feat(validation): WFA tooling + multi-fold results phase1-100 runs

Aggiunge 2 script di analisi per validare i top-K genomi cross-fold: - scripts/analyze_btc_winners.py: per-trade dump (wins/losses/winrate/ avg_win/avg_loss/return/maxDD/Sharpe) per ogni top-K × 4 fold expanding-window WFA. Usato per identificare i winner robusti vs i lucky-shot overfit. - scripts/compare_winners.py: cross-run comparison di 5 winner candidate (BTC 1h + ETH 1h + BTC 5m + ETH 5m) sui medesimi 4 fold, con totali cumulativi. Risultati WFA freezati: - validation-btc-100-001.json: BTC 1h baseline (undertrading=10) - validation-btc-100-001-thr3.json: BTC 1h con threshold=3 (rilassato per strategie ultra-selettive) - validation-btc-100-5m-thr3.json: BTC 5m con threshold=3 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-16 21:48:55 +00:00
parent 8b767da5e7
commit 9c871d1d86
5 changed files with 2166 additions and 0 deletions
@@ -0,0 +1,125 @@
+"""Analisi per-trade dei top-K candidate del run BTC.
+
+Per ciascun genome top-K, ri-esegue il backtest su ogni fold WFA e raccoglie:
+- n_trades, n_wins, n_losses, win_rate
+- max_drawdown
+- return, sharpe
+- list trade pnl summary
+
+Output stampato a stdout, non scrive su disco.
+"""
+from __future__ import annotations
+
+import argparse
+from datetime import datetime
+
+import pandas as pd  # type: ignore[import-untyped]
+
+from multi_swarm_core.agents.hypothesis import _try_parse
+from multi_swarm_core.backtest.engine import BacktestEngine
+from multi_swarm_core.cerbero.client import CerberoClient
+from multi_swarm_core.config import load_settings
+from multi_swarm_core.data.cerbero_ohlcv import CerberoOHLCVLoader, OHLCVRequest
+from multi_swarm_core.data.splits import expanding_walk_forward
+from multi_swarm_core.metrics.basic import max_drawdown, sharpe_ratio, total_return
+from multi_swarm_core.persistence.repository import Repository
+from multi_swarm_core.protocol.compiler import compile_strategy
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--run-id", required=True)
+    ap.add_argument("--top-k", type=int, default=10)
+    ap.add_argument("--n-folds", type=int, default=4)
+    ap.add_argument("--train-ratio", type=float, default=0.5)
+    ap.add_argument("--symbol", default="BTC-PERPETUAL")
+    ap.add_argument("--timeframe", default="1h")
+    ap.add_argument("--start", default="2018-09-01T00:00:00+00:00")
+    ap.add_argument("--end", default="2026-01-01T00:00:00+00:00")
+    ap.add_argument("--fees-bp", type=float, default=5.0)
+    args = ap.parse_args()
+
+    settings = load_settings()
+    repo = Repository(settings.ga_db_path)
+    repo.init_schema()
+    all_evals = repo.list_evaluations(args.run_id)
+    parseable = [
+        e for e in all_evals
+        if e.get("raw_text") and not e.get("parse_error") and e["fitness"] > 0
+    ]
+    parseable.sort(key=lambda e: e["fitness"], reverse=True)
+    seen: set[str] = set()
+    top: list[dict] = []
+    for e in parseable:
+        if e["genome_id"] in seen:
+            continue
+        seen.add(e["genome_id"])
+        top.append(e)
+        if len(top) >= args.top_k:
+            break
+
+    token = (
+        settings.cerbero_mainnet_token.get_secret_value()
+        if settings.cerbero_mainnet_token
+        else settings.cerbero_testnet_token.get_secret_value()
+    )
+    cerbero = CerberoClient(
+        base_url=settings.cerbero_base_url,
+        token=token,
+        bot_tag=settings.cerbero_bot_tag,
+    )
+    loader = CerberoOHLCVLoader(client=cerbero, cache_dir=settings.series_dir)
+    ohlcv = loader.load(OHLCVRequest(
+        symbol=args.symbol,
+        timeframe=args.timeframe,
+        start=datetime.fromisoformat(args.start),
+        end=datetime.fromisoformat(args.end),
+    ))
+    splits = expanding_walk_forward(ohlcv.index, train_ratio=args.train_ratio, n_folds=args.n_folds)
+    engine = BacktestEngine(fees_bp=args.fees_bp)
+
+    print(f"\n{'=' * 110}")
+    print(f"PER-TRADE ANALYSIS — top-{len(top)} genomes × {len(splits)} folds")
+    print(f"{'=' * 110}")
+
+    for ev in top:
+        strat, err = _try_parse(ev["raw_text"] or "")
+        if strat is None:
+            print(f"\n>>> {ev['genome_id'][:16]} — parse error: {err}")
+            continue
+
+        print(f"\n>>> {ev['genome_id']} (fit_IS={ev['fitness']:.4f}, sharpe_IS={ev['sharpe']:.3f})")
+        print(f"{'fold':<5} {'period':<26} {'trades':>7} {'wins':>5} {'losses':>7} {'win%':>6} {'avg_w':>9} {'avg_l':>9} {'ret':>7} {'maxDD':>7} {'sharpe':>7}")
+        for s in splits:
+            test_df = ohlcv.loc[s.test_idx]
+            try:
+                signal_fn = compile_strategy(strat)
+                signals = signal_fn(test_df)
+                bt = engine.run(test_df, signals)
+            except Exception as e:
+                print(f"  fold {s.fold}: error {e}")
+                continue
+            trades = bt.trades
+            n_trades = len(trades)
+            wins = [t.net_pnl for t in trades if t.net_pnl > 0]
+            losses = [t.net_pnl for t in trades if t.net_pnl <= 0]
+            n_wins = len(wins)
+            n_losses = len(losses)
+            win_rate = (n_wins / n_trades * 100) if n_trades else 0.0
+            avg_w = (sum(wins) / n_wins) if n_wins else 0.0
+            avg_l = (sum(losses) / n_losses) if n_losses else 0.0
+            # Normalize equity for DD/return
+            if n_trades > 0:
+                notional = float(test_df["close"].iloc[0])
+                equity_pos = (bt.equity_curve / notional) + 1.0
+                ret_pct = total_return(equity_pos)
+                dd = max_drawdown(equity_pos)
+                sr = sharpe_ratio(bt.returns, periods_per_year=8760)
+            else:
+                ret_pct = dd = sr = 0.0
+            period = f"{str(s.test_idx[0])[:10]}..{str(s.test_idx[-1])[:10]}"
+            print(f"{s.fold:<5} {period:<26} {n_trades:>7} {n_wins:>5} {n_losses:>7} {win_rate:>5.1f}% {avg_w:>9.1f} {avg_l:>9.1f} {ret_pct:>6.2%} {dd:>6.2%} {sr:>7.3f}")
+
+
+if __name__ == "__main__":
+    main()