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 +1,165 @@
|
||||
"""CMB06 — Trend + Seasonality Combo
|
||||
IDEA: TSMOM long-flat (multi-horizon, vol-targeted, like TP01) but scale the
|
||||
exposure UP in historically strong calendar windows (day-of-week + month-of-year
|
||||
expanding expanding expectancy). Causal only: expectancy estimated on expanding window
|
||||
using data BEFORE the current bar.
|
||||
|
||||
Design:
|
||||
- Base signal: TSMOM multi-horizon (1M / 3M / 6M), long-flat, vote-then-sign
|
||||
- Volatility targeting: 20% target, 2x lev cap (same as TP01)
|
||||
- Seasonality multiplier: expand-window daily/monthly return expectancy,
|
||||
normalised to [scale_min, scale_max] so it's a scalar boost, not a flip.
|
||||
The multiplier is always >= 0 (never inverts the trend).
|
||||
|
||||
Causal guarantee:
|
||||
- Day-of-week expectancy at bar i uses only past bars (strict shift: computed on
|
||||
data up to bar i-1, applied at bar i).
|
||||
- Month-of-year same.
|
||||
- Both use EXPANDING window (not rolling) -> no future-data leak, and it
|
||||
gradually stabilises as history accumulates.
|
||||
|
||||
Grid (4 params): 2 scale ranges × 2 TFs = 4 cells total.
|
||||
"""
|
||||
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 _expanding_dow_expectancy(df: pd.DataFrame) -> np.ndarray:
|
||||
"""For each bar, return the expanding-window mean return of the same day-of-week,
|
||||
computed on PAST bars only (shift 1). Returns NaN until at least 4 samples exist."""
|
||||
c = df["close"].values.astype(float)
|
||||
r = al.simple_returns(c) # r[i] = return realized at bar i
|
||||
dt = pd.to_datetime(df["datetime"], utc=True)
|
||||
dow = dt.dt.dayofweek.values # 0=Mon..6=Sun
|
||||
|
||||
exp = np.full(len(r), np.nan)
|
||||
# For each bar i, compute mean return of same DOW for all bars j < i
|
||||
# Use expanding sum by DOW category
|
||||
dow_sum = np.zeros(7, dtype=float)
|
||||
dow_cnt = np.zeros(7, dtype=int)
|
||||
|
||||
for i in range(1, len(r)):
|
||||
# update with bar i-1 (strictly past)
|
||||
d_prev = dow[i - 1]
|
||||
dow_sum[d_prev] += r[i - 1]
|
||||
dow_cnt[d_prev] += 1
|
||||
|
||||
d = dow[i]
|
||||
if dow_cnt[d] >= 4:
|
||||
exp[i] = dow_sum[d] / dow_cnt[d]
|
||||
|
||||
return exp
|
||||
|
||||
|
||||
def _expanding_month_expectancy(df: pd.DataFrame) -> np.ndarray:
|
||||
"""Same but for month-of-year (1..12). Requires >= 4 past bars in same month."""
|
||||
c = df["close"].values.astype(float)
|
||||
r = al.simple_returns(c)
|
||||
dt = pd.to_datetime(df["datetime"], utc=True)
|
||||
moy = dt.dt.month.values # 1..12
|
||||
|
||||
exp = np.full(len(r), np.nan)
|
||||
mo_sum = np.zeros(13, dtype=float)
|
||||
mo_cnt = np.zeros(13, dtype=int)
|
||||
|
||||
for i in range(1, len(r)):
|
||||
m_prev = moy[i - 1]
|
||||
mo_sum[m_prev] += r[i - 1]
|
||||
mo_cnt[m_prev] += 1
|
||||
|
||||
m = moy[i]
|
||||
if mo_cnt[m] >= 4:
|
||||
exp[i] = mo_sum[m] / mo_cnt[m]
|
||||
|
||||
return exp
|
||||
|
||||
|
||||
def _seasonality_multiplier(df: pd.DataFrame, scale_min: float, scale_max: float) -> np.ndarray:
|
||||
"""Combine DOW + month expanding expectancy into a [scale_min, scale_max] multiplier.
|
||||
When either is NaN (early history), default to 1.0 (neutral)."""
|
||||
dow_exp = _expanding_dow_expectancy(df)
|
||||
mon_exp = _expanding_month_expectancy(df)
|
||||
|
||||
# Normalise each to [-1, +1] range using the expanding-window min/max seen so far.
|
||||
# We use a causal expanding percentile: zscore is simpler and avoids percentile loop.
|
||||
# Use zscore over an expanding window instead (pandas expanding).
|
||||
dow_s = pd.Series(dow_exp)
|
||||
mon_s = pd.Series(mon_exp)
|
||||
|
||||
# Causal z-score (expanding)
|
||||
dow_z = (dow_s - dow_s.expanding().mean()) / dow_s.expanding().std().replace(0, np.nan)
|
||||
mon_z = (mon_s - mon_s.expanding().mean()) / mon_s.expanding().std().replace(0, np.nan)
|
||||
|
||||
# Blend (equal weight)
|
||||
combined = (dow_z.fillna(0.0) + mon_z.fillna(0.0)).values / 2.0
|
||||
|
||||
# Map to [scale_min, scale_max] via sigmoid-like clamp
|
||||
# clip to [-2, 2] sigma, then linearly map
|
||||
combined_clipped = np.clip(combined, -2.0, 2.0)
|
||||
mid = (scale_min + scale_max) / 2.0
|
||||
half_range = (scale_max - scale_min) / 2.0
|
||||
mult = mid + half_range * (combined_clipped / 2.0)
|
||||
|
||||
# Where both were NaN (very early bars), use neutral = 1.0
|
||||
both_nan = dow_s.isna().values & mon_s.isna().values
|
||||
mult[both_nan] = 1.0
|
||||
|
||||
return mult
|
||||
|
||||
|
||||
def _tsmom_base(df: pd.DataFrame) -> np.ndarray:
|
||||
"""Multi-horizon TSMOM: 1M/3M/6M vote, long-flat, vol-targeted."""
|
||||
c = df["close"].values.astype(float)
|
||||
bpd = al.bars_per_day(df)
|
||||
d = np.zeros(len(c))
|
||||
for months in (1, 3, 6):
|
||||
h = int(months * 30 * bpd)
|
||||
if h >= len(c):
|
||||
continue
|
||||
s = np.full(len(c), np.nan)
|
||||
s[h:] = np.sign(c[h:] / c[:-h] - 1.0)
|
||||
d = d + np.nan_to_num(s)
|
||||
direction = np.clip(np.sign(d), 0, None) # long-flat only
|
||||
return al.vol_target(direction, df, target_vol=0.20, vol_win_days=30, leverage_cap=2.0)
|
||||
|
||||
|
||||
def make_target(scale_min: float, scale_max: float):
|
||||
"""Return a target_fn that applies the seasonality multiplier."""
|
||||
def target_fn(df: pd.DataFrame) -> np.ndarray:
|
||||
base = _tsmom_base(df)
|
||||
mult = _seasonality_multiplier(df, scale_min, scale_max)
|
||||
combined = base * mult
|
||||
# Keep within leverage cap
|
||||
combined = np.clip(combined, 0.0, 2.0)
|
||||
combined = np.nan_to_num(combined, nan=0.0)
|
||||
return combined
|
||||
return target_fn
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
# Grid: 2 scale ranges × 2 TFs = 4 cells
|
||||
# scale_min/max: how much to reduce/boost position in weak/strong seasons
|
||||
# (0.5, 1.5) = modest 50% swing; (0.25, 1.75) = aggressive 150% swing
|
||||
configs = [
|
||||
("CMB06-modest", 0.5, 1.5),
|
||||
("CMB06-aggr", 0.25, 1.75),
|
||||
]
|
||||
|
||||
all_reps = []
|
||||
for name, smin, smax in configs:
|
||||
print(f"\n=== Running {name} (scale [{smin},{smax}]) ===")
|
||||
rep = al.study_weights(name, make_target(smin, smax), tfs=("1d", "12h"))
|
||||
print(al.fmt(rep))
|
||||
all_reps.append((name, rep))
|
||||
|
||||
# Pick best by min_asset_holdout_sharpe at best TF
|
||||
def best_holdout(rep):
|
||||
return max(c["min_asset_holdout_sharpe"] for c in rep["cells"])
|
||||
|
||||
best_name, best_rep = max(all_reps, key=lambda x: best_holdout(x[1]))
|
||||
print(f"\n>>> BEST CONFIG: {best_name}")
|
||||
print(al.fmt(best_rep))
|
||||
print("JSON:", al.as_json(best_rep))
|
||||
Reference in New Issue
Block a user