feat(analysis): 3 strategie oneste validate OOS multi-crypto (DIP/TR/ROT)

Ricerca onesta post-squeeze su 8 crypto (2018-2026), engine fee-aware con ingresso eseguibile a close[i], uscita TP/SL intrabar, OOS held-out, sweep fee. Lezione madre: shortare cripto perde OOS sistematicamente (campione net-bull) -> tutte le strategie robuste sono long-biased. Tre meccanismi distinti e complementari: - DIP01 dip-buy z-score reversion (long-only, 1h) robusto BTC/ETH/SOL - TR01 EMA 20/100 trend-following (long-only, 4h) robusto su 5/8 asset - ROT01 rotazione cross-sectional momentum sul paniere (1d) OOS +44%, param-insensitive Engine e validazione: scripts/analysis/honest_lab.py + honest_final.py (+ honest_candidates/diag/diag2/trend/rotation). Diario in docs/diary/. Onesto sull'obiettivo: €50/giorno su €1000 in pochi mesi non e' raggiungibile a rischio sano (~1825%/anno); edge reali 30-60% OOS pluriennale. Via realistica: portafoglio delle 3, leva moderata, crescita composta. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-05-28 23:28:00 +02:00
parent 48435f6858
commit 212427ffa1
11 changed files with 1157 additions and 0 deletions
@@ -0,0 +1,96 @@
+"""Strategia #3 candidata: ROTAZIONE cross-sectional momentum (multi-crypto).
+Una sola strategia che usa l'INTERO paniere: ad ogni ribilanciamento alloca il
+capitale agli asset con momentum migliore (long-only). Cattura la dispersione tra
+crypto (gli alt forti corrono molto piu' di BTC nei bull) senza shortare nulla.
+
+Onesto: i pesi a close[i] usano solo rendimenti passati; il rendimento del bar
+i->i+1 e' realizzato con quei pesi. Fee sul turnover. Allineamento per timestamp.
+"""
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+PROJECT_ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(PROJECT_ROOT))
+
+from scripts.analysis.honest_lab import get_df, available_assets, FEE_RT  # noqa: E402
+
+LEV = 3.0
+GROSS = 0.45        # esposizione lorda = LEV*POS del singolo (0.15*3) per confronto equo
+
+
+def build_panel(assets: list[str], tf: str) -> pd.DataFrame:
+    """Matrice close allineata per timestamp (inner join)."""
+    closes = {}
+    for a in assets:
+        df = get_df(a, tf)
+        s = pd.Series(df["close"].values,
+                      index=pd.to_datetime(df["timestamp"], unit="ms", utc=True))
+        closes[a] = s[~s.index.duplicated()]
+    panel = pd.DataFrame(closes).dropna()
+    return panel
+
+
+def simulate_rotation(panel: pd.DataFrame, lookback=30, top_k=2,
+                      fee_rt=FEE_RT, gross=GROSS, abs_filter=True,
+                      oos_frac=0.0) -> dict:
+    """Ad ogni barra: ranking per rendimento passato `lookback`; pesi uguali sui
+    top_k con momentum>0 (se abs_filter); altrimenti cash. gross = esposizione tot.
+    oos_frac>0: parte a investire solo dall'ultimo frac del campione."""
+    P = panel.values
+    T, N = P.shape
+    rets = np.zeros_like(P)
+    rets[1:] = P[1:] / P[:-1] - 1
+    years = panel.index.year.values
+    start = max(lookback + 1, int(T * (1 - oos_frac)) if oos_frac else lookback + 1)
+    cap = peak = 1000.0
+    max_dd = 0.0
+    w = np.zeros(N)
+    yearly: dict[int, float] = {}
+    turn_total = 0.0
+    for i in range(start, T - 1):
+        mom = P[i] / P[i - lookback] - 1
+        order = np.argsort(mom)[::-1]
+        new_w = np.zeros(N)
+        chosen = [j for j in order if (mom[j] > 0 or not abs_filter)][:top_k]
+        if chosen:
+            for j in chosen:
+                new_w[j] = gross / len(chosen)
+        # fee sul turnover (one-way = fee_rt/2 su ogni variazione di peso)
+        turnover = np.abs(new_w - w).sum()
+        cap -= cap * turnover * (fee_rt / 2)
+        turn_total += turnover
+        w = new_w
+        port_ret = float(np.dot(w, rets[i + 1]))      # rendimento bar i->i+1
+        cap = max(cap * (1 + port_ret), 10.0)
+        peak = max(peak, cap); max_dd = max(max_dd, (peak - cap) / peak)
+        yearly[years[i]] = yearly.get(years[i], 0.0) + port_ret * 100
+    return {
+        "ret": (cap / 1000 - 1) * 100,
+        "dd": max_dd * 100,
+        "turnover": turn_total,
+        "yearly": yearly,
+        "pos_years": sum(1 for v in yearly.values() if v > 0),
+        "n_years": len(yearly),
+    }
+
+
+if __name__ == "__main__":
+    assets = available_assets()
+    print(f"ROTATION cross-sectional momentum — fee {FEE_RT*100:.2f}% RT, gross {GROSS} | OOS 30%")
+    print(f"  Paniere: {assets}")
+    for tf in ["1d", "4h"]:
+        panel = build_panel(assets, tf)
+        print(f"\n  === {tf} === panel {panel.shape[0]} barre, {panel.index[0].date()} -> {panel.index[-1].date()}")
+        print(f"  {'config':<22s}{'FULL%':>9s}{'OOS%':>9s}{'DD%':>6s}{'Turn':>7s}{'AnniP':>8s}")
+        for lb in [20, 30, 60, 90]:
+            for k in [1, 2, 3]:
+                full = simulate_rotation(panel, lookback=lb, top_k=k)
+                oos = simulate_rotation(panel, lookback=lb, top_k=k, oos_frac=0.30)
+                anni = f"{full['pos_years']}/{full['n_years']}"
+                print(f"  lb{lb:<3d} top{k:<14d}{full['ret']:>+9.0f}{oos['ret']:>+9.0f}"
+                      f"{full['dd']:>6.0f}{full['turnover']:>7.0f}{anni:>8s}")