearn_available = False
    logger.warning("scikit-learn not installed -- quant model disabled")


# ═══════════════════════════════════════════════════════════════════════════════
# FEATURE ENGINEERING
# ═══════════════════════════════════════════════════════════════════════════════

def _encode_cyclic(value: float, period: float) -> tuple[float, float]:
    angle = 2 * math.pi * value / period
    return math.sin(angle), math.cos(angle)


def _safe_float(val, default=0.0) -> float:
    if val is None:
        return default
    try:
        f = float(val)
        return default if (math.isnan(f) or math.isinf(f)) else f
    except (ValueError, TypeError):
        return default


def _engineer_features_from_row(row: dict, asset: str = "") -> dict:
    """Transform a DB row into the 25-feature vector."""
    features = {}
    is_gold = "XAU" in asset.upper() or "GOLD" in asset.upper()

    features["vix_close"] = _safe_float(row.get("vix_close"), 15.0)
    features["vix_change_5d"] = _safe_float(row.get("vix_change_5d"))
    features["volume_spike"] = _safe_float(row.get("volume_spike"), 1.0)
    features["weekly_return"] = _safe_float(row.get("weekly_return"))
    features["rsi_14"] = _safe_float(row.get("rsi_14"), 50.0)
    features["macd_hist"] = _safe_float(row.get("macd_hist"))
    features["ema_spread"] = _safe_float(row.get("ema_spread"))
    features["bb_width"] = _safe_float(row.get("bb_width"), 0.02)
    features["dxy_change"] = _safe_float(row.get("dxy_change"))

    prev = row.get("prev_gap_dir", "NONE")
    features["prev_gap_dir"] = 1.0 if prev == "BULLISH" else (-1.0 if prev == "BEARISH" else 0.0)

    friday_date = row.get("friday_date", "2020-01-01")
    try:
        month = pd.Timestamp(friday_date).month
    except Exception:
        month = 1
    sin_m, cos_m = _encode_cyclic(month, 12)
    features["month_sin"] = sin_m
    features["month_cos"] = cos_m

    features["atr_14"] = _safe_float(row.get("atr_14"), 1.0)
    features["gap_atr_ratio"] = _safe_float(row.get("gap_atr_ratio"))
    dom = _safe_float(row.get("day_of_month"), 15)
    features["day_of_month_norm"] = dom / 31.0
    features["prev_3_gaps_mean"] = _safe_float(row.get("prev_3_gaps_mean"))
    features["gap_fill_rate"] = _safe_float(row.get("gap_fill_rate"), 0.6)

    features["gold_dxy_ratio"] = _safe_float(row.get("gold_dxy_ratio")) if is_gold else 0.0
    features["gld_momentum"] = _safe_float(row.get("gld_momentum")) if is_gold else 0.0
    features["real_yield_proxy"] = _safe_float(row.get("real_yield_proxy"))

    # 5 interaction features
    features["vix_x_dxy"] = features["vix_close"] * features["dxy_change"]
    rsi_extreme = max(0, features["rsi_14"] - 70) + max(0, 30 - features["rsi_14"])
    features["rsi_extreme_x_momentum"] = rsi_extreme * features["weekly_return"]
    features["squeeze_x_trend"] = features["bb_width"] * abs(features["ema_spread"])
    features["vix_fear_regime"] = 1.0 if features["vix_close"] > 25 else 0.0
    features["month_end"] = 1.0 if dom >= 28 else 0.0

    return features


FEATURE_COLUMNS = [
    "vix_close", "vix_change_5d", "volume_spike", "weekly_return",
    "rsi_14", "macd_hist", "ema_spread", "bb_width",
    "dxy_change", "prev_gap_dir", "month_sin", "month_cos",
    "atr_14", "gap_atr_ratio", "day_of_month_norm",
    "prev_3_gaps_mean", "gap_fill_rate",
    "gold_dxy_ratio", "gld_momentum", "real_yield_proxy",
    "vix_x_dxy", "rsi_extreme_x_momentum", "squeeze_x_trend",
    "vix_fear_regime", "month_end",
]


# ═══════════════════════════════════════════════════════════════════════════════
# LOT MANAGER — Anti-Martingale with Tiered Reduction
# ═══════════════════════════════════════════════════════════════════════════════

class LotManager:
    """
    Manages lot sizing with anti-martingale (reduce on loss).

    Rules:
      - Base lot = 0.01 per $20 capital
      - Below 0.20 lot: each loss reduces by 0.01
      - At/above 0.20 lot: each loss reduces by 0.02
      - After each win: restore to calculated level
      - 15% risk cap per trade
      - Hard cap at 1.00 lot
    """

    def __init__(self, starting_capital: float = 50.0,
                 capital_per_step: float = 20.0,
                 base_lot: float = 0.01,
                 max_lot: float = 0.50,
                 max_risk_pct: float = 0.15,
                 max_loss_per_micro: float = 15.0):
        self.capital = starting_capital
        self.capital_per_step = capital_per_step
        self.base_lot = base_lot
        self.max_lot = max_lot
        self.max_risk_pct = max_risk_pct
        self.max_loss_per_micro = max_loss_per_micro
        self._consecutive_losses = 0
        self._last_result = None
        self._last_lot = base_lot

    def get_lot(self) -> float:
        """Calculate current lot size based on capital, loss streak, and risk cap."""
        # Base lot from capital steps (+0.01 per $20)
        calculated = self.base_lot * max(1, int(self.capital / self.capital_per_step))

        # Tiered loss reduction (anti-martingale)
        if calculated >= 0.20:
            reduction_per_loss = 0.02  # bigger lots = bigger reduction
        else:
            reduction_per_loss = 0.01
        reduction = reduction_per_loss * self._consecutive_losses
        adjusted = max(self.base_lot, calculated - reduction)

        # RISK CAP: max loss per trade must not exceed max_risk_pct of capital
        risk_capped = adjusted
        if self.capital > 0 and self.max_loss_per_micro > 0:
            max_risk_dollars = self.capital * self.max_risk_pct
            max_safe_lots = max_risk_dollars / self.max_loss_per_micro
            risk_capped = max(self.base_lot, round(max_safe_lots * 100) / 100 * self.base_lot)
            adjusted = min(adjusted, risk_capped)

        # Apply hard cap
        final = min(adjusted, self.max_lot)
        self._last_lot = final

        logger.info(
            "LotManager: capital=$%.2f, calc=%.2f, losses=%d, risk_cap=%.2f, final=%.2f",
            self.capital, calculated, self._consecutive_losses, risk_capped, final
        )
        return round(final, 2)

    def record_result(self, profit: float):
        """Record trade result to adjust future lot sizing."""
        self.capital += profit
        self.capital = max(0, self.capital)

        if profit > 0:
            self._consecutive_losses = 0
            self._last_result = "WIN"
        elif profit < 0:
            self._consecutive_losses += 1
            self._last_result = "LOSS"
        else:
            self._last_result = "BREAKEVEN"

        logger.info(
            "LotManager: result=%s ($%.2f), capital=$%.2f, streak=%d",
            self._last_result, profit, self.capital, self._consecutive_losses
        )

    def get_status(self) -> dict:
        return {
            "capital": self.capital,
            "current_lot": self.get_lot(),
            "consecutive_losses": self._consecutive_losses,
            "last_result": self._last_result,
        }


# ═══════════════════════════════════════════════════════════════════════════════
# QUANT MODEL — 3-Model Ensemble
# ═══════════════════════════════════════════════════════════════════════════════

class QuantGapModel:
    """
    3-model ensemble: GradientBoosting + LogisticRegression + RandomForest.
    Averages probabilities from all 3 for the final prediction.
    """

    def __init__(self, db_path: Path | None = None):
        self.db_path = db_path
        self.models = {}
        self.scaler = None
        self.is_trained = False
        self.train_accuracy = 0.0
        self.train_samples = 0
        self.asset_models: dict[str, dict] = {}
        self._feat_cols = FEATURE_COLUMNS
        self.lot_manager = LotManager()

    async def train(self) -> dict:
        """Train 3-model ensemble from historical_gaps table."""
        if not _sklearn_available:
            return {"status": "fallback", "reason": "sklearn_not_installed"}

        if self.db_path is None or not self.db_path.exists():
            return {"status": "skipped", "reason": "no_db"}

        import aiosqlite

        try:
            async with aiosqlite.connect(str(self.db_path)) as db:
                db.row_factory = aiosqlite.Row

                cursor = await db.execute("PRAGMA table_info(historical_gaps)")
                columns = [row[1] for row in await cursor.fetchall()]
                has_v3 = "atr_14" in columns
                has_v2 = "rsi_14" in columns

                if not has_v2:
                    return await self._train_basic(db)

                cursor = await db.execute("""
                    SELECT * FROM historical_gaps
                    WHERE gap_direction IN ('BULLISH', 'BEARISH')
                    ORDER BY friday_date ASC
                """)
                rows = [dict(r) for r in await cursor.fetchall()]

            if len(rows) < 30:
                return {"status": "insufficient_data", "samples": len(rows)}

            feat_cols = FEATURE_COLUMNS if has_v3 else [
                c for c in FEATURE_COLUMNS
                if c not in ("atr_14", "gap_atr_ratio", "day_of_month_norm",
                             "prev_3_gaps_mean", "gap_fill_rate",
                             "gold_dxy_ratio", "gld_momentum", "real_yield_proxy",
                             "vix_x_dxy", "rsi_extreme_x_momentum",
                             "squeeze_x_trend", "vix_fear_regime", "month_end")
            ]
            self._feat_cols = feat_cols

            X_rows, y = [], []
            for row in rows:
                asset = row.get("asset", "")
                feats = _engineer_features_from_row(row, asset)
                X_rows.append([feats.get(c, 0.0) for c in feat_cols])
                y.append(1 if row["gap_direction"] == "BULLISH" else 0)

            X = np.nan_to_num(np.array(X_rows, dtype=np.float64), nan=0.0, posinf=0.0, neginf=0.0)
            y = np.array(y)

            self.scaler = StandardScaler()
            X_scaled = self.scaler.fit_transform(X)

            # Train 3 global models
            self.models["gbm"] = GradientBoostingClassifier(
                n_estimators=200, max_depth=4, learning_rate=0.05,
                subsample=0.8, min_samples_leaf=10, random_state=42)
            self.models["gbm"].fit(X_scaled, y)

            self.models["lr"] = LogisticRegression(
                C=1.0, penalty="l2", solver="lbfgs",
                max_iter=1000, class_weight="balanced", random_state=42)
            self.models["lr"].fit(X_scaled, y)

            self.models["rf"] = RandomForestClassifier(
                n_estimators=300, max_depth=6,
                min_samples_leaf=10, class_weight="balanced", random_state=42)
            self.models["rf"].fit(X_scaled, y)

            self.is_trained = True

            probs = self._ensemble_predict(X_scaled)
            y_pred = (probs > 0.5).astype(int)
            self.train_accuracy = accuracy_score(y, y_pred)
            self.train_samples = len(y)

            try:
                brier = brier_score_loss(y, probs)
                logloss = log_loss(y, probs)
            except Exception:
                brier, logloss = 0.0, 0.0

            # Per-asset 3-model ensembles
            assets = sorted(set(r["asset"] for r in rows))
            for asset in assets:
                asset_rows = [r for r in rows if r["asset"] == asset]
                if len(asset_rows) < 20:
                    continue

                Xa = np.array([
                    [_engineer_features_from_row(r, asset).get(c, 0.0) for c in feat_cols]
                    for r in asset_rows
                ], dtype=np.float64)
                Xa = np.nan_to_num(Xa, nan=0.0, posinf=0.0, neginf=0.0)
                ya = np.array([1 if r["gap_direction"] == "BULLISH" else 0 for r in asset_rows])

                scaler_a = StandardScaler()
                Xa_s = scaler_a.fit_transform(Xa)
                msl = max(5, len(ya) // 50)

                asset_ens = {}
                asset_ens["gbm"] = GradientBoostingClassifier(
                    n_estimators=200, max_depth=4, learning_rate=0.05,
                    subsample=0.8, min_samples_leaf=msl, random_state=42)
                asset_ens["gbm"].fit(Xa_s, ya)

                asset_ens["lr"] = LogisticRegression(
                    C=1.0, penalty="l2", solver="lbfgs",
                    max_iter=1000, class_weight="balanced", random_state=42)
                asset_ens["lr"].fit(Xa_s, ya)

                asset_ens["rf"] = RandomForestClassifier(
                    n_estimators=300, max_depth=6, min_samples_leaf=msl,
                    class_weight="balanced", random_state=42)
                asset_ens["rf"].fit(Xa_s, ya)

                probs_a = np.mean([m.predict_proba(Xa_s)[:, 1] for m in asset_ens.values()], axis=0)
                acc_a = accuracy_score(ya, (probs_a > 0.5).astype(int))

                self.asset_models[asset] = {
                    "models": asset_ens, "scaler": scaler_a,
                    "accuracy": acc_a, "samples": len(ya),
                }
                logger.info("Ensemble [%s]: %.1f%% on %d samples", asset, acc_a * 100, len(ya))

            metrics = {
                "status": "trained",
                "model_type": "3-model ensemble (GBM + LR + RF)",
                "global_accuracy": round(self.train_accuracy, 4),
                "brier_score": round(brier, 4),
                "log_loss": round(logloss, 4),
                "samples": self.train_samples,
                "features": len(feat_cols),
                "assets": {a: {"acc": round(m["accuracy"], 4), "n": m["samples"]}
                           for a, m in self.asset_models.items()},
            }
            logger.info("Ensemble trained: %.1f%% accuracy (%d samples, %d features)",
                         self.train_accuracy * 100, self.train_samples, len(feat_cols))
            return metrics

        except Exception as e:
            logger.error("Model training failed: %s", e, exc_info=True)
            return {"status": "error", "error": str(e)}

    def _ensemble_predict(self, X_scaled: np.ndarray) -> np.ndarray:
        """Average probabilities from all 3 models."""
        probs = []
        for name, model in self.models.items():
            try:
                probs.append(model.predict_proba(X_scaled)[:, 1])
            except Exception:
                pass
        if not probs:
            return np.full(X_scaled.shape[0], 0.5)
        return np.mean(probs, axis=0)

    async def _train_basic(self, db) -> dict:
        try:
            cursor = await db.execute("""
                SELECT asset, gap_direction, COUNT(*) as cnt
                FROM historical_gaps WHERE gap_direction IN ('BULLISH', 'BEARISH')
                GROUP BY asset, gap_direction
            """)
            totals = await cursor.fetchall()
            self._basic_priors = {}
            for row in totals:
                asset, direction, count = row[0], row[1], row[2]
                if asset not in self._basic_priors:
                    self._basic_priors[asset] = {"BULLISH": 0, "BEARISH": 0}
                self._basic_priors[asset][direction] = count
            total = sum(v["BULLISH"] + v["BEARISH"] for v in self._basic_priors.values())
            self.is_trained = True
            self.train_samples = total
            return {"status": "basic_fallback", "samples": total}
        except Exception as e:
            return {"status": "error", "error": str(e)}

    async def predict_live(self, asset: str) -> dict:
        """Generate a live prediction using the 3-model ensemble."""
        default = {
            "bull_prob": 0.5, "bear_prob": 0.5, "confidence": 0.0,
            "direction": "NEUTRAL", "samples": 0,
            "explanation": "No trained model available",
            "lot_size": self.lot_manager.get_lot(),
            "model_agreement": {},
        }
        if not self.is_trained:
            return default

        model_asset = "XAUUSD" if asset == "GOLD" else asset

        try:
            features = await self._get_live_features(model_asset)
            feat_cols = self._feat_cols

            X = np.array([[features.get(c, 0.0) for c in feat_cols]], dtype=np.float64)
            X = np.nan_to_num(X, nan=0.0, posinf=0.0, neginf=0.0)

            if _sklearn_available and self.models and self.scaler is not None:
                if model_asset in self.asset_models:
                    am = self.asset_models[model_asset]
                    X_s = am["scaler"].transform(X)
                    model_set = am["models"]
                    model_label = f"asset-specific ({am['samples']})"
                else:
                    X_s = self.scaler.transform(X)
                    model_set = self.models
                    model_label = f"global ({self.train_samples})"

                individual = {}
                probs_list = []
                for name, model in model_set.items():
                    try:
                        p = model.predict_proba(X_s)[0]
                        bp = float(p[1]) if len(p) > 1 else 0.5
                        individual[name] = {
                            "bull_prob": round(bp, 4),
                            "direction": "BULLISH" if bp > 0.5 else "BEARISH",
                        }
                        probs_list.append(bp)
                    except Exception:
                        pass

                if not probs_list:
                    return default

                bull_prob = float(np.mean(probs_list))
                bear_prob = 1.0 - bull_prob

            elif hasattr(self, "_basic_priors") and model_asset in self._basic_priors:
                counts = self._basic_priors[model_asset]
                total = counts["BULLISH"] + counts["BEARISH"]
                bull_prob = (counts["BULLISH"] + 1) / (total + 2)
                bear_prob = 1.0 - bull_prob
                individual = {}
                model_label = f"frequency ({total})"
            else:
                return default

            if bull_prob > 0.50:
                direction = "BULLISH"
                confidence = bull_prob
            else:
                direction = "BEARISH"
                confidence = bear_prob

            agree_count = sum(1 for m in individual.values() if m["direction"] == direction)
            total_models = len(individual)

            samples = self.asset_models.get(model_asset, {}).get("samples", self.train_samples)
            explanation = (
                f"Ensemble v7 ({model_label}): P(bull)={bull_prob:.3f}, P(bear)={bear_prob:.3f}. "
                f"Agreement: {agree_count}/{total_models} models. "
                f"VIX={features.get('vix_close', 0):.1f}, RSI={features.get('rsi_14', 50):.1f}"
            )

            return {
                "bull_prob": round(bull_prob, 4),
                "bear_prob": round(bear_prob, 4),
                "confidence": round(confidence, 4),
                "direction": direction,
                "samples": samples,
                "explanation": explanation,
                "lot_size": self.lot_manager.get_lot(),
                "model_agreement": individual,
                "agree_count": agree_count,
                "total_models": total_models,
            }

        except Exception as e:
            logger.error("Live prediction failed for %s: %s", asset, e)
            default["explanation"] = f"Prediction error: {e}"
            return default

    async def _get_live_features(self, asset: str) -> dict:
        """Fetch current market features from yfinance."""
        from data.price_data import fetch_candles, get_vix, get_volume_spike
        from analysis.technical import ema, rsi, macd, bollinger_bands, atr as compute_atr

        features = {c: 0.0 for c in FEATURE_COLUMNS}
        is_gold = "XAU" in asset.upper() or "GOLD" in asset.upper()

        try:
            vix_val = await get_vix()
            features["vix_close"] = vix_val

            try:
                import yfinance as yf
                vix_hist = await asyncio.to_thread(
                    lambda: yf.download("^VIX", period="10d", interval="1d", progress=False))
                if len(vix_hist) >= 6:
                    if isinstance(vix_hist.columns, pd.MultiIndex):
                        vix_hist.columns = vix_hist.columns.get_level_values(0)
                    vix_hist.columns = [c.lower() for c in vix_hist.columns]
                    features["vix_change_5d"] = _safe_float(
                        (vix_hist["close"].iloc[-1] - vix_hist["close"].iloc[-6]) / vix_hist["close"].iloc[-6] * 100)
            except Exception:
                pass

            features["volume_spike"] = await get_volume_spike(asset)

            df = await fetch_candles(asset, "1d", 50)
            if not df.empty and len(df) >= 20:
                close, high, low = df["close"], df["high"], df["low"]

                if len(close) >= 6:
                    features["weekly_return"] = _safe_float(
                        (close.iloc[-1] - close.iloc[-6]) / close.iloc[-6] * 100)

                rsi_series = rsi(close)
                features["rsi_14"] = _safe_float(rsi_series.iloc[-1], 50.0)

                _, _, hist = macd(close)
                features["macd_hist"] = _safe_float(hist.iloc[-1])

                ema10, ema20 = ema(close, 10), ema(close, 20)
                price = float(close.iloc[-1])
                if price > 0:
                    features["ema_spread"] = _safe_float((ema10.iloc[-1] - ema20.iloc[-1]) / price * 100)

                upper, mid, lower = bollinger_bands(close)
                mid_val = _safe_float(mid.iloc[-1])
                if mid_val > 0:
                    features["bb_width"] = _safe_float((upper.iloc[-1] - lower.iloc[-1]) / mid_val, 0.02)

                atr_series = compute_atr(high, low, close)
                features["atr_14"] = _safe_float(atr_series.iloc[-1], 1.0)

            # DXY
            try:
                import yfinance as yf
                dxy_hist = await asyncio.to_thread(
                    lambda: yf.download("DX-Y.NYB", period="10d", interval="1d", progress=False))
                if len(dxy_hist) >= 6:
                    if isinstance(dxy_hist.columns, pd.MultiIndex):
                        dxy_hist.columns = dxy_hist.columns.get_level_values(0)
                    dxy_hist.columns = [c.lower() for c in dxy_hist.columns]
                    features["dxy_change"] = _safe_float(
                        (dxy_hist["close"].iloc[-1] - dxy_hist["close"].iloc[-6]) / dxy_hist["close"].iloc[-6] * 100)
            except Exception:
                pass

            # Time features
            from datetime import datetime, timezone
            now = datetime.now(timezone.utc)
            sin_m, cos_m = _encode_cyclic(now.month, 12)
            features["month_sin"] = sin_m
            features["month_cos"] = cos_m
            features["day_of_month_norm"] = now.day / 31.0
            features["month_end"] = 1.0 if now.day >= 28 else 0.0
            features["vix_fear_regime"] = 1.0 if features["vix_close"] > 25 else 0.0

            # Previous gaps from DB
            try:
                import aiosqlite
                if self.db_path and self.db_path.exists():
                    async with aiosqlite.connect(str(self.db_path)) as db:
                        cursor = await db.execute(
                            "SELECT gap_direction, gap_pct, gap_fill_rate FROM historical_gaps WHERE asset=? ORDER BY friday_date DESC LIMIT 3",
                            (asset,))
                        rows = await cursor.fetchall()
                        if rows:
                            features["prev_gap_dir"] = 1.0 if rows[0][0] == "BULLISH" else (-1.0 if rows[0][0] == "BEARISH" else 0.0)
                            features["prev_3_gaps_mean"] = _safe_float(np.mean([r[1] for r in rows if r[1]]))
                            features["gap_fill_rate"] = _safe_float(rows[0][2], 0.6)
            except Exception:
                pass

            # Gold-specific
            if is_gold:
                try:
                    import yfinance as yf
                    gld = await asyncio.to_thread(
                        lambda: yf.download("GLD", period="10d", interval="1d", progress=False))
                    if len(gld) >= 6:
                        if isinstance(gld.columns, pd.MultiIndex):
                            gld.columns = gld.columns.get_level_values(0)
                        gld.columns = [c.lower() for c in gld.columns]
                        features["gld_momentum"] = _safe_float(
                            (gld["close"].iloc[-1] - gld["close"].iloc[-6]) / gld["close"].iloc[-6] * 100)
                except Exception:
                    pass

                try:
                    import yfinance as yf
                    tip = await asyncio.to_thread(
                        lambda: yf.download("TIP", period="10d", interval="1d", progress=False))
                    if len(tip) >= 6:
                        if isinstance(tip.columns, pd.MultiIndex):
                            tip.columns = tip.columns.get_level_values(0)
                        tip.columns = [c.lower() for c in tip.columns]
                        features["real_yield_proxy"] = _safe_float(
                            (tip["close"].iloc[-1] - tip["close"].iloc[-6]) / tip["close"].iloc[-6] * 100)
                except Exception:
                    pass

            # Interaction features
            features["vix_x_dxy"] = features["vix_close"] * features["dxy_change"]
            rsi_extreme = max(0, features["rsi_14"] - 70) + max(0, 30 - features["rsi_14"])
            features["rsi_extreme_x_momentum"] = rsi_extreme * features["weekly_return"]
            features["squeeze_x_trend"] = features["bb_width"] * abs(features["ema_spread"])

        except Exception as e:
            logger.error("Feature extraction failed for %s: %s", asset, e)

        return features

    def get_metrics(self) -> dict:
        return {
            "is_trained": self.is_trained,
            "model_type": "3-model ensemble (GBM + LR + RF)",
            "global_accuracy": round(self.train_accuracy, 4) if self.is_trained else None,
            "train_samples": self.train_samples,
            "feature_count": len(self._feat_cols),
            "features": self._feat_cols,
            "asset_models": {
                a: {"accuracy": round(m["accuracy"], 4), "samples": m["samples"]}
                for a, m in self.asset_models.items()
            },
            "lot_status": self.lot_manager.get_status(),
        }