yirugi
/
fithia2
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Diagnose RSI-14/BB%B axis for V27: ABORT — redundant with OBV-slope

Browse Source 
RSI-14: Pearson=+0.157, G2 passes (+0.316R gap, direction confirmed).
BUT ρ(RSI-14, obv_slope_20)=+0.726 — G5a FAIL. RSI captures same momentum
information as V24's OBV-slope. BB %B is ρ=0.889 with RSI — identical axis.
Both fail G1 (n=98 < 120 threshold). V24 remains champion.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
main
I Luk Kim 4 months ago
parent 137b413084
commit ab34c3ef1b
1 changed files with 462 additions and 0 deletions
Show Stats Download Patch File Download Diff File

462
apps/intraday_bt/scripts/diag_orb_momentum_features.py
Unescape Escape View File

@ -0,0 +1,462 @@
"""
V27 ORB Momentum Features Diagnostic — RSI-14 and Bollinger Band %B
Hypotheses:
RSI-14 (prior day):
Low RSI (<30 = oversold): post-selloff gap-up → strong reversion drive
High RSI (>70 = overbought): momentum continuation → gap adds to trend
Mid RSI (40-60): neutral — empirically uncertain
BB %B (prior day):
Low %B (<0.3): near lower band = compressed/reset before gap-up → explosive
High %B (>0.7): near/above upper band = existing momentum + gap = continuation
Direction: empirically determined, not assumed
Both features use daily bars from `prev_bars` (strictly before entry date — lookahead-free).
These are the next natural axes after:
- V25: FINRA short-volume axis NULL (Pearson max 0.062)
- V26: ORB tape ignition NULL (range_coil_orb Pearson=-0.128 but p~0.10, n=101 < 120)
Method:
1. Run V24 simulation over 200d window → 101 trades with r_multiple
2. For each trade, compute RSI-14 and BB %B from prev_bars (daily bars before entry)
3. Also compute RSI-OBV interaction: RSI × obv_slope_20 / sqrt(2) — additive signal?
4. Report Pearson, tercile WR/avg_R, gate outcomes
5. Pairwise: vs obv_slope_20 and avg_daily_vol_14d (G5a/G5b)
Gates:
G1: |Pearson| ≥ 0.07 on ≥ 120 trades (relaxed to 60 if coverage < 80%)
G2: |top − bottom tercile avg_R| ≥ 0.30R
G3: |top − bottom tercile WR| ≥ 5pp
G4: feature coverage ≥ 50%
G5a: |ρ(feature, obv_slope_20)| < 0.70
G5b: |ρ(feature, avg_daily_vol_14d)| < 0.70
"""
from __future__ import annotations
import concurrent.futures
import datetime as dt
import os
import sys
from pathlib import Path
import pyarrow.parquet as pq
import yaml
from zoneinfo import ZoneInfo
sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../../..")))
from libs.common.time_utils import trading_days_between
from libs.intraday.domain import ORBStrategyParams
from libs.intraday.features import compute_obv_slope_approx, enrich_daily_bars
from libs.intraday.orb_simulator import ORBSimulationState, run_orb_simulation_with_state
from libs.intraday.screener import orb_pre_screen_candidates
# ── Config ──────────────────────────────────────────────────────────────────
V24_CONFIG = "configs/intraday/strategies/orb_gainers_v24_quality_overlay.yaml"
UNIVERSE_FILE = "configs/symbols_midlarge_snapshot_exact.yaml"
INTRADAY_CACHE_DIR = "data/cache/intraday"
LOOKBACK_DAYS = 200
FEATURE_LOOKBACK_TRADING = 30 # RSI-14 needs 15 prior bars; BB needs 20
_ET = ZoneInfo("America/New_York")
_MKT_OPEN = dt.time(9, 30)
_MKT_CLOSE = dt.time(16, 0)
# ── Daily Bar Builder ─────────────────────────────────────────────────────────
def _parse_ts(ts_raw: object) -> dt.datetime:
s = str(ts_raw)
if s.endswith("Z"):
s = s[:-1] + "+00:00"
return dt.datetime.fromisoformat(s).astimezone(_ET)
def _build_daily_bar_from_intraday(path: Path, date: str) -> dict | None:
try:
table = pq.read_table(str(path))
rows = table.to_pydict()
except Exception:
return None
opens, highs, lows, closes, vols = [], [], [], [], []
for i, ts_raw in enumerate(rows.get("timestamp", [])):
try:
ts = _parse_ts(ts_raw)
except Exception:
continue
if _MKT_OPEN <= ts.time() < _MKT_CLOSE:
opens.append(float(rows["open"][i] or 0))
highs.append(float(rows["high"][i] or 0))
lows.append(float(rows["low"][i] or 0))
closes.append(float(rows["close"][i] or 0))
vols.append(float(rows["volume"][i] or 0))
if not opens:
return None
return {
"date": date, "open": opens[0], "high": max(highs),
"low": min(lows), "close": closes[-1], "volume": sum(vols),
}
def build_daily_bars(tickers: list[str], dates: list[str], workers: int = 8) -> dict[str, list[dict]]:
root = Path(INTRADAY_CACHE_DIR)
def _load(ticker: str) -> tuple[str, list[dict]]:
d_path = root / ticker
if not d_path.is_dir():
return ticker, []
bars: list[dict] = []
for date in dates:
p = d_path / f"{date}.parquet"
if not p.exists():
continue
bar = _build_daily_bar_from_intraday(p, date)
if bar and bar["close"] > 0:
bars.append(bar)
return ticker, bars
result: dict[str, list[dict]] = {}
with concurrent.futures.ThreadPoolExecutor(max_workers=workers) as ex:
for ticker, bars in ex.map(_load, tickers):
if bars:
result[ticker] = bars
return result
def load_intraday_bulk(candidates: dict[str, list[str]]) -> dict[str, dict[str, list[dict]]]:
import pandas as pd
result: dict[str, dict[str, list[dict]]] = {}
for date, tickers in candidates.items():
day_bars: dict[str, list[dict]] = {}
for ticker in tickers:
p = Path(INTRADAY_CACHE_DIR) / ticker / f"{date}.parquet"
if not p.exists():
continue
try:
df = pd.read_parquet(str(p))
if not df.empty and len(df) >= 5:
day_bars[ticker] = df.to_dict("records")
except Exception:
pass
if day_bars:
result[date] = day_bars
return result
# ── Feature Computations (dict-based daily bars, lookahead-free) ──────────────
def compute_rsi_14(prev_bars: list[dict]) -> float | None:
"""RSI(14) from last 15 bars in prev_bars (14 price changes). Lookahead-free."""
if len(prev_bars) < 15:
return None
recent = sorted(prev_bars, key=lambda b: b["date"])[-15:]
gains, losses = [], []
for i in range(14):
chg = recent[i + 1]["close"] - recent[i]["close"]
if chg >= 0:
gains.append(chg); losses.append(0.0)
else:
gains.append(0.0); losses.append(abs(chg))
avg_gain = sum(gains) / 14
avg_loss = sum(losses) / 14
if avg_loss == 0:
return 100.0
rs = avg_gain / avg_loss
return 100.0 - (100.0 / (1.0 + rs))
def compute_bb_pct_b(prev_bars: list[dict], window: int = 20) -> float | None:
"""BB %B: (last_close - lower_band) / (upper_band - lower_band). >1 = above upper."""
if len(prev_bars) < window:
return None
recent = sorted(prev_bars, key=lambda b: b["date"])[-window:]
closes = [b["close"] for b in recent]
sma = sum(closes) / window
if sma <= 0:
return None
std = (sum((c - sma) ** 2 for c in closes) / window) ** 0.5
if std <= 0:
return 0.5
band_w = 4 * std # upper - lower = 4σ
if band_w <= 0:
return 0.5
return (closes[-1] - (sma - 2 * std)) / band_w
# ── Stats Helpers ─────────────────────────────────────────────────────────────
def pearson(xs: list[float], ys: list[float]) -> float | None:
if len(xs) != len(ys) or len(xs) < 2:
return None
n = len(xs)
xm, ym = sum(xs) / n, sum(ys) / n
num = sum((xs[i] - xm) * (ys[i] - ym) for i in range(n))
dx = (sum((x - xm) ** 2 for x in xs)) ** 0.5
dy = (sum((y - ym) ** 2 for y in ys)) ** 0.5
if dx <= 0 or dy <= 0:
return None
return num / (dx * dy)
def tercile_stats(vals: list[float], outcomes_r: list[float]) -> dict:
if len(vals) < 6:
return {}
pairs = sorted(zip(vals, outcomes_r), key=lambda p: p[0])
n = len(pairs)
t = n // 3
def stats(sub):
ys = [p[1] for p in sub]
wr = sum(1 for y in ys if y > 0) / len(ys) if ys else 0.0
avg = sum(ys) / len(ys) if ys else 0.0
return {"n": len(ys), "wr": wr, "avg_r": avg}
return {"low": stats(pairs[:t]), "mid": stats(pairs[t:2*t]), "high": stats(pairs[2*t:])}
# ── Main ─────────────────────────────────────────────────────────────────────
def main() -> None:
print("=== V27 ORB Momentum Features Diagnostic (RSI-14 + BB %B) ===\n")
# 1. Load V24 config
with open(V24_CONFIG) as f:
raw = yaml.safe_load(f)
params = ORBStrategyParams(**raw["orb_strategy"])
print(f"V24 config loaded. weight_obv_slope={params.weight_obv_slope}")
# 2. Determine 200d trading window
today = dt.date(2026, 4, 21)
all_td = trading_days_between(today - dt.timedelta(days=400), today)
trading_days_list = [d.isoformat() for d in all_td[-LOOKBACK_DAYS:]]
print(f"Window: {trading_days_list[0]} → {trading_days_list[-1]} ({len(trading_days_list)} trading days)")
extended_td = [d.isoformat() for d in all_td[-(LOOKBACK_DAYS + FEATURE_LOOKBACK_TRADING + 10):]]
first_cal = dt.date.fromisoformat(extended_td[0])
last_cal = dt.date.fromisoformat(trading_days_list[-1])
needed_dates: list[str] = []
d = first_cal
while d <= last_cal:
needed_dates.append(d.isoformat())
d += dt.timedelta(days=1)
# 3. Load universe
with open(UNIVERSE_FILE) as f:
udata = yaml.safe_load(f)
universe = udata.get("symbols", udata) if isinstance(udata, dict) else udata
if "QQQ" not in universe:
universe = list(universe) + ["QQQ"]
print(f"Universe: {len(universe)} tickers")
# 4. Build daily bars + enrichment
print(f"\nBuilding daily bars ({len(needed_dates)} calendar days)...")
daily_bars = build_daily_bars(universe, needed_dates)
print(f"Built daily bars for {len(daily_bars)} tickers")
print("Computing enrichment...")
enrichment = enrich_daily_bars(daily_bars, trading_days_list)
print(f"Enrichment for {len(enrichment)} tickers")
# 5. Pre-screen + load intraday
candidates = orb_pre_screen_candidates(
daily_bars, trading_days_list, enrichment,
min_price=params.min_price,
min_atr=params.min_atr_14,
min_avg_dollar_vol=params.min_avg_dollar_volume,
max_per_day=None,
)
total_pairs = sum(len(v) for v in candidates.values())
print(f"Pre-screened: {total_pairs} ticker-days")
print("Loading intraday bars...")
all_intraday = load_intraday_bulk(candidates)
print(f"Loaded: {sum(len(v) for v in all_intraday.values())} ticker-days")
# 6. Run V24 simulation
print("\nRunning V24 simulation...")
state = ORBSimulationState(equity=params.initial_capital)
day_results, _ = run_orb_simulation_with_state(
all_intraday, trading_days_list, params, enrichment, state=state,
)
all_trades = [t for dr in day_results for t in dr.trades]
trades_with_r = [t for t in all_trades if getattr(t, "r_multiple_at_exit", None) is not None]
print(f"Total trades: {len(all_trades)}, with r_multiple: {len(trades_with_r)}")
if len(trades_with_r) < 20:
print("ABORT: fewer than 20 trades with r_multiple")
return
# 7. Compute features per trade
print("\nComputing RSI-14 and BB %B features per trade...")
sorted_daily: dict[str, list[dict]] = {
t: sorted(bars, key=lambda b: b["date"])
for t, bars in daily_bars.items()
}
annotated: list[dict] = []
missing: dict[str, int] = {"rsi_14": 0, "bb_pct_b": 0, "obv_slope": 0}
for trade in trades_with_r:
ticker = trade.ticker
date = str(trade.date)[:10]
r = float(trade.r_multiple_at_exit)
bars_t = sorted_daily.get(ticker, [])
prev_bars = [b for b in bars_t if b["date"][:10] < date]
f_rsi = compute_rsi_14(prev_bars)
f_bb = compute_bb_pct_b(prev_bars)
f_obv = compute_obv_slope_approx(prev_bars, lookback=20)
avg_vol = enrichment.get(ticker, {}).get(date, {}).get("avg_daily_vol_14d")
for fname, fval in [("rsi_14", f_rsi), ("bb_pct_b", f_bb), ("obv_slope", f_obv)]:
if fval is None:
missing[fname] += 1
annotated.append({
"ticker": ticker, "date": date, "r": r, "win": r > 0,
"rsi_14": f_rsi,
"bb_pct_b": f_bb,
"obv_slope": f_obv,
"avg_daily_vol": avg_vol,
})
total = len(annotated)
print(f"Annotated: {total} trades")
for fname in ["rsi_14", "bb_pct_b", "obv_slope"]:
print(f" Missing {fname}: {missing[fname]}/{total}")
# 8. Feature analysis — empirically determine best tercile direction
feature_defs = [
# (key, description, hypothesized best tercile — empirically verified)
("rsi_14", "RSI-14 prior day: momentum state before gap-up", "high"),
("bb_pct_b", "BB %B prior day: price position in Bollinger Band", "low"),
]
print("\n" + "=" * 90)
print("FEATURE ANALYSIS — V24 200d trade set")
print("=" * 90)
results: dict[str, dict | None] = {}
for feat_name, description, hypothesized_best in feature_defs:
valid = [(a[feat_name], a["r"]) for a in annotated if a[feat_name] is not None]
if len(valid) < 20:
print(f"\n{feat_name}: SKIP — only {len(valid)} valid trades (need ≥20)")
results[feat_name] = None
continue
vals = [v[0] for v in valid]
rs = [v[1] for v in valid]
overall_wr = sum(1 for v in valid if v[1] > 0) / len(valid)
rho = pearson(vals, rs)
tstat = tercile_stats(vals, rs)
coverage_pct = len(valid) / total
print(f"\n{'─'*60}")
print(f"FEATURE: {feat_name}")
print(f" Description: {description}")
print(f" n={len(valid)}/{total} ({coverage_pct*100:.0f}% coverage), overall WR={overall_wr*100:.1f}%")
rho_abs = abs(rho) if rho is not None else 0.0
print(f" Pearson(feature, r_multiple) = {rho:.4f}" if rho is not None else " Pearson = n/a")
if tstat:
h, m, lo = tstat["high"], tstat["mid"], tstat["low"]
print(f" Tercile breakdown (low→high feature value):")
print(f" Bottom: n={lo['n']}, WR={lo['wr']*100:.1f}%, avg_R={lo['avg_r']:+.3f}")
print(f" Middle: n={m['n']}, WR={m['wr']*100:.1f}%, avg_R={m['avg_r']:+.3f}")
print(f" Top: n={h['n']}, WR={h['wr']*100:.1f}%, avg_R={h['avg_r']:+.3f}")
# Empirically determine best tercile (higher avg_R wins)
if h["avg_r"] >= lo["avg_r"]:
best_tercile = "high"
worst_tercile = "low"
else:
best_tercile = "low"
worst_tercile = "high"
direction_match = best_tercile == hypothesized_best
print(f" Empirical best: '{best_tercile}' tercile (hypothesis: '{hypothesized_best}' → {'✓ confirmed' if direction_match else '✗ INVERTED'})")
best = tstat[best_tercile]
worst = tstat[worst_tercile]
min_trades = 120 if coverage_pct >= 0.80 else 60
g1 = rho_abs >= 0.07 and len(valid) >= min_trades
g2 = best["avg_r"] - worst["avg_r"] >= 0.30
g3 = best["wr"] >= worst["wr"] + 0.05
g4 = coverage_pct >= 0.50
n_failed = sum([not g1, not g2, not g3, not g4])
overall_pass = n_failed == 0
print(f" Gates (best='{best_tercile}' tercile):")
print(f" G1 |Pearson|≥0.07 + n≥{min_trades}: {rho_abs:.4f}, n={len(valid)} → {'PASS ✓' if g1 else 'FAIL ✗'}")
print(f" G2 avg_R gap ≥ 0.30R: {best['avg_r'] - worst['avg_r']:+.3f} → {'PASS ✓' if g2 else 'FAIL ✗'}")
print(f" G3 WR gap ≥ 5pp: {(best['wr'] - worst['wr'])*100:+.1f}pp → {'PASS ✓' if g3 else 'FAIL ✗'}")
print(f" G4 coverage ≥ 50%: {coverage_pct*100:.0f}% → {'PASS ✓' if g4 else 'FAIL ✗'}")
print(f" VERDICT: {'ALL GATES PASS → PROCEED TO PHASE 2' if overall_pass else f'FAIL ({n_failed} gate(s) failed)'}")
results[feat_name] = {
"pass": overall_pass, "pearson": rho, "stats": tstat,
"n": len(valid), "coverage": coverage_pct,
"best_tercile": best_tercile, "hypothesized_best": hypothesized_best,
}
# 9. Pairwise correlations
print(f"\n{'─'*60}")
print("PAIRWISE CORRELATIONS:")
for feat_name in [fd[0] for fd in feature_defs]:
for ref_key, ref_label, gate_name in [
("obv_slope", "obv_slope_20", "G5a"),
("avg_daily_vol", "avg_daily_vol_14d", "G5b"),
]:
combined = [
(a[feat_name], a[ref_key]) for a in annotated
if a[feat_name] is not None and a[ref_key] is not None
]
if len(combined) >= 10:
rho_x = pearson([c[0] for c in combined], [c[1] for c in combined])
if rho_x is not None:
gate_pass = abs(rho_x) < 0.70
print(f" ρ({feat_name[:18]:18s}, {ref_label}): {rho_x:+.4f} {gate_name}: {'PASS ✓' if gate_pass else 'FAIL ✗'}")
print("\n Feature intercorrelation:")
combined_rr = [
(a["rsi_14"], a["bb_pct_b"]) for a in annotated
if a["rsi_14"] is not None and a["bb_pct_b"] is not None
]
if combined_rr:
rho_cross = pearson([c[0] for c in combined_rr], [c[1] for c in combined_rr])
print(f" ρ(rsi_14, bb_pct_b): {rho_cross:.4f}" if rho_cross is not None else " ρ(rsi_14, bb_pct_b): n/a")
# 10. Summary
passing = [name for name, r in results.items() if r is not None and r["pass"]]
failed = [name for name, r in results.items() if r is not None and not r["pass"]]
skipped = [name for name, r in results.items() if r is None]
print(f"\n{'='*90}")
print("SUMMARY")
print(f"{'='*90}")
print(f"Features passing all gates: {passing if passing else 'NONE'}")
print(f"Features failing gates: {failed if failed else 'NONE'}")
print(f"Features skipped: {skipped if skipped else 'NONE'}")
if passing:
best = max(passing, key=lambda n: abs(results[n]["pearson"] or 0))
pearson_val = results[best]["pearson"]
obv_pearson_ref = 0.2349
weight_magnitude = round(0.05 * min(1.0, abs(pearson_val) / obv_pearson_ref), 2)
weight_magnitude = max(weight_magnitude, 0.02)
best_direction = results[best]["best_tercile"]
weight_sign = +1 if best_direction == "high" else -1
print(f"\nVERDICT: PROCEED TO PHASE 2")
print(f" Best feature: {best}")
print(f" Pearson: {pearson_val:.4f}")
print(f" Direction: '{best_direction}' tercile is best → weight = {weight_sign * weight_magnitude:+.3f}")
print(f" Suggested weight_momentum_signal in V27 config: {weight_sign * weight_magnitude:+.3f}")
print(f" → Wire '{best}' into enrich_daily_bars as momentum signal")
print(f" → Add weight_momentum_signal to V27 config (parent V24)")
else:
print(f"\nVERDICT: ABORT — RSI/BB momentum axis null on V24 200d trade set")
print(" → V24 remains champion. Momentum overlay exhausted.")
print(" → Next Ralph iteration: gravitational pull or gap z-score from libs/features/market_features.py")
if __name__ == "__main__":
main()
Write Preview
Loading…
Cancel
Save