research(alt): sweep 104 strategie alternative su Deribit (153 agenti) + marginal scorer

Ondata di ricerca onesta a largo spettro su BTC/ETH+DVOL certificati: 104 ipotesi distinte (11 famiglie), un agente-finder per ipotesi, verifica avversariale a 3 scettici sui promettenti, sintesi (153 agenti totali). Esito: NIENTE di nuovo regge -> conferma del soffitto strutturale ~1.3 BTC/ETH-direzionale; lo stack TP01+XS01+VRP01 resta imbattuto. - altlib.py: harness condiviso vettoriale leak-free (eval_weights/study_weights, fee-sweep, both-asset + hold-out 2025+). Riproduce i numeri canonici di TP01. - MARGINAL SCORER (study_marginal/marginal_vs_tp01): Sharpe INCREMENTALE vs baseline TP01 (corr, blend uplift OOS, alpha residua) + jackknife OOS (clean-year + drop-best-month). earns_slot = abs!=FAIL & ADDS & robust_oos. Smaschera gli overlay su TSMOM con PASS assoluti fasulli (CMB04, VOL11, ...) e il falso positivo KAMA (ADDS ma muore al jackknife). - runs/*.py (104) script riproducibili per ipotesi; wf_altstrat.js workflow. - Verdetto: 0 candidati deployabili; 2 LEAD fragili (VOL08, STA05_LS) da forward-monitor. - test_marginal_scorer.py blocca baseline + invarianti. Suite: 32 verde. Diario: docs/diary/2026-06-20-alt-strategies-100agent-sweep.md Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-20 19:50:39 +00:00
parent bf84bc91e2
commit 5ac4e16af8
111 changed files with 16924 additions and 0 deletions
@@ -0,0 +1,110 @@
+"""VOL03 — DVOL-gated TSMOM
+HYPOTHESIS: TP01-style multi-horizon TSMOM (vol-targeted, long-flat) but ONLY active when
+DVOL is BELOW its expanding median. When DVOL is elevated (above median), go flat.
+Rationale: in calm regimes (low DVOL), trend tends to persist; in high-vol regimes,
+momentum can reverse or get choppy. Gating on DVOL below median may improve risk-adjusted returns.
+
+NOTE: DVOL history starts 2021-03, so full backtest (2019+) will have NaN DVOL for early bars.
+We handle this by defaulting to ACTIVE (no gate) when DVOL is NaN, so pre-2021 bars
+are the same as vanilla TSMOM. This avoids burning early history on a look-ahead free gate.
+
+Internal grid (4 configs, total 2 TFs x 2 configs = 4 backtests within study_weights per TF):
+- VOL03-A: multi-horizon (1,3,6 month) TSMOM + DVOL < expanding median
+- VOL03-B: multi-horizon (1,3,6 month) TSMOM + DVOL < expanding 40th pctile (stricter gate)
+We test on 1d and 12h -> 2 TFs x 2 configs = 4 study_weights calls total (each covers BTC+ETH).
+Pick best config by min_asset_holdout_sharpe.
+"""
+import sys
+sys.path.insert(0, "/opt/docker/PythagorasGoal/scripts/research/alt")
+import altlib as al
+import numpy as np
+import pandas as pd
+
+
+def tsmom_dvol_gated(df: pd.DataFrame, dvol_pctile: float = 0.50) -> np.ndarray:
+    """
+    Multi-horizon TSMOM (1,3,6 month) long-flat, vol-targeted.
+    Gate: position is ZERO when DVOL >= expanding percentile threshold.
+    When DVOL is NaN (pre-2021), treat as gate=OFF (keep TSMOM signal).
+
+    dvol_pctile: gate triggers (flat) when DVOL >= this expanding pctile of historical DVOL.
+    """
+    c = df["close"].values.astype(float)
+    bpd = al.bars_per_day(df)
+    asset = None
+    # Detect asset from data (try BTC first, then ETH)
+    # We'll use a closure over the caller's asset name - but since target_fn(df) is called
+    # from study_weights which passes df, we need to infer asset from DVOL data availability.
+    # Try BTC DVOL first, then ETH.
+    dv = None
+    for a in ("BTC", "ETH"):
+        try:
+            dv = al.dvol(df, a)
+            asset = a
+            break
+        except Exception:
+            continue
+
+    # Multi-horizon TSMOM signal: sum of sign over 1m, 3m, 6m
+    h1 = int(30 * bpd)
+    h3 = int(90 * bpd)
+    h6 = int(180 * bpd)
+    direction = np.zeros(len(c))
+    for h in (h1, h3, h6):
+        sig = np.full(len(c), np.nan)
+        sig[h:] = np.sign(c[h:] / c[:-h] - 1)
+        direction += np.nan_to_num(sig, nan=0.0)
+    # Long-flat: only go long (direction > 0), else flat
+    direction = np.clip(np.sign(direction), 0.0, 1.0)
+
+    # DVOL gate: compute expanding percentile of DVOL causally
+    if dv is not None:
+        dvol_series = pd.Series(dv)
+        # Expanding percentile (causal)
+        gate_active = np.zeros(len(c), dtype=bool)  # True = be active (below threshold)
+        # Use rolling expanding quantile: pandas expanding().quantile() is causal
+        dvol_thresh = dvol_series.expanding(min_periods=30).quantile(dvol_pctile)
+        # Gate: active when dvol < threshold (below median = calm regime)
+        # NaN dvol (pre-2021): treat as gate=OFF -> still active (no penalty)
+        dvol_nan = dvol_series.isna() | dvol_thresh.isna()
+        gate_active = dvol_nan | (dvol_series < dvol_thresh)
+        direction = direction * gate_active.values.astype(float)
+
+    # Vol-target
+    return al.vol_target(direction, df, target_vol=0.20, vol_win_days=30, leverage_cap=2.0)
+
+
+def make_target_fn(dvol_pctile: float):
+    """Create a target function with given DVOL percentile gate."""
+    def target_fn(df: pd.DataFrame) -> np.ndarray:
+        return tsmom_dvol_gated(df, dvol_pctile=dvol_pctile)
+    return target_fn
+
+
+# --- Run 4 configs: 2 pctile thresholds x 2 TFs ---
+# But study_weights handles 2 TFs internally, so we need 2 separate calls.
+# Total: 2 configs x 1 call each (each covers both TFs) = 2 study_weights calls
+# Each call tests 2 TFs x 2 assets = 4 backtests per call -> 8 total. OK.
+
+configs = [
+    ("VOL03-A-median", 0.50),   # flat when DVOL >= expanding median
+    ("VOL03-B-p40",    0.40),   # flat when DVOL >= expanding 40th pctile (stricter gate)
+]
+
+reports = []
+for name, pctile in configs:
+    fn = make_target_fn(pctile)
+    rep = al.study_weights(name, fn, tfs=("1d", "12h"))
+    print(al.fmt(rep))
+    print("JSON:", al.as_json(rep))
+    print()
+    reports.append((name, pctile, rep))
+
+# Pick best config by min_asset_holdout_sharpe across all cells
+best_name, best_pctile, best_rep = max(
+    reports,
+    key=lambda x: x[2]["verdict"].get("best_holdout_sharpe", -99)
+)
+print(f"\n=== BEST CONFIG: {best_name} (dvol_pctile={best_pctile}) ===")
+print(al.fmt(best_rep))
+print("JSON:", al.as_json(best_rep))