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>
This commit is contained in:
Adriano Dal Pastro
@@ -0,0 +1,194 @@
|
||||
"""STA04 — K-means regime -> trend gating.
|
||||
|
||||
IDEA: cluster causal (vol, return, range) features using K-means with expanding
|
||||
statistics (z-scored causally), then enable TSMOM only in the historically-bullish/
|
||||
trending cluster. No future labels. Fully causal.
|
||||
|
||||
APPROACH:
|
||||
- Features (causal at bar i):
|
||||
1. realized_vol (30-day annualized)
|
||||
2. momentum return (lookback days)
|
||||
3. normalized range = ATR / close (relative range)
|
||||
- Expanding z-score: we don't know the distribution of features ahead of time.
|
||||
We compute expanding mean/std up to bar i for each feature, then z-score.
|
||||
This is causal: uses data[0..i] only.
|
||||
- K-means: we run offline K-means on the TRAINING portion (full history up to a
|
||||
burn-in), then use the fitted centroids to classify new bars causally.
|
||||
Strategy: classify each bar, determine which cluster(s) historically have
|
||||
been bullish/trending (positive mean return), gate TSMOM only in those clusters.
|
||||
- TSMOM signal: sign of 3-month return, vol-targeted.
|
||||
|
||||
GRID (<=4 combos to keep total backtests <=6 with 2 TFs):
|
||||
- (n_clusters=3, lookback_months=3) <- canonical
|
||||
- (n_clusters=4, lookback_months=3) <- more granular clusters
|
||||
Keep TFs = (1d, 12h).
|
||||
"""
|
||||
import sys
|
||||
sys.path.insert(0, "/opt/docker/PythagorasGoal/scripts/research/alt")
|
||||
import altlib as al
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
from sklearn.cluster import KMeans
|
||||
|
||||
|
||||
def expanding_zscore(x: np.ndarray, min_periods: int = 30) -> np.ndarray:
|
||||
"""Causal expanding z-score: at bar i, use data[0..i] to compute mean/std."""
|
||||
out = np.full(len(x), np.nan)
|
||||
for i in range(min_periods, len(x)):
|
||||
window = x[:i+1]
|
||||
m = np.nanmean(window)
|
||||
s = np.nanstd(window)
|
||||
if s > 0:
|
||||
out[i] = (x[i] - m) / s
|
||||
else:
|
||||
out[i] = 0.0
|
||||
return out
|
||||
|
||||
|
||||
def build_features(df: pd.DataFrame, lookback_months: int) -> np.ndarray:
|
||||
"""Build causal feature matrix [vol_z, momentum_z, range_z] for each bar."""
|
||||
c = df["close"].values.astype(float)
|
||||
bpd = al.bars_per_day(df)
|
||||
bpy = bpd * 365.25
|
||||
|
||||
# Feature 1: realized vol (30d)
|
||||
r = al.simple_returns(c)
|
||||
rv = al.realized_vol(r, max(2, 30 * bpd), bpy)
|
||||
|
||||
# Feature 2: momentum return over lookback_months
|
||||
lb_bars = int(lookback_months * 30.44 * bpd)
|
||||
mom = np.zeros(len(c))
|
||||
for i in range(lb_bars, len(c)):
|
||||
mom[i] = c[i] / c[i - lb_bars] - 1.0
|
||||
|
||||
# Feature 3: normalized range (ATR / close)
|
||||
at = al.atr(df, win=max(2, 14))
|
||||
rng = np.where(c > 0, at / c, 0.0)
|
||||
|
||||
# Expanding z-score (causal)
|
||||
rv_z = expanding_zscore(rv, min_periods=60)
|
||||
mom_z = expanding_zscore(mom, min_periods=60)
|
||||
rng_z = expanding_zscore(rng, min_periods=60)
|
||||
|
||||
feat = np.column_stack([rv_z, mom_z, rng_z])
|
||||
return feat
|
||||
|
||||
|
||||
def make_target(df: pd.DataFrame, n_clusters: int, lookback_months: int,
|
||||
train_frac: float = 0.5) -> np.ndarray:
|
||||
"""
|
||||
K-means regime-gated TSMOM.
|
||||
|
||||
1. Build causal features.
|
||||
2. Use the first train_frac of valid data to fit K-means.
|
||||
3. Label each cluster: positive if mean forward return (in training) is positive.
|
||||
4. Gate TSMOM: position = vol_targeted_tsmom * in_bullish_cluster.
|
||||
"""
|
||||
c = df["close"].values.astype(float)
|
||||
bpd = al.bars_per_day(df)
|
||||
n = len(df)
|
||||
|
||||
# Build features
|
||||
feat = build_features(df, lookback_months)
|
||||
|
||||
# Identify valid (non-NaN) rows
|
||||
valid_mask = np.all(np.isfinite(feat), axis=1)
|
||||
|
||||
# TSMOM signal: sign of lookback_months return, vol-targeted, long-only (flat on negative)
|
||||
lb_bars = int(lookback_months * 30.44 * bpd)
|
||||
tsmom_dir = np.zeros(n)
|
||||
for i in range(lb_bars, n):
|
||||
ret = c[i] / c[i - lb_bars] - 1.0
|
||||
tsmom_dir[i] = 1.0 if ret > 0 else 0.0 # long-flat (no short, consistent with TP01)
|
||||
|
||||
tsmom_pos = al.vol_target(tsmom_dir, df, target_vol=0.20, vol_win_days=30, leverage_cap=2.0)
|
||||
|
||||
# Find the training cutoff (first train_frac of valid bars)
|
||||
valid_idx = np.where(valid_mask)[0]
|
||||
if len(valid_idx) < n_clusters * 20:
|
||||
# Not enough data, return raw tsmom
|
||||
return tsmom_pos
|
||||
|
||||
train_end_idx = valid_idx[int(len(valid_idx) * train_frac)]
|
||||
|
||||
# Fit K-means on training portion
|
||||
train_feat = feat[valid_idx[valid_idx <= train_end_idx]]
|
||||
if len(train_feat) < n_clusters * 10:
|
||||
return tsmom_pos
|
||||
|
||||
km = KMeans(n_clusters=n_clusters, n_init=10, random_state=42)
|
||||
km.fit(train_feat)
|
||||
|
||||
# Determine cluster "bullishness" from training data:
|
||||
# For each training bar, check if the next bar's return is positive.
|
||||
# A cluster is "bullish" if mean(next_return | cluster) > 0.
|
||||
r = al.simple_returns(c)
|
||||
train_labels = km.labels_
|
||||
train_valid_indices = valid_idx[valid_idx <= train_end_idx]
|
||||
|
||||
cluster_returns = {k: [] for k in range(n_clusters)}
|
||||
for i_pos, idx_i in enumerate(train_valid_indices):
|
||||
if idx_i + 1 < n:
|
||||
cluster_returns[train_labels[i_pos]].append(r[idx_i + 1])
|
||||
|
||||
bullish_clusters = set()
|
||||
for k, rets in cluster_returns.items():
|
||||
if len(rets) > 5 and np.mean(rets) > 0:
|
||||
bullish_clusters.add(k)
|
||||
|
||||
# If no bullish cluster found, use all clusters (fall back to pure TSMOM)
|
||||
if not bullish_clusters:
|
||||
bullish_clusters = set(range(n_clusters))
|
||||
|
||||
# Classify ALL valid bars causally using fitted centroids
|
||||
all_valid_feat = feat[valid_mask]
|
||||
all_labels = km.predict(all_valid_feat)
|
||||
|
||||
# Build gate array
|
||||
gate = np.zeros(n)
|
||||
for i_pos, idx_i in enumerate(np.where(valid_mask)[0]):
|
||||
if all_labels[i_pos] in bullish_clusters:
|
||||
gate[idx_i] = 1.0
|
||||
|
||||
# Final position: TSMOM gated by regime
|
||||
target = tsmom_pos * gate
|
||||
target = np.nan_to_num(target, nan=0.0)
|
||||
return target
|
||||
|
||||
|
||||
def run_config(n_clusters: int, lookback_months: int):
|
||||
name = f"STA04_k{n_clusters}_lb{lookback_months}m"
|
||||
fn = lambda df: make_target(df, n_clusters=n_clusters, lookback_months=lookback_months)
|
||||
rep = al.study_weights(name, fn, tfs=("1d", "12h"))
|
||||
return rep
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
# Grid: 2 configs x 2 TFs = 4 backtests per asset x 2 assets = 8 backtests total.
|
||||
# Keep it small: just 2 configs.
|
||||
configs = [
|
||||
(3, 3), # 3 clusters, 3-month lookback
|
||||
(4, 3), # 4 clusters, 3-month lookback
|
||||
]
|
||||
|
||||
best_rep = None
|
||||
best_score = -999.0
|
||||
|
||||
for n_clusters, lookback_months in configs:
|
||||
print(f"\n{'='*60}")
|
||||
print(f"CONFIG: n_clusters={n_clusters}, lookback_months={lookback_months}")
|
||||
print('='*60)
|
||||
rep = run_config(n_clusters, lookback_months)
|
||||
print(al.fmt(rep))
|
||||
print("JSON:", al.as_json(rep))
|
||||
|
||||
score = rep.get("verdict", {}).get("best_holdout_sharpe", -999.0) or -999.0
|
||||
if score > best_score:
|
||||
best_score = score
|
||||
best_rep = rep
|
||||
|
||||
print("\n" + "="*60)
|
||||
print("BEST CONFIG:")
|
||||
print("="*60)
|
||||
print(al.fmt(best_rep))
|
||||
print("JSON:", al.as_json(best_rep))
|
||||
Reference in New Issue
Block a user