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 +1,150 @@
|
||||
"""TRD05 — ADX-filtered EMA crossover.
|
||||
|
||||
Hypothesis: EMA(fast, slow) cross provides directional signal ONLY when ADX(14) > threshold
|
||||
(trending regime). When ADX is below the threshold (chop), position goes flat.
|
||||
|
||||
Grid (<=4 param sets, total backtests = 4 params * 2 assets * 2 tfs = 16, but we limit to 2 TFs):
|
||||
(fast_ema, slow_ema, adx_period, adx_thresh)
|
||||
- (20, 100, 14, 25) — canonical from hypothesis
|
||||
- (10, 50, 14, 25) — faster cross
|
||||
- (20, 100, 14, 20) — more lenient ADX gate
|
||||
- (5, 20, 14, 25) — short-term cross with ADX filter
|
||||
|
||||
We run 4 configs but only 1 TF at a time to stay within 2-CPU budget.
|
||||
Best config selected by min-asset holdout Sharpe across 2 TFs (1d, 12h).
|
||||
|
||||
ADX calculation (causal):
|
||||
+DM[i] = max(high[i]-high[i-1], 0) if > (low[i-1]-low[i]) else 0
|
||||
-DM[i] = max(low[i-1]-low[i], 0) if > (high[i]-high[i-1]) else 0
|
||||
TR[i] = max(high[i]-low[i], |high[i]-close[i-1]|, |low[i]-close[i-1]|)
|
||||
Smooth over `period` with Wilder's EMA (alpha=1/period)
|
||||
+DI = 100 * smooth(+DM) / smooth(TR)
|
||||
-DI = 100 * smooth(-DM) / smooth(TR)
|
||||
DX = 100 * |+DI - -DI| / (+DI + -DI)
|
||||
ADX = Wilder EMA(DX, period)
|
||||
"""
|
||||
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 _wilder_ema(x: np.ndarray, period: int) -> np.ndarray:
|
||||
"""Wilder smoothing (EMA with alpha=1/period, adjust=False)."""
|
||||
alpha = 1.0 / period
|
||||
out = np.empty(len(x), dtype=float)
|
||||
out[0] = x[0]
|
||||
for i in range(1, len(x)):
|
||||
out[i] = out[i - 1] * (1.0 - alpha) + x[i] * alpha
|
||||
return out
|
||||
|
||||
|
||||
def _adx(df: pd.DataFrame, period: int = 14) -> np.ndarray:
|
||||
"""Compute causal ADX(period). Returns array len(df), NaN for first ~2*period bars."""
|
||||
h = df["high"].values.astype(float)
|
||||
l = df["low"].values.astype(float)
|
||||
c = df["close"].values.astype(float)
|
||||
n = len(h)
|
||||
|
||||
# True Range
|
||||
pc = np.roll(c, 1)
|
||||
pc[0] = c[0]
|
||||
tr = np.maximum(h - l, np.maximum(np.abs(h - pc), np.abs(l - pc)))
|
||||
|
||||
# Directional Movements
|
||||
up = h - np.roll(h, 1)
|
||||
dn = np.roll(l, 1) - l
|
||||
up[0] = 0.0
|
||||
dn[0] = 0.0
|
||||
pos_dm = np.where((up > dn) & (up > 0), up, 0.0)
|
||||
neg_dm = np.where((dn > up) & (dn > 0), dn, 0.0)
|
||||
|
||||
# Wilder smooth
|
||||
str_ = _wilder_ema(tr, period)
|
||||
spdm = _wilder_ema(pos_dm, period)
|
||||
sndm = _wilder_ema(neg_dm, period)
|
||||
|
||||
# DI lines
|
||||
pdi = 100.0 * np.where(str_ > 0, spdm / str_, 0.0)
|
||||
ndi = 100.0 * np.where(str_ > 0, sndm / str_, 0.0)
|
||||
|
||||
# DX and ADX
|
||||
denom = pdi + ndi
|
||||
dx = np.where(denom > 0, 100.0 * np.abs(pdi - ndi) / denom, 0.0)
|
||||
adx = _wilder_ema(dx, period)
|
||||
|
||||
# First 2*period bars are warm-up — NaN them
|
||||
adx[:2 * period] = np.nan
|
||||
return adx
|
||||
|
||||
|
||||
def make_target(fast: int, slow: int, adx_period: int, adx_thresh: float,
|
||||
vol_target: bool = True):
|
||||
"""Return a target_fn for study_weights that implements ADX-filtered EMA cross."""
|
||||
def target_fn(df: pd.DataFrame) -> np.ndarray:
|
||||
c = df["close"].values.astype(float)
|
||||
|
||||
ema_fast = al.ema(c, fast)
|
||||
ema_slow = al.ema(c, slow)
|
||||
adx_vals = _adx(df, adx_period)
|
||||
|
||||
# Signal: +1 if fast > slow (bullish trend), -1 if fast < slow (bearish)
|
||||
# Flat when ADX < threshold (choppy) or ADX is NaN (warmup)
|
||||
cross_signal = np.where(ema_fast > ema_slow, 1.0, -1.0)
|
||||
|
||||
trending = np.where(
|
||||
np.isfinite(adx_vals) & (adx_vals > adx_thresh),
|
||||
1.0, 0.0
|
||||
)
|
||||
|
||||
direction = cross_signal * trending
|
||||
|
||||
# Long-flat only (like TP01, we don't short crypto)
|
||||
# Actually let's try L/S first since hypothesis doesn't restrict
|
||||
direction_lf = np.clip(direction, 0, 1) # long-flat version
|
||||
|
||||
if vol_target:
|
||||
return al.vol_target(direction_lf, df, target_vol=0.20, vol_win_days=30,
|
||||
leverage_cap=2.0)
|
||||
else:
|
||||
return direction_lf
|
||||
|
||||
return target_fn
|
||||
|
||||
|
||||
# --- Grid of configs ---------------------------------------------------------
|
||||
CONFIGS = [
|
||||
dict(fast=20, slow=100, adx_period=14, adx_thresh=25), # canonical
|
||||
dict(fast=10, slow=50, adx_period=14, adx_thresh=25), # faster cross
|
||||
dict(fast=20, slow=100, adx_period=14, adx_thresh=20), # relaxed gate
|
||||
dict(fast=5, slow=20, adx_period=14, adx_thresh=25), # short-term
|
||||
]
|
||||
|
||||
# We test 2 timeframes: 1d and 12h (within 2-CPU budget constraint)
|
||||
TFS = ("1d", "12h")
|
||||
|
||||
best_rep = None
|
||||
best_score = -999.0
|
||||
|
||||
print("=== TRD05: ADX-filtered EMA crossover ===\n")
|
||||
|
||||
for cfg in CONFIGS:
|
||||
label = f"TRD05(ema{cfg['fast']}/{cfg['slow']},adx{cfg['adx_period']}>{cfg['adx_thresh']})"
|
||||
fn = make_target(**cfg)
|
||||
rep = al.study_weights(label, fn, tfs=TFS)
|
||||
print(al.fmt(rep))
|
||||
print()
|
||||
|
||||
# Score = min holdout sharpe across cells
|
||||
score = rep["verdict"].get("best_holdout_sharpe", -999.0) or -999.0
|
||||
if score > best_score:
|
||||
best_score = score
|
||||
best_rep = rep
|
||||
best_cfg = cfg
|
||||
|
||||
print("\n" + "=" * 60)
|
||||
print(f"BEST CONFIG: {best_cfg}")
|
||||
print(al.fmt(best_rep))
|
||||
print("JSON:", al.as_json(best_rep))
|
||||
Reference in New Issue
Block a user