File size: 12,670 Bytes
d699ea7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 | #!/usr/bin/env python3
"""Batched (GPU) backtest for the Stage-3 flash-crash detector.
The full cascade scores Stage-3 once per tick (a TCN forward per tick), which is
correct in production but makes offline backtests impractically slow. This script
precomputes every stride-1 200-tick window score in one batched GPU pass, then
replays the (fixed) cascade decision cheaply:
alert(t) <= s3(t) >= stage3_threshold AND Stage-5 posterior >= alert_threshold
Stage-5 fusion uses the trained Stage-2 score (cheap) + neutral Stage-4 (0.5).
Stage-1 is advisory (never a veto). Crash windows use the same 2%-drop-in-60s
label rule as the cascade backtest, and alerts are coalesced by --cooldown-ms.
Usage:
PYTHONPATH=ml python scripts/run_backtest_batched.py \
--data data/parquet/LUNAUSDT_2022-05-10.parquet --model models/stage3_tcn_prod.pt
"""
import argparse
import logging
import sys
from pathlib import Path
import numpy as np
import pandas as pd
import torch
ML_DIR = Path(__file__).resolve().parent.parent / "ml"
sys.path.insert(0, str(ML_DIR))
from flash_crash_watchdog.data.historical_loader import df_to_ticks # noqa: E402
from flash_crash_watchdog.data.labels import label_crashes # noqa: E402
from flash_crash_watchdog.features import FEATURE_NAMES, FeatureExtractor # noqa: E402
from flash_crash_watchdog.models.stage3_tcn import normalize_z, TCNConfig, TCNDetector # noqa: E402
torch.serialization.add_safe_globals([TCNConfig])
from flash_crash_watchdog.models.stage2_isolation_forest import Stage2IsolationForest # noqa: E402
from flash_crash_watchdog.models.stage5_bayesian import Stage5Config, Stage5Bayesian # noqa: E402
logger = logging.getLogger(__name__)
TCN_FEATURES = FEATURE_NAMES[:17]
W = 200 # trained window length
NORM_WINDOW = 500 # rolling-z window (matches training / Stage3TCN.feed)
CH = 4096 # GPU batch per forward chunk
def device() -> str:
return "cuda" if torch.cuda.is_available() else "cpu"
def stage3_scores(model: TCNDetector, F: np.ndarray, dev: str) -> np.ndarray:
"""Every stride-1 200-window Stage-3 score (batched GPU)."""
if len(F) <= W:
return np.zeros(0, dtype=np.float32)
norm = normalize_z(F, NORM_WINDOW)
n = len(norm) - W + 1
out = np.empty(n, dtype=np.float32)
for s in range(0, n, CH):
e = min(s + CH, n)
win = np.stack([norm[i:i + W] for i in range(s, e)])
x = torch.from_numpy(np.ascontiguousarray(win)).permute(0, 2, 1).float().to(dev)
with torch.no_grad():
out[s:e] = model(x)[:, -1].cpu().numpy()
return out
def score_day(model: TCNDetector, dev: str, parquet: str | Path, max_ticks: int, crash_pct: float):
"""Feature-extract + score one day. Returns (scores, times, ticks, mids, crash_ranges)."""
df = pd.read_parquet(parquet)
if max_ticks > 0 and len(df) > max_ticks:
df = df.iloc[: max_ticks]
ticks = list(df_to_ticks(df, symbol="DAY"))
extractor = FeatureExtractor()
F = np.zeros((len(ticks), 17), dtype=np.float32)
times = np.empty(len(ticks), dtype=np.int64)
mids = np.full(len(ticks), np.nan, dtype=np.float64)
for i, t in enumerate(ticks):
fd = extractor.extract(t)
F[i] = [float(fd.get(k, 0.0)) or 0.0 for k in TCN_FEATURES]
times[i] = t.book.timestamp_ms
mp = t.book.mid_price
if mp is not None:
mids[i] = mp
F = np.nan_to_num(F, nan=0.0, posinf=0.0, neginf=0.0)
filled = pd.Series(mids).ffill().bfill().to_numpy()
mids = np.where(np.isnan(filled), 1.0, filled)
scores = stage3_scores(model, F, dev) # scores[j] => window ending at tick j+W-1
crashes = label_crashes(ticks, drop_threshold_pct=crash_pct)
crash_ranges = [(c.start_ts, c.end_ts) for c in crashes]
return scores, times, ticks, mids, crash_ranges
def evaluate(scores, times, ticks, s2, s5, threshold, crash_ranges, cooldown_ms,
mids=None, confirmation_drop_pct=0.0, drop_lookback=0, min_tv_bps=0.0):
"""Replay + metrics for one threshold. Returns (alerts,tp,fp,fn,prec,rec,f1,med_ttd,rate).
When confirmation_drop_pct > 0, an alert fires only if the mid is in a REAL
sustained decline: mids[alert_tick] <= mids[alert_tick - drop_lookback] *
(1 - confirmation_drop_pct/100). This rejects transient normal-day spikes
while keeping crash-day sustained drops. drop_lookback==0 => W//2 (100 ticks).
When min_tv_bps > 0, an alert fires only if the window's trailing realized
volatility (std/mean of the mid over the 200-tick window, in bps) is above
the threshold. Normal-day false positives are on ~0-vol flat windows; crash
onsets are on high-vol windows — so this suppresses calm-day chatter while
keeping crash detections (and it is observable in real time).
"""
n_scores = len(scores)
if drop_lookback <= 0:
drop_lookback = W // 2
alerts = []
last_ts = None
for j, tick_idx in enumerate(np.arange(n_scores) + W - 1):
s3v = float(scores[j])
if s3v < threshold:
continue
tick = ticks[tick_idx]
s2v, _ = s2.score(tick)
alert = s5.aggregate(tick, s2v, s3v, 0.5)
if alert is None:
continue
ts = times[tick_idx]
if min_tv_bps > 0:
if mids is None or tick_idx < W:
continue
seg = mids[tick_idx - W + 1:tick_idx + 1]
mn = float(seg.mean())
tv = float(seg.std() / mn) * 10000.0 if mn > 0 else 0.0
if tv < min_tv_bps:
continue
if confirmation_drop_pct > 0:
if mids is None or tick_idx < drop_lookback:
continue
if not (mids[tick_idx] <= mids[tick_idx - drop_lookback] * (1.0 - confirmation_drop_pct / 100.0)):
continue
if last_ts is not None and ts - last_ts < cooldown_ms:
continue
last_ts = ts
alerts.append(ts)
matched = set()
ttd = []
tp = 0
for ts in alerts:
hit = [i for i, (a, b) in enumerate(crash_ranges) if a <= ts <= b]
if hit:
tp += 1
matched.add(hit[0])
ttd.append(crash_ranges[hit[0]][1] - ts)
fn = len(crash_ranges) - len(matched)
fp = len(alerts) - tp
prec = tp / (tp + fp) if tp + fp else 0.0
rec = tp / (tp + fn) if tp + fn else 0.0
f1 = 2 * prec * rec / (prec + rec) if prec + rec else 0.0
return (len(alerts), tp, fp, fn, prec, rec, f1,
float(np.median(ttd)) if ttd else 0.0, 100.0 * len(alerts) / max(1, len(ticks)))
def main() -> int:
parser = argparse.ArgumentParser(description="Batched GPU Stage-3 backtest")
parser.add_argument("--data", required=True, help="Tick parquet")
parser.add_argument("--model", default="models/stage3_tcn_prod.pt")
parser.add_argument("--stage2", default="models/stage2_isolation_forest.joblib")
parser.add_argument("--threshold", type=float, default=0.5,
help="Stage-3 pass gate (operating point: 0.5; or use --thresholds)")
parser.add_argument("--thresholds", default=None, help="comma list to sweep, e.g. '0.3,0.4,0.5,0.6,0.7'")
parser.add_argument("--stage5-threshold", type=float, default=0.5)
parser.add_argument("--cooldown-ms", type=int, default=10000)
parser.add_argument("--confirmation-drop-pct", type=float, default=0.0,
help="0.0 = OFF. Confirmation gate: an alert counts only when the "
"mid is in a REAL sustained decline: "
"mids[alert_tick] <= mids[alert_tick - drop_lookback] * (1 - pct/100). "
"Rejects transient normal-day spikes, keeps crash-day sustained drops.")
parser.add_argument("--drop-lookback", type=int, default=0,
help="ticks back to compare mid for the confirmation gate; 0 = W//2 (100). "
"Ignored when --confirmation-drop-pct == 0.")
parser.add_argument("--min-trailing-vol-bps", type=float, default=0.0,
help="Regime gate (bps): only alert if the window's trailing realized "
"volatility (mid std/mean over 200 ticks) exceeds this. Suppresses "
"calm-day chatter (normal FPs are ~0 bps; crash onsets are 10s of bps). "
"0.0 = off.")
parser.add_argument("--out-json", default=None,
help="Write the threshold-sweep table to a JSON file (for the leaderboard)")
parser.add_argument("--crash-pct", type=float, default=2.0)
parser.add_argument("--max-ticks", type=int, default=0, help="0 = all")
parser.add_argument("--calibrate-on", default=None,
help="RSR-11: a SEPARATE parquet day used ONLY to select the "
"threshold (max F1). The eval day (--data) is never used "
"for threshold selection, which would overfit the metric.")
args = parser.parse_args()
logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
dev = device()
st = torch.load(args.model, map_location="cpu", weights_only=True) # MLOPS-06
cfg = st["config"] if isinstance(st["config"], TCNConfig) else TCNConfig(**st["config"])
model = TCNDetector(cfg).to(dev)
model.load_state_dict(st["model_state"])
model.eval()
s2 = Stage2IsolationForest()
if Path(args.stage2).exists():
s2.load(args.stage2)
s5 = Stage5Bayesian(Stage5Config(alert_threshold=args.stage5_threshold))
scores, times, ticks, mids, crash_ranges = score_day(model, dev, args.data, args.max_ticks, args.crash_pct)
logger.info("Found %d ground-truth crash windows on eval day", len(crash_ranges))
thrs = [float(t) for t in args.thresholds.split(",")] if args.thresholds else [args.threshold]
# RSR-11: threshold selection must never happen on the evaluation day.
if args.calibrate_on:
cal = score_day(model, dev, args.calibrate_on, args.max_ticks, args.crash_pct)
best_f1, best_thr = -1.0, thrs[0]
for thr in thrs:
_an, tp, fp, fn, _prec, rec, f1, _med, _rate = evaluate(
cal[0], cal[1], cal[2], s2, s5, thr, cal[4], args.cooldown_ms,
mids=cal[3], confirmation_drop_pct=args.confirmation_drop_pct,
drop_lookback=args.drop_lookback, min_tv_bps=args.min_trailing_vol_bps)
if f1 > best_f1:
best_f1, best_thr = f1, thr
logger.info("RSR-11: selected threshold %.3f on calibration day %s (F1=%.3f); "
"eval day %s is untouched by selection",
best_thr, Path(args.calibrate_on).name, best_f1, Path(args.data).name)
thrs = [best_thr]
elif len(thrs) > 1:
logger.warning(
"RSR-11: sweeping %d thresholds ON the evaluation day overfits the metric — "
"use --calibrate-on to select the threshold on a separate day", len(thrs))
print("day=%s ticks=%d crashes=%d gate_drop_pct=%.2f drop_lookback=%d min_tv_bps=%.1f"
% (Path(args.data).name, len(ticks), len(crash_ranges),
args.confirmation_drop_pct, (args.drop_lookback or W // 2), args.min_trailing_vol_bps))
rows = []
print("threshold,alerts,tp,fp,fn,precision,recall,f1,alert_rate_pct,median_ttd_ms")
for thr in thrs:
alerts_n, tp, fp, fn, prec, rec, f1, med, rate = evaluate(
scores, times, ticks, s2, s5, thr, crash_ranges, args.cooldown_ms,
mids=mids, confirmation_drop_pct=args.confirmation_drop_pct,
drop_lookback=args.drop_lookback, min_tv_bps=args.min_trailing_vol_bps)
print("%.3f,%d,%d,%d,%d,%.3f,%.3f,%.3f,%.3f,%.0f" % (thr, alerts_n, tp, fp, fn, prec, rec, f1, rate, med))
rows.append({"threshold": thr, "alerts": alerts_n, "tp": tp, "fp": fp, "fn": fn,
"precision": prec, "recall": rec, "f1": f1,
"alert_rate_pct": rate, "median_ttd_ms": med})
if args.out_json:
import json as _json
payload = {
"day": Path(args.data).name,
"ticks": len(ticks),
"crashes": len(crash_ranges),
"gate_bps": args.min_trailing_vol_bps,
"cooldown_ms": args.cooldown_ms,
"model": Path(args.model).name,
"thresholds": rows,
}
Path(args.out_json).write_text(_json.dumps(payload, indent=2))
logger.info("wrote leaderboard table -> %s", args.out_json)
return 0
if __name__ == "__main__":
raise SystemExit(main()) |