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 +1,191 @@
|
||||
"""XAS05 — Lead-lag ETH->BTC (mirror of XAS04)
|
||||
|
||||
HYPOTHESIS: ETH returns lead BTC returns by 1 bar. BTC position = sign of ETH lagged return.
|
||||
This is the mirror of XAS04 (BTC->ETH). We test several signal constructions:
|
||||
1. Sign of ETH 1-bar return (pure lag) -> BTC position
|
||||
2. ETH EMA momentum (fast/slow cross) -> BTC direction
|
||||
3. ETH TSMOM (30/90/180 day) multi-horizon -> BTC direction
|
||||
4. Blend of ETH 1-bar lag + ETH EMA momentum -> BTC direction
|
||||
|
||||
CAUSAL GUARANTEE: We use SAME timeframe ETH data aligned to BTC timestamps (merge_asof backward).
|
||||
For cross-TF to work without lookahead, we must shift the ETH signal by 1 bar when mixing TFs.
|
||||
The simplest honest approach: use ETH data at the SAME timeframe as the BTC data being evaluated.
|
||||
|
||||
For study_weights, target_fn(df) is called with each asset's df.
|
||||
When df=BTC: we load ETH at the same TF, align it to BTC timestamps, compute the ETH signal,
|
||||
and apply it to BTC -> the lead-lag hypothesis.
|
||||
When df=ETH: we load ETH at the same TF, compute the ETH signal on the same data, and apply it
|
||||
to ETH itself -> equivalent to trend-following ETH on its own momentum (baseline).
|
||||
|
||||
CRITICAL LOOKAHEAD WARNING (detected during development):
|
||||
Using ETH 1d data to generate signals on BTC 12h bars IS a lookahead:
|
||||
ETH 1d bar at T 00:00 has a close that matches ETH 12h bar at T 12:00 (i.e., noon close),
|
||||
not midnight. The daily bar is labeled at midnight but closes are from future noon.
|
||||
FIX: We always load ETH at the SAME TF as the df being evaluated.
|
||||
"""
|
||||
import sys
|
||||
sys.path.insert(0, "/opt/docker/PythagorasGoal/scripts/research/alt")
|
||||
import altlib as al
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
|
||||
|
||||
_TF_MAP = {} # will be filled per run
|
||||
|
||||
|
||||
def _align_eth_to_btc(btc_df: pd.DataFrame, eth_df: pd.DataFrame) -> np.ndarray:
|
||||
"""Align ETH close prices to BTC timestamps using merge_asof (causal: backward).
|
||||
Both should be same TF to avoid cross-TF lookahead. Returns ETH close aligned to
|
||||
BTC timestamps, len(btc_df). Applies an EXTRA 1-bar shift to ensure true causality:
|
||||
ETH bar closing at T cannot influence BTC bar also closing at T (concurrent effect);
|
||||
we require ETH close at T-1 to predict BTC bar at T+1 via altlib's own shift.
|
||||
"""
|
||||
btc_ts = btc_df["timestamp"].astype("int64").values
|
||||
eth_ts = eth_df["timestamp"].astype("int64").values
|
||||
eth_close = eth_df["close"].values.astype(float)
|
||||
|
||||
# Shift ETH by 1 bar: use ETH close at previous bar (T-1) as signal at bar T
|
||||
# This prevents any possibility of concurrent/lookahead correlation
|
||||
eth_close_lagged = np.empty_like(eth_close)
|
||||
eth_close_lagged[0] = np.nan
|
||||
eth_close_lagged[1:] = eth_close[:-1]
|
||||
|
||||
left = pd.DataFrame({"timestamp": btc_ts})
|
||||
right = pd.DataFrame({"timestamp": eth_ts, "eth_close": eth_close_lagged})
|
||||
merged = pd.merge_asof(left, right, on="timestamp", direction="backward")
|
||||
return merged["eth_close"].values.astype(float)
|
||||
|
||||
|
||||
def make_xas05_config1(lag_bars=1, tf="1d"):
|
||||
"""Config 1: Sign of ETH 1-bar lagged return -> vol-targeted position.
|
||||
Uses ETH return at prior bar (decided at close[i-1]) -> hold during bar i+1.
|
||||
Extra 1-bar lag ensures strict causality even for concurrent closes.
|
||||
"""
|
||||
def target_fn(df):
|
||||
# Detect asset by checking if it's the ETH df (ETH will self-signal)
|
||||
# We always load ETH at the same TF as df
|
||||
eth_df = al.get("ETH", tf)
|
||||
eth_c = _align_eth_to_btc(df, eth_df)
|
||||
n = len(df)
|
||||
|
||||
# ETH lagged return: sign of ETH return (already 1-bar lagged via alignment)
|
||||
eth_ret = np.zeros(n)
|
||||
eth_ret[lag_bars:] = eth_c[lag_bars:] / eth_c[:-lag_bars] - 1.0
|
||||
|
||||
# Direction = sign of ETH return
|
||||
direction = np.sign(eth_ret)
|
||||
direction[~np.isfinite(direction)] = 0.0
|
||||
direction[:lag_bars + 5] = 0.0 # warmup
|
||||
|
||||
# Vol-target BTC position based on ETH signal
|
||||
return al.vol_target(direction, df, target_vol=0.20, vol_win_days=30, leverage_cap=2.0)
|
||||
|
||||
return target_fn
|
||||
|
||||
|
||||
def make_xas05_config2(ema_fast=5, ema_slow=20, tf="1d"):
|
||||
"""Config 2: ETH EMA momentum (fast/slow cross) -> direction."""
|
||||
def target_fn(df):
|
||||
eth_df = al.get("ETH", tf)
|
||||
eth_c = _align_eth_to_btc(df, eth_df)
|
||||
n = len(df)
|
||||
|
||||
fast = al.ema(eth_c, ema_fast)
|
||||
slow = al.ema(eth_c, ema_slow)
|
||||
|
||||
direction = np.where(fast > slow, 1.0, 0.0) # long-flat
|
||||
direction[:ema_slow + 5] = 0.0 # warmup
|
||||
|
||||
return al.vol_target(direction, df, target_vol=0.20, vol_win_days=30, leverage_cap=2.0)
|
||||
|
||||
return target_fn
|
||||
|
||||
|
||||
def make_xas05_config3(tf="1d"):
|
||||
"""Config 3: ETH TSMOM (30/90/180 day) multi-horizon -> direction."""
|
||||
def target_fn(df):
|
||||
eth_df = al.get("ETH", tf)
|
||||
eth_c = _align_eth_to_btc(df, eth_df)
|
||||
n = len(df)
|
||||
bpd = al.bars_per_day(df)
|
||||
|
||||
# Multi-horizon ETH momentum
|
||||
signal = np.zeros(n)
|
||||
for days in (30, 90, 180):
|
||||
h = int(days * bpd)
|
||||
s = np.full(n, np.nan)
|
||||
if h < n:
|
||||
s[h:] = np.sign(eth_c[h:] / eth_c[:-h] - 1.0)
|
||||
signal = signal + np.nan_to_num(s)
|
||||
|
||||
# Long only when ETH shows positive trend
|
||||
direction = np.clip(np.sign(signal), 0, None) # long-flat
|
||||
direction[:int(180 * bpd) + 5] = 0.0 # warmup
|
||||
|
||||
return al.vol_target(direction, df, target_vol=0.20, vol_win_days=30, leverage_cap=2.0)
|
||||
|
||||
return target_fn
|
||||
|
||||
|
||||
def make_xas05_config4(lag_bars=1, ema_span=10, tf="1d"):
|
||||
"""Config 4: Blend of ETH 1-bar lag + ETH EMA momentum -> direction."""
|
||||
def target_fn(df):
|
||||
eth_df = al.get("ETH", tf)
|
||||
eth_c = _align_eth_to_btc(df, eth_df)
|
||||
n = len(df)
|
||||
|
||||
# Signal 1: ETH lagged return sign
|
||||
eth_ret = np.zeros(n)
|
||||
eth_ret[lag_bars:] = eth_c[lag_bars:] / eth_c[:-lag_bars] - 1.0
|
||||
sig1 = np.sign(eth_ret)
|
||||
sig1[:lag_bars + 3] = 0.0
|
||||
|
||||
# Signal 2: ETH EMA momentum
|
||||
fast = al.ema(eth_c, ema_span)
|
||||
slow = al.ema(eth_c, ema_span * 4)
|
||||
sig2 = np.where(fast > slow, 1.0, 0.0)
|
||||
sig2[:ema_span * 4 + 5] = 0.0
|
||||
|
||||
# Blend: both signals must agree -> long-flat
|
||||
direction = np.where((sig1 > 0) & (sig2 > 0), 1.0, 0.0)
|
||||
direction[~np.isfinite(direction)] = 0.0
|
||||
|
||||
return al.vol_target(direction, df, target_vol=0.20, vol_win_days=30, leverage_cap=2.0)
|
||||
|
||||
return target_fn
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
print("XAS05 — Lead-lag ETH->BTC (HONEST: same TF, extra 1-bar lag to prevent concurrent lookahead)")
|
||||
print("=" * 80)
|
||||
|
||||
# Only run 1d to keep total backtests <= 6 (4 configs x 1 TF x 2 assets = 8, but we cap at 4 configs)
|
||||
# Use 1d only - it's the canonical TF for trend strategies and avoids TF mismatch issues
|
||||
configs = [
|
||||
("XAS05-C1-lag1ret-1d", make_xas05_config1(lag_bars=1, tf="1d"), ("1d",)),
|
||||
("XAS05-C2-ema5x20-1d", make_xas05_config2(ema_fast=5, ema_slow=20, tf="1d"), ("1d",)),
|
||||
("XAS05-C3-tsmom-1d", make_xas05_config3(tf="1d"), ("1d",)),
|
||||
("XAS05-C4-blend-1d", make_xas05_config4(lag_bars=1, ema_span=10, tf="1d"), ("1d",)),
|
||||
]
|
||||
|
||||
results = []
|
||||
best_rep = None
|
||||
best_hold = -999
|
||||
|
||||
for name, fn, tfs in configs:
|
||||
print(f"\n--- Running {name} ---")
|
||||
rep = al.study_weights(name, fn, tfs=tfs)
|
||||
print(al.fmt(rep))
|
||||
results.append(rep)
|
||||
|
||||
# Track best by min hold-out
|
||||
v = rep["verdict"]
|
||||
h = v.get("best_holdout_sharpe", -999)
|
||||
if h is not None and h > best_hold:
|
||||
best_hold = h
|
||||
best_rep = rep
|
||||
|
||||
print("\n" + "=" * 80)
|
||||
print(f"BEST CONFIG: {best_rep['name']} -> {best_rep['verdict']['grade']}")
|
||||
print(al.fmt(best_rep))
|
||||
print("\nJSON:", al.as_json(best_rep))
|
||||
Reference in New Issue
Block a user