feat: continuous strategy generator (WF, MTF, stability, ML models, auto-ensemble)

2026-07-27 15:37:44 +00:00 · 2026-05-03 22:42:29 +02:00
parent 982b902b1a
commit 8409451ac4
1 changed files with 184 additions and 0 deletions
@@ -0,0 +1,184 @@
+#!/usr/bin/env python
+"""
+Continuous Strategy Generator — runs indefinitely, improving over time.
+
+Features:
+- Infinite loop: generate → optimize → ensemble → repeat
+- Walk-Forward validation required (OOS Sharpe > 0)
+- Multi-Timeframe check (1min, 5min, 15min, 1h)
+- Rolling stability check (12-month Sharpe never negative)
+- ML model training when LLM suggests it's beneficial
+- Auto-ensemble from top strategies
+- Daytrading AND swing style alternating
+
+Usage:
+    python scripts/predix_continuous_strategies.py
+    python scripts/predix_continuous_strategies.py --style daytrading --rounds 100
+    python scripts/predix_continuous_strategies.py --style both --workers 4
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import logging
+import os
+import sys
+import time
+from datetime import datetime
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+from rdagent.scenarios.qlib.local.strategy_orchestrator import StrategyOrchestrator
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
+logger = logging.getLogger(__name__)
+
+BATCH_SIZE = 5
+COOLDOWN_SECONDS = 30
+
+
+def build_ml_model(factor_values: pd.DataFrame, close: pd.Series, style: str) -> dict | None:
+    """Train ML model if data is sufficient, return strategy dict or None."""
+    from sklearn.ensemble import GradientBoostingRegressor
+
+    df = factor_values.ffill().dropna()
+    close_aligned = close.reindex(df.index).ffill()
+
+    common = df.index.intersection(close_aligned.index)
+    if len(common) < 5000:
+        logger.info("ML: insufficient data (<5000 rows)")
+        return None
+
+    X = df.loc[common].values
+    y = close_aligned.loc[common].pct_change(96).shift(-96).fillna(0).values  # forward 96-bar return
+
+    split = int(len(X) * 0.7)
+    X_train, X_test = X[:split], X[split:]
+    y_train, y_test = y[:split], y[split:]
+
+    model = GradientBoostingRegressor(n_estimators=100, max_depth=5, random_state=42)
+    model.fit(X_train, y_train)
+
+    # Generate signal on test data
+    preds = model.predict(X_test)
+    signal = pd.Series(np.sign(preds), index=common[split:])
+
+    # Backtest
+    from rdagent.components.backtesting.vbt_backtest import backtest_signal_ftmo
+    bt = backtest_signal_ftmo(
+        close=close_aligned.loc[common[split:]],
+        signal=signal,
+        txn_cost_bps=2.14,
+        wf_rolling=True,
+    )
+
+    is_oos_sharpe = bt.get("wf_oos_sharpe_mean", 0)
+    if is_oos_sharpe <= 0:
+        logger.info(f"ML model rejected: OOS Sharpe={is_oos_sharpe:.2f}")
+        return None
+
+    logger.info(f"ML model accepted: Sharpe={bt['sharpe']:.2f} OOS={is_oos_sharpe:.2f}")
+    return {
+        "strategy_name": f"ML_GradientBoost_{style}_{int(time.time())}",
+        "status": "accepted",
+        "sharpe_ratio": round(bt["sharpe"], 4),
+        "max_drawdown": round(bt["max_drawdown"], 4),
+        "win_rate": round(bt["win_rate"], 4),
+        "n_trades": bt["n_trades"],
+        "oos_sharpe": round(is_oos_sharpe, 4),
+        "type": "ml_model",
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--style", default="both", choices=["daytrading", "swing", "both"])
+    parser.add_argument("--workers", type=int, default=2)
+    parser.add_argument("--rounds", type=int, default=0, help="Stop after N rounds (0=infinite)")
+    parser.add_argument("--min-sharpe", type=float, default=1.5)
+    parser.add_argument("--batch-size", type=int, default=5)
+    parser.add_argument("--ml-rounds", type=int, default=3, help="Train ML model every N rounds")
+    args = parser.parse_args()
+
+    print(f"\n{'='*60}")
+    print(f"  Predix Continuous Strategy Generator")
+    print(f"  Style: {args.style} | Workers: {args.workers}")
+    print(f"  Min Sharpe: {args.min_sharpe} | Batch: {args.batch_size}")
+    print(f"  ML every {args.ml_rounds} rounds")
+    print(f"{'='*60}\n")
+
+    round_num = 0
+    total_accepted = 0
+    total_ml_accepted = 0
+    start_time = datetime.now()
+
+    while True:
+        round_num += 1
+        styles = [args.style] if args.style != "both" else (["swing", "daytrading"] if round_num % 2 == 1 else ["daytrading", "swing"])
+
+        for style in styles:
+            print(f"\n--- Round {round_num} | Style: {style} ---")
+
+            orch = StrategyOrchestrator(
+                top_factors=20, trading_style=style,
+                min_sharpe=args.min_sharpe,
+                use_optuna=True, optuna_trials=30,
+            )
+
+            try:
+                results = orch.generate_strategies(count=BATCH_SIZE, workers=args.workers)
+            except Exception as e:
+                logger.error(f"Round {round_num} {style} failed: {e}")
+                continue
+
+            accepted = [r for r in results if r.get("status") == "accepted"]
+            total_accepted += len(accepted)
+            print(f"  Accepted: {len(accepted)}/{len(results)}  (Total: {total_accepted})")
+
+            for r in accepted[:3]:
+                print(f"    {r.get('strategy_name', '?')[:40]:40s} S={r.get('sharpe_ratio',0):.1f} OOS={r.get('oos_sharpe',0):.1f}")
+
+            # Ensemble after every round
+            ensemble = orch.build_ensemble(results)
+            if ensemble and ensemble.get("status") == "success":
+                print(f"  Ensemble: S={ensemble['sharpe_ratio']:.1f} OOS={ensemble['oos_sharpe']:.1f} ({len(ensemble['members'])} members)")
+
+            # ML model every N rounds
+            if round_num % args.ml_rounds == 0:
+                print(f"\n  [ML] Training model on all factors...")
+                factors = orch.load_top_factors()
+                if factors:
+                    factor_values = {}
+                    for f in factors:
+                        series = orch.load_factor_values(f["factor_name"])
+                        if series is not None:
+                            factor_values[f["factor_name"]] = series
+                    if len(factor_values) >= 3:
+                        df = pd.DataFrame(factor_values)
+                        if isinstance(df.index, pd.MultiIndex):
+                            df = df.droplevel(-1)
+                        ml_result = build_ml_model(df, orch.ohlcv_close, style)
+                        if ml_result:
+                            total_ml_accepted += 1
+                            print(f"  [ML] Accepted! S={ml_result['sharpe_ratio']:.1f} OOS={ml_result['oos_sharpe']:.1f}")
+
+        elapsed = (datetime.now() - start_time).total_seconds()
+        print(f"\n  Elapsed: {elapsed/60:.0f}min | Accepted: {total_accepted} (+{total_ml_accepted} ML) | Rate: {total_accepted/(elapsed/3600):.1f}/h")
+
+        if args.rounds > 0 and round_num >= args.rounds:
+            break
+
+        time.sleep(COOLDOWN_SECONDS)
+
+    print(f"\n{'='*60}")
+    print(f"  DONE: {total_accepted} strategies + {total_ml_accepted} ML models")
+    print(f"  Total time: {(datetime.now()-start_time).total_seconds()/3600:.1f}h")
+    print(f"{'='*60}")
+
+
+if __name__ == "__main__":
+    main()