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 +1,278 @@
|
||||
"""XAS02 — ETH/BTC ratio momentum (TSMOM on the spread).
|
||||
|
||||
IDEA: Trend-follow the ETH/BTC ratio using time-series momentum (TSMOM).
|
||||
If the ETH/BTC ratio is rising (ETH outperforming BTC), go long ETH / short BTC.
|
||||
If it's falling, go short ETH / long BTC (or flat, long-only variant).
|
||||
We use multi-horizon momentum blending (30/90/180 day lookbacks) and vol-target.
|
||||
|
||||
IMPLEMENTATION NOTE:
|
||||
- The ratio ETH/BTC is constructed from the two certified price series.
|
||||
- Each asset gets a position: +1 on the outperformer, -1 on underperformer (market-neutral)
|
||||
OR long-flat (long outperformer, flat underperformer).
|
||||
- We test 4 parameter configs on 2 TFs = 8 backtests total (fits in 2-CPU budget).
|
||||
|
||||
For the multi-asset study_weights framework, we run each asset independently
|
||||
but the TARGET for each asset is derived from the ETH/BTC ratio signal:
|
||||
- BTC target: -1 * ratio_signal (short BTC when ETH is outperforming)
|
||||
- ETH target: +1 * ratio_signal (long ETH when ETH is outperforming)
|
||||
|
||||
We test both market-neutral (clip to [-1, 1]) and long-flat (clip to [0, 1]) variants,
|
||||
plus different momentum horizons.
|
||||
"""
|
||||
import sys
|
||||
sys.path.insert(0, "/opt/docker/PythagorasGoal/scripts/research/alt")
|
||||
import altlib as al
|
||||
import numpy as np
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# Core ratio momentum function
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
def make_ratio_target(horizons=(30, 90, 180), long_flat=False, vol_tgt=True):
|
||||
"""
|
||||
Build a target function for one asset given:
|
||||
- horizons: lookback days for TSMOM on the ETH/BTC ratio
|
||||
- long_flat: if True, clip to [0, 1] (long-flat); if False, [-1, 1] (market-neutral)
|
||||
- vol_tgt: apply vol targeting
|
||||
|
||||
Returns a tuple (btc_fn, eth_fn) where each fn(df) -> target array.
|
||||
"""
|
||||
def make_fn(asset):
|
||||
def fn(df):
|
||||
# Load BTC and ETH at the same TF
|
||||
# We need to infer the TF from df's bar spacing
|
||||
dt_secs = (df["datetime"].iloc[-1] - df["datetime"].iloc[0]).total_seconds() / (len(df) - 1)
|
||||
if dt_secs < 3600 * 2:
|
||||
tf_str = "1h"
|
||||
elif dt_secs < 3600 * 6:
|
||||
tf_str = "4h"
|
||||
elif dt_secs < 3600 * 10:
|
||||
tf_str = "8h"
|
||||
elif dt_secs < 3600 * 14:
|
||||
tf_str = "12h"
|
||||
else:
|
||||
tf_str = "1d"
|
||||
|
||||
btc = al.get("BTC", tf_str)
|
||||
eth = al.get("ETH", tf_str)
|
||||
|
||||
# Align on timestamps (inner join)
|
||||
import pandas as pd
|
||||
btc_s = pd.Series(btc["close"].values, index=btc["timestamp"].values)
|
||||
eth_s = pd.Series(eth["close"].values, index=eth["timestamp"].values)
|
||||
common = btc_s.index.intersection(eth_s.index)
|
||||
btc_c = btc_s.loc[common].values
|
||||
eth_c = eth_s.loc[common].values
|
||||
|
||||
# Build the ETH/BTC ratio
|
||||
ratio = eth_c / btc_c
|
||||
|
||||
bpd = al.bars_per_day(btc)
|
||||
|
||||
# Multi-horizon TSMOM on the ratio
|
||||
signal = np.zeros(len(ratio))
|
||||
for h_days in horizons:
|
||||
h = int(h_days * bpd)
|
||||
if h >= len(ratio):
|
||||
continue
|
||||
s = np.full(len(ratio), np.nan)
|
||||
s[h:] = np.sign(ratio[h:] / ratio[:-h] - 1)
|
||||
signal = signal + np.nan_to_num(s)
|
||||
|
||||
# Normalize to [-1, 1]
|
||||
max_votes = len(horizons)
|
||||
signal = signal / max_votes # now in [-1, 1]
|
||||
|
||||
# Long-flat: only take the outperformer side
|
||||
if long_flat:
|
||||
signal = np.clip(signal, 0, 1)
|
||||
|
||||
# Map to this asset's direction:
|
||||
# ETH/BTC ratio up -> long ETH, short BTC
|
||||
if asset == "ETH":
|
||||
dir_signal = signal
|
||||
else: # BTC
|
||||
dir_signal = -signal
|
||||
|
||||
# Align to df (which may be btc or eth df)
|
||||
# df timestamps may include bars not in common — need to align
|
||||
df_ts = df["timestamp"].values
|
||||
# Create a full signal array aligned to df
|
||||
aligned = np.zeros(len(df_ts))
|
||||
# Map common timestamps back to df indices
|
||||
common_set = set(common.tolist())
|
||||
ts_to_signal = dict(zip(common.tolist(), dir_signal.tolist()))
|
||||
for i, ts in enumerate(df_ts):
|
||||
if ts in ts_to_signal:
|
||||
aligned[i] = ts_to_signal[ts]
|
||||
|
||||
# Apply vol targeting on the current asset's df
|
||||
if vol_tgt:
|
||||
return al.vol_target(aligned, df, target_vol=0.20, vol_win_days=30, leverage_cap=2.0)
|
||||
else:
|
||||
return aligned
|
||||
|
||||
return fn
|
||||
|
||||
return make_fn("BTC"), make_fn("ETH")
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# Internal mini-grid: 4 configs on 2 TFs = 8 backtests
|
||||
# Config A: multi-horizon [30,90,180] long-flat vol-targeted
|
||||
# Config B: multi-horizon [30,90,180] market-neutral vol-targeted
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
configs = {
|
||||
"A_longflat": {
|
||||
"horizons": (30, 90, 180),
|
||||
"long_flat": True,
|
||||
"vol_tgt": True,
|
||||
"desc": "multi-hz [30,90,180] long-flat vol-target"
|
||||
},
|
||||
"B_neutral": {
|
||||
"horizons": (30, 90, 180),
|
||||
"long_flat": False,
|
||||
"vol_tgt": True,
|
||||
"desc": "multi-hz [30,90,180] market-neutral vol-target"
|
||||
},
|
||||
}
|
||||
|
||||
# We need a custom study function because both assets use the ratio signal
|
||||
# (not independent signals). But the library's study_weights passes the same
|
||||
# target_fn to each asset. We hack this by using closures that look up the
|
||||
# correct asset direction.
|
||||
|
||||
def run_config(config_name, cfg):
|
||||
btc_fn, eth_fn = make_ratio_target(
|
||||
horizons=cfg["horizons"],
|
||||
long_flat=cfg["long_flat"],
|
||||
vol_tgt=cfg["vol_tgt"]
|
||||
)
|
||||
|
||||
# Run on both TFs
|
||||
results_by_tf = {}
|
||||
for tf in ("1d", "12h"):
|
||||
df_btc = al.get("BTC", tf)
|
||||
df_eth = al.get("ETH", tf)
|
||||
|
||||
tgt_btc = btc_fn(df_btc)
|
||||
tgt_eth = eth_fn(df_eth)
|
||||
|
||||
res_btc = al.eval_weights(df_btc, tgt_btc)
|
||||
res_eth = al.eval_weights(df_eth, tgt_eth)
|
||||
|
||||
# Fee sweep for both
|
||||
sweep_btc = {}
|
||||
sweep_eth = {}
|
||||
for f in al.FEE_SWEEP:
|
||||
key = f"{2*f*100:.2f}%RT"
|
||||
sweep_btc[key] = al.eval_weights(df_btc, tgt_btc, fee_side=f)["full"]["sharpe"]
|
||||
sweep_eth[key] = al.eval_weights(df_eth, tgt_eth, fee_side=f)["full"]["sharpe"]
|
||||
|
||||
fee_ok_btc = sweep_btc.get("0.20%RT", -9) > 0
|
||||
fee_ok_eth = sweep_eth.get("0.20%RT", -9) > 0
|
||||
|
||||
min_full = min(res_btc["full"]["sharpe"], res_eth["full"]["sharpe"])
|
||||
min_hold = min(
|
||||
res_btc["holdout"].get("sharpe", 0.0),
|
||||
res_eth["holdout"].get("sharpe", 0.0)
|
||||
)
|
||||
|
||||
results_by_tf[tf] = {
|
||||
"tf": tf,
|
||||
"min_asset_full_sharpe": round(min_full, 3),
|
||||
"min_asset_holdout_sharpe": round(min_hold, 3),
|
||||
"fee_survives": fee_ok_btc and fee_ok_eth,
|
||||
"full_sharpe": round((res_btc["full"]["sharpe"] + res_eth["full"]["sharpe"]) / 2, 3),
|
||||
"per_asset": {
|
||||
"BTC": {
|
||||
"full": res_btc["full"],
|
||||
"holdout": res_btc["holdout"],
|
||||
"tim": res_btc["time_in_market"],
|
||||
"turnover": res_btc["turnover_per_year"],
|
||||
"fee_sweep": sweep_btc,
|
||||
"yearly": res_btc["yearly"]
|
||||
},
|
||||
"ETH": {
|
||||
"full": res_eth["full"],
|
||||
"holdout": res_eth["holdout"],
|
||||
"tim": res_eth["time_in_market"],
|
||||
"turnover": res_eth["turnover_per_year"],
|
||||
"fee_sweep": sweep_eth,
|
||||
"yearly": res_eth["yearly"]
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return results_by_tf
|
||||
|
||||
|
||||
print("XAS02 — ETH/BTC Ratio Momentum (TSMOM on ratio)")
|
||||
print("="*60)
|
||||
|
||||
all_results = {}
|
||||
for cfg_name, cfg in configs.items():
|
||||
print(f"\nRunning config {cfg_name}: {cfg['desc']} ...")
|
||||
all_results[cfg_name] = run_config(cfg_name, cfg)
|
||||
for tf, cell in all_results[cfg_name].items():
|
||||
print(f" TF={tf}: minFull={cell['min_asset_full_sharpe']:+.2f} "
|
||||
f"minHold={cell['min_asset_holdout_sharpe']:+.2f} "
|
||||
f"feeOK={cell['fee_survives']}")
|
||||
for a, pa in cell["per_asset"].items():
|
||||
print(f" {a}: full Sh={pa['full']['sharpe']:+.2f} "
|
||||
f"DD={pa['full']['maxdd']*100:.0f}% "
|
||||
f"hold Sh={pa['holdout'].get('sharpe', 0):+.2f}")
|
||||
|
||||
# Pick best config (best min_asset_holdout_sharpe)
|
||||
best_cfg = None
|
||||
best_score = -999
|
||||
best_tf_name = None
|
||||
for cfg_name, tf_results in all_results.items():
|
||||
for tf, cell in tf_results.items():
|
||||
score = cell["min_asset_holdout_sharpe"]
|
||||
if score > best_score:
|
||||
best_score = score
|
||||
best_cfg = cfg_name
|
||||
best_tf_name = tf
|
||||
|
||||
print(f"\nBest config: {best_cfg} @ TF={best_tf_name}")
|
||||
|
||||
# Build the final report in al format
|
||||
best_cells = []
|
||||
for cfg_name, tf_results in all_results.items():
|
||||
for tf, cell in tf_results.items():
|
||||
if cfg_name == best_cfg:
|
||||
best_cells.append(cell)
|
||||
|
||||
# Use a simple verdict function
|
||||
def verdict(cells):
|
||||
if not cells:
|
||||
return {"grade": "FAIL", "reason": "no cells"}
|
||||
best = max(cells, key=lambda c: c.get("min_asset_holdout_sharpe", -9))
|
||||
pass_ = (best.get("min_asset_full_sharpe", -9) >= 0.5 and
|
||||
best.get("min_asset_holdout_sharpe", -9) >= 0.2 and
|
||||
best.get("fee_survives", False))
|
||||
weak = (best.get("min_asset_full_sharpe", -9) >= 0.3 and
|
||||
best.get("min_asset_holdout_sharpe", -9) >= 0.0)
|
||||
grade = "PASS" if pass_ else ("WEAK" if weak else "FAIL")
|
||||
return {
|
||||
"grade": grade,
|
||||
"best_tf": best.get("tf"),
|
||||
"best_full_sharpe": best.get("min_asset_full_sharpe"),
|
||||
"best_holdout_sharpe": best.get("min_asset_holdout_sharpe"),
|
||||
"n_positive_cells": sum(1 for c in cells if c.get("min_asset_full_sharpe", -9) > 0),
|
||||
"n_cells": len(cells)
|
||||
}
|
||||
|
||||
rep = {
|
||||
"name": "XAS02",
|
||||
"kind": "weights",
|
||||
"cells": best_cells,
|
||||
"verdict": verdict(best_cells),
|
||||
"best_config": best_cfg,
|
||||
"configs_tested": list(configs.keys()),
|
||||
}
|
||||
|
||||
print("\n" + al.fmt(rep))
|
||||
print("JSON:", al.as_json(rep))
|
||||
Reference in New Issue
Block a user