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 +1,139 @@
|
||||
"""RSK07 — Drawdown-scaled exposure
|
||||
HYPOTHESIS: Exposure proportional to (1 - recent rolling drawdown) on a long-only base.
|
||||
De-risk into weakness: when the asset is in a large drawdown, reduce position size.
|
||||
|
||||
Style: al.study_weights (continuous position / vol-targeted)
|
||||
|
||||
Idea:
|
||||
- Compute the rolling drawdown over a lookback window.
|
||||
- Target exposure = (1 - drawdown_fraction) where drawdown_fraction in [0, 1].
|
||||
- Apply vol-targeting on top to keep risk constant.
|
||||
- Long-only base (no shorting).
|
||||
|
||||
The rolling drawdown at bar i = (rolling_max(close, dd_win) - close[i]) / rolling_max(close, dd_win)
|
||||
This is causal: uses close[i] and prior highs.
|
||||
|
||||
Exposure(i) = max(0, 1 - drawdown(i))
|
||||
With vol-targeting, this scales by (target_vol / realized_vol).
|
||||
|
||||
Small grid (<=4 configs, total backtests = 4 * 2 assets <= 8):
|
||||
A: dd_win=20, vol_target=0.20
|
||||
B: dd_win=60, vol_target=0.20
|
||||
C: dd_win=120, vol_target=0.20
|
||||
D: dd_win=60, vol_target=0.15
|
||||
|
||||
TFs tested: 1d, 12h (2 TFs * 4 configs * 2 assets = 16 total — but study_weights
|
||||
runs per config, so we do 4 configs across 2 TFs = 8 backtest calls)
|
||||
"""
|
||||
|
||||
import sys
|
||||
sys.path.insert(0, "/opt/docker/PythagorasGoal/scripts/research/alt")
|
||||
import altlib as al
|
||||
import numpy as np
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Core target function
|
||||
# ---------------------------------------------------------------------------
|
||||
def make_target(df, dd_win: int = 60, target_vol: float = 0.20) -> np.ndarray:
|
||||
"""
|
||||
Long-only drawdown-scaled exposure with vol-targeting.
|
||||
|
||||
Steps:
|
||||
1. Compute rolling max of close over dd_win bars (causal: max(close[i-dd_win:i+1]))
|
||||
2. Drawdown fraction = (rolling_max - close) / rolling_max
|
||||
3. Raw exposure = max(0, 1 - drawdown_fraction) in [0, 1]
|
||||
4. Apply vol-target scaling: multiply by (target_vol / realized_vol), cap at 2x
|
||||
5. Result: long-only position in [0, 2], decided with data <= close[i]
|
||||
"""
|
||||
c = df["close"].values.astype(float)
|
||||
n = len(c)
|
||||
|
||||
# Causal rolling maximum: max of close over [i-dd_win+1 .. i]
|
||||
# Use pandas rolling with min_periods=1
|
||||
c_series = df["close"].astype(float)
|
||||
roll_max = c_series.rolling(dd_win, min_periods=1).max().values
|
||||
|
||||
# Drawdown fraction (0 = at high-water mark, 1 = fully drawn down)
|
||||
dd_frac = np.where(roll_max > 0, (roll_max - c) / roll_max, 0.0)
|
||||
dd_frac = np.clip(dd_frac, 0.0, 1.0)
|
||||
|
||||
# Raw direction/size: (1 - drawdown), always long [0, 1]
|
||||
raw_exposure = 1.0 - dd_frac # 1.0 at HWM, 0.0 at full drawdown
|
||||
|
||||
# Vol-targeting: scale so expected volatility = target_vol
|
||||
# Use al.vol_target with direction=raw_exposure (already in [0,1])
|
||||
# But al.vol_target expects direction in {-1, 0, 1}; we'll do manual vol-scaling
|
||||
# Realized vol: rolling std of log returns
|
||||
log_ret = np.diff(np.log(c), prepend=np.nan)
|
||||
vol_win = int(30 * al.bars_per_day(df))
|
||||
vol_win = max(vol_win, 5)
|
||||
r_series = pd.Series(log_ret) if False else __import__('pandas').Series(log_ret)
|
||||
|
||||
# Realized vol: annualized
|
||||
log_ret_arr = al.log_returns(c)
|
||||
bpy = al.bars_per_year(df)
|
||||
rv = al.realized_vol(log_ret_arr, vol_win, bpy)
|
||||
|
||||
# Vol-target scaling
|
||||
lev = np.where((rv > 0) & np.isfinite(rv), target_vol / rv, 1.0)
|
||||
lev = np.clip(lev, 0.0, 2.0)
|
||||
|
||||
# Final target: drawdown-scaled exposure * vol lever
|
||||
target = raw_exposure * lev
|
||||
|
||||
# Cap at 2.0 (leverage cap)
|
||||
target = np.clip(target, 0.0, 2.0)
|
||||
|
||||
# First few bars: NaN until we have enough data
|
||||
warmup = max(dd_win, vol_win)
|
||||
target[:warmup] = np.nan
|
||||
|
||||
return target
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Grid search
|
||||
# ---------------------------------------------------------------------------
|
||||
import pandas as pd # noqa: E402 (needed above via __import__, explicit now)
|
||||
|
||||
GRID = [
|
||||
{"dd_win": 20, "target_vol": 0.20, "label": "dd=20 vol=20%"},
|
||||
{"dd_win": 60, "target_vol": 0.20, "label": "dd=60 vol=20%"},
|
||||
{"dd_win": 120, "target_vol": 0.20, "label": "dd=120 vol=20%"},
|
||||
{"dd_win": 60, "target_vol": 0.15, "label": "dd=60 vol=15%"},
|
||||
]
|
||||
|
||||
best_rep = None
|
||||
best_score = -999.0
|
||||
best_label = ""
|
||||
|
||||
for params in GRID:
|
||||
dd_win = params["dd_win"]
|
||||
target_vol = params["target_vol"]
|
||||
label = f"RSK07 {params['label']}"
|
||||
|
||||
print(f"\n--- Testing {label} ---")
|
||||
|
||||
rep = al.study_weights(
|
||||
label,
|
||||
lambda df, dw=dd_win, tv=target_vol: make_target(df, dd_win=dw, target_vol=tv),
|
||||
tfs=("1d", "12h"),
|
||||
)
|
||||
print(al.fmt(rep))
|
||||
|
||||
# Score by min-asset hold-out Sharpe
|
||||
best_cell = max(rep["cells"], key=lambda c: c.get("min_asset_holdout_sharpe", -9))
|
||||
score = best_cell.get("min_asset_holdout_sharpe", -9.0)
|
||||
if score > best_score:
|
||||
best_score = score
|
||||
best_rep = rep
|
||||
best_label = label
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Final report
|
||||
# ---------------------------------------------------------------------------
|
||||
print("\n" + "=" * 60)
|
||||
print(f"BEST CONFIG: {best_label}")
|
||||
print(al.fmt(best_rep))
|
||||
print("JSON:", al.as_json(best_rep))
|
||||
Reference in New Issue
Block a user