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 +1,166 @@
|
||||
"""STA07 — Online SGD Logistic Regression (next-bar sign prediction)
|
||||
Hypothesis: An online logistic classifier (sklearn SGDClassifier with partial_fit) is
|
||||
updated bar-by-bar using causal features and predicts the sign of the NEXT bar's return.
|
||||
The prediction confidence (decision_function score) is used as a continuous position
|
||||
(long if positive score, short/flat if negative — but long-only via clip to [0,1]).
|
||||
|
||||
Features (all causal at bar i):
|
||||
- short EMA vs long EMA ratio (trend)
|
||||
- RSI(14) normalized to [-1,1]
|
||||
- z-score of close over 20 bars
|
||||
- realized vol ratio (fast / slow) as regime indicator
|
||||
- log return of last bar (momentum/mean-reversion signal)
|
||||
- ATR normalized (relative volatility)
|
||||
|
||||
The label for bar i is: sign(close[i+1] / close[i] - 1)
|
||||
-> at decision time i we don't have i+1 yet, but we use PAST labels to train.
|
||||
-> Specifically, we do partial_fit at bar i using features[i-1] and label[i-1]
|
||||
(the actual outcome that just resolved), then predict at bar i using features[i].
|
||||
-> This is fully causal: model at bar i trained only on history ending at close[i-1].
|
||||
|
||||
Grid: 2 warmup periods (60 / 120 bars) × 2 TFs (1d / 12h) = 4 total cells (<=6 limit).
|
||||
Best config selected by min_asset_holdout_sharpe across all cells.
|
||||
"""
|
||||
import sys
|
||||
sys.path.insert(0, "/opt/docker/PythagorasGoal/scripts/research/alt")
|
||||
import altlib as al
|
||||
import numpy as np
|
||||
from sklearn.linear_model import SGDClassifier
|
||||
from sklearn.preprocessing import StandardScaler
|
||||
|
||||
|
||||
def online_sgd_logistic_target(df: "pd.DataFrame", warmup: int = 60) -> np.ndarray:
|
||||
"""
|
||||
Online SGD logistic regression updated each bar.
|
||||
|
||||
Causality:
|
||||
At bar i:
|
||||
1. We receive outcome from bar i-1 (sign of return from close[i-2] to close[i-1]).
|
||||
2. We do partial_fit(features[i-1], label[i-1]) — update model.
|
||||
3. We predict at features[i] -> continuous score via decision_function.
|
||||
4. Position = clip(score, 0, 1) to stay long-flat, then vol-target.
|
||||
|
||||
The model is never trained on data beyond close[i-1] when producing the position for
|
||||
bar i+1 (altlib shifts pos by 1 internally). So there is no look-ahead.
|
||||
"""
|
||||
c = df["close"].values.astype(float)
|
||||
n = len(c)
|
||||
|
||||
# --- Causal features computed once vectorially ---
|
||||
r = al.log_returns(c)
|
||||
ema_fast = al.ema(c, 10)
|
||||
ema_slow = al.ema(c, 40)
|
||||
ema_ratio = np.where(ema_slow > 0, ema_fast / ema_slow - 1.0, 0.0)
|
||||
|
||||
rsi14 = al.rsi(c, 14)
|
||||
rsi_norm = (rsi14 - 50.0) / 50.0 # normalize to [-1, 1]
|
||||
|
||||
zsc = al.zscore(c, 20)
|
||||
zsc = np.nan_to_num(zsc, nan=0.0)
|
||||
|
||||
rv_fast = al.realized_vol(r, 5, al.bars_per_year(df))
|
||||
rv_slow = al.realized_vol(r, 20, al.bars_per_year(df))
|
||||
rv_ratio = np.where((rv_slow > 0) & np.isfinite(rv_slow) & np.isfinite(rv_fast),
|
||||
rv_fast / rv_slow - 1.0, 0.0)
|
||||
|
||||
atr14 = al.atr(df, 14)
|
||||
atr_norm = np.where(c > 0, atr14 / c, 0.0)
|
||||
|
||||
# Feature matrix [n, 6]
|
||||
X = np.column_stack([
|
||||
ema_ratio,
|
||||
rsi_norm,
|
||||
zsc,
|
||||
rv_ratio,
|
||||
r, # last bar return (known at bar i)
|
||||
atr_norm,
|
||||
])
|
||||
X = np.nan_to_num(X, nan=0.0, posinf=0.0, neginf=0.0)
|
||||
|
||||
# Labels: sign of NEXT return (for training only; not used in prediction)
|
||||
# label[i] = sign(r[i+1]): known at bar i+1, used to update model at bar i+1
|
||||
labels = np.sign(np.roll(r, -1)) # peek-ahead in labels array only
|
||||
# But we access labels[i-1] at bar i -> labels[i-1] = sign(r[i]) which is known at i
|
||||
# So: when we update at bar i, we use label[i-1] = sign(r[i-1+1]) = sign(r[i])
|
||||
# r[i] = log(close[i]/close[i-1]) — fully known at bar i. Causal. ✓
|
||||
|
||||
# Online SGD Logistic
|
||||
clf = SGDClassifier(
|
||||
loss="log_loss",
|
||||
penalty="l2",
|
||||
alpha=1e-4,
|
||||
learning_rate="optimal",
|
||||
random_state=42,
|
||||
max_iter=1,
|
||||
warm_start=True,
|
||||
)
|
||||
|
||||
scores = np.zeros(n)
|
||||
classes = np.array([-1, 1])
|
||||
|
||||
for i in range(1, n):
|
||||
# Update model: use features[i-1] and label[i-1] (=sign(r[i]), known at i)
|
||||
label_i_minus_1 = int(np.sign(r[i])) # sign of return from close[i-1] to close[i]
|
||||
if label_i_minus_1 == 0:
|
||||
label_i_minus_1 = 1 # tie-break: treat flat as up
|
||||
|
||||
feat = X[i - 1].reshape(1, -1)
|
||||
|
||||
# Only partial_fit after warmup — before that, accumulate without predicting
|
||||
try:
|
||||
clf.partial_fit(feat, [label_i_minus_1], classes=classes)
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
# Predict at bar i if model has been fitted (after warmup)
|
||||
if i >= warmup:
|
||||
try:
|
||||
score = clf.decision_function(X[i].reshape(1, -1))[0]
|
||||
scores[i] = score
|
||||
except Exception:
|
||||
scores[i] = 0.0
|
||||
else:
|
||||
scores[i] = 0.0
|
||||
|
||||
# Convert decision score to long-flat position in [0, 1]
|
||||
# Use tanh to squash to (-1, 1), then clip to [0, 1] for long-flat
|
||||
pos_raw = np.tanh(scores) # in (-1, 1)
|
||||
pos_lf = np.clip(pos_raw, 0.0, 1.0) # long-flat
|
||||
|
||||
# Vol-target the position
|
||||
pos = al.vol_target(pos_lf, df, target_vol=0.20, vol_win_days=30, leverage_cap=2.0)
|
||||
return pos
|
||||
|
||||
|
||||
def make_target(warmup: int):
|
||||
def target_fn(df):
|
||||
return online_sgd_logistic_target(df, warmup=warmup)
|
||||
return target_fn
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
configs = [
|
||||
("warmup60", 60),
|
||||
("warmup120", 120),
|
||||
]
|
||||
|
||||
results = []
|
||||
for label, warmup in configs:
|
||||
print(f"\n--- Running STA07 config: {label} ---")
|
||||
rep = al.study_weights(
|
||||
f"STA07-OnlineSGD-{label}",
|
||||
make_target(warmup),
|
||||
tfs=("1d", "12h"),
|
||||
)
|
||||
print(al.fmt(rep))
|
||||
print("JSON:", al.as_json(rep))
|
||||
results.append((label, warmup, rep))
|
||||
|
||||
# Pick best config by best_holdout_sharpe from verdict
|
||||
best_label, best_warmup, best_rep = max(
|
||||
results,
|
||||
key=lambda x: x[2]["verdict"].get("best_holdout_sharpe", -99)
|
||||
)
|
||||
print(f"\n=== BEST CONFIG: {best_label} ===")
|
||||
print(al.fmt(best_rep))
|
||||
print("JSON:", al.as_json(best_rep))
|
||||
Reference in New Issue
Block a user