Spaces:
Running
Running
File size: 18,951 Bytes
b0d9ebf 8af7f2e b0d9ebf 8af7f2e c789230 b0d9ebf 8af7f2e c789230 b0d9ebf c789230 b0d9ebf c789230 b0d9ebf | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 | """ml_predictor/features.py β shared point-in-time feature builder.
ONE source of truth for the ML model's feature vector, used identically by:
β’ dataset.py (training-CSV construction over historical dates)
β’ infer.py (live inference on the latest bar)
so training features exactly equal production features (no train/serve skew).
Every feature is computed point-in-time from data up to and including `date`
(no lookahead), mirroring research/backtest.py::_compute_indicators and the
11-weight ML sub-feature math in predictor_core.get_ml_feature_score (L667-753).
News sentiment is deliberately EXCLUDED β it is live-only and cannot be
backfilled to historical dates (see CLAUDE.md). It is applied as a live
inference-time confidence adjustment inside infer.py, not as a trained feature.
"""
from __future__ import annotations
import os
import numpy as np
import pandas as pd
# Lightweight indicator primitives (no heavy imports).
from trial_run import rsi, atr, obv, adx_s, macd_h
from ml_combiner import bollinger_position, ema_stack_score, shadow_flag
TIMEFRAMES = ["INTRADAY", "1D", "3D"]
# Experiment flag: when ML_STRATEGY_FEATURES=1, the best NSE-backtested S-signals are
# appended as binary features (computed in dataset.py per ticker). Off by default.
STRATEGY_FEATURE_COLS = ["sig_s8", "sig_ctrio", "sig_s4v2", "sig_s6", "sig_s16", "sig_s1"]
_USE_STRATEGY_FEATS = os.environ.get("ML_STRATEGY_FEATURES") == "1"
# Extra indicator + Monte-Carlo features. PRODUCTION DEFAULT (validated on the 5-month OOS
# holdout, research/ml_backtest.py): vs the 37-feature baseline they lift INTRADAY direction
# 71%β82% and price-target hit 64%β84%, cut the estimated-high MAE 1.54β1.07, and raise 1D/3D
# per-trade P&L (+0.86β+1.02, +1.29β+1.51, PF 1.74β1.93 / 1.79β1.96) with NO regression. The
# channel positions, extra oscillators, realized vol, and MC hit-probability/excursion estimates
# are computed identically in compute_features (live) and research/augment_features.py (training
# CSV) via the shared _extra_feature_series / _mc_features helpers β no train/serve skew.
# Set ML_EXTRA_FEATURES=0 to revert to the 37-feature model (must retrain to match).
EXTRA_FEATURE_COLS = [
"bb_bandwidth", "bb_squeeze", # Bollinger volatility regime / squeeze
"keltner_pct", "donchian_pct", # channel positions
"stoch_k", "cci20", "mfi14", "williams_r", # extra oscillators
"hist_vol_20", # realized volatility
"mc_up_prob_3d", "mc_exp_maxup_3d", "mc_exp_maxdn_3d", # Monte-Carlo path features
]
_USE_EXTRA_FEATS = os.environ.get("ML_EXTRA_FEATURES", "1") != "0"
# Ordered feature list β the manifest stores this so infer.py builds columns in
# the exact order the models were trained on. Keep additions APPEND-ONLY.
FEATURE_COLUMNS: list[str] = [
# Momentum / oscillators
"rsi14", "rsi5", "rsi2", "macd_hist", "adx14",
# Trend / EMA
"price_vs_ema20", "price_vs_ema50", "price_vs_ema200", "ema_stack", "supertrend",
# Bollinger
"bb_pct", "bb_lower_dist", "bb_upper_dist",
# Relative strength vs Nifty
"rs_3m",
# Volume
"vol_ratio", "obv_z", "vol_trend_5d",
# Volatility / range
"atr_pct", "shadow",
# Returns / streak
"return_10d", "return_20d", "return_90d", "dist_52w_high", "consec_days",
# Domain trigger flags
"trigger_T1", "trigger_T2", "trigger_T3", "trigger_T4", "trigger_T5",
"trigger_T6", "trigger_T7", "trigger_B1", "trigger_B2", "trigger_B3",
# Macro regime
"vix_level", "nifty_ok", "vix_decl",
]
if _USE_STRATEGY_FEATS:
FEATURE_COLUMNS = FEATURE_COLUMNS + STRATEGY_FEATURE_COLS
if _USE_EXTRA_FEATS:
FEATURE_COLUMNS = FEATURE_COLUMNS + EXTRA_FEATURE_COLS
_NAN = float("nan")
def _supertrend_dir(c: pd.Series, h: pd.Series, l: pd.Series, period: int = 10, mult: float = 3.0) -> int:
"""Supertrend(10,3) final direction: +1 bullish, -1 bearish. Mirrors
predictor_core._st_dir_last / backtest._compute_indicators supertrend block."""
try:
if len(c) < period + 1:
return 0
tr = pd.concat([h - l, (h - c.shift(1)).abs(), (l - c.shift(1)).abs()], axis=1).max(axis=1)
_atr = tr.rolling(period).mean()
hl2 = (h + l) / 2
up_raw = (hl2 + mult * _atr).values
dn_raw = (hl2 - mult * _atr).values
cv = c.values
n = len(cv)
upper, lower = up_raw.copy(), dn_raw.copy()
dirn = 1
for i in range(1, n):
if not (pd.isna(up_raw[i]) or pd.isna(dn_raw[i])):
upper[i] = min(up_raw[i], upper[i - 1]) if cv[i - 1] <= upper[i - 1] else up_raw[i]
lower[i] = max(dn_raw[i], lower[i - 1]) if cv[i - 1] >= lower[i - 1] else dn_raw[i]
if cv[i] > upper[i - 1]:
dirn = 1
elif cv[i] < lower[i - 1]:
dirn = -1
return dirn
except Exception:
return 0
def _consec_days(c: pd.Series) -> float:
"""Signed consecutive-day streak: +n up, -n down, 0 otherwise."""
try:
if len(c) < 6:
return 0.0
diffs = c.iloc[-6:].diff().dropna()
up = dn = 0
for d in reversed(diffs.values):
if d > 0 and dn == 0:
up += 1
elif d < 0 and up == 0:
dn += 1
else:
break
if up >= 1:
return float(up)
if dn >= 1:
return float(-dn)
return 0.0
except Exception:
return 0.0
def _trigger_flags(rsi14, bb_pct, r10, r20, macd, above_ema50, above_ema200, consec) -> dict:
"""Replicates research/backtest.py::_compute_trigger_flags (1D canonical triggers)."""
consec_up = int(consec) if consec and consec > 0 else 0
crash_exhausted = bool(r10 < -6.0 or r20 < -8.0)
overbought_extreme = bool(rsi14 > 70)
T1 = bool(above_ema50 and macd > 0 and not overbought_extreme)
T2 = bool(above_ema50 and r10 > 3.0 and bb_pct < 85.0)
T3 = bool(above_ema50 and consec_up >= 3 and r20 > 0.0)
T4 = bool(rsi14 < 50 and bb_pct < 45.0 and r10 > -2.0 and not crash_exhausted)
T5 = bool(r10 > 7.0 and bb_pct < 80.0)
T6 = bool(rsi14 < 44 and bb_pct < 35.0 and not crash_exhausted)
T7 = bool(above_ema50 and 1.0 <= r20 <= 5.0 and rsi14 < 62.0)
B1 = False # removed from production; kept for schema stability
B2 = bool((not above_ema50) and macd < 0 and r10 < -4.0 and rsi14 > 42 and bb_pct > 40.0)
B3 = bool(crash_exhausted and macd < 0)
return {
"trigger_T1": int(T1), "trigger_T2": int(T2), "trigger_T3": int(T3),
"trigger_T4": int(T4), "trigger_T5": int(T5), "trigger_T6": int(T6),
"trigger_T7": int(T7), "trigger_B1": int(B1), "trigger_B2": int(B2),
"trigger_B3": int(B3),
}
def _last(series: pd.Series, default=_NAN) -> float:
try:
v = float(series.iloc[-1])
return v if np.isfinite(v) else default
except Exception:
return default
# ββ Extra indicator features (ML_EXTRA_FEATURES=1) β vectorized, point-in-time ββ
# These return full backward-looking Series so the SAME math serves both live inference
# (compute_features β last value at `date`) and the training-CSV augmenter
# (research/augment_features.py β value indexed at each sampled date). All windows are
# trailing (rolling/ewm), so the last value of the full series == the value at that date
# with no lookahead. Keep in ONE place to avoid train/serve skew.
def _extra_feature_series(c: pd.Series, h: pd.Series, l: pd.Series, v: pd.Series) -> dict:
out: dict[str, pd.Series] = {}
tp = (h + l + c) / 3.0 # typical price
# Bollinger bandwidth + squeeze regime.
sma20 = c.rolling(20).mean()
std20 = c.rolling(20).std()
bb_up = sma20 + 2 * std20
bb_lo = sma20 - 2 * std20
bandwidth = (bb_up - bb_lo) / sma20.replace(0, np.nan) * 100.0
out["bb_bandwidth"] = bandwidth
# squeeze = bandwidth in the bottom 20% of its trailing 126-bar range (breakout setup).
bw_q20 = bandwidth.rolling(126, min_periods=30).quantile(0.20)
out["bb_squeeze"] = (bandwidth <= bw_q20).astype(float)
# Keltner channel position (EMA20 Β± 2Β·ATR20).
_atr20 = atr(h, l, c, 20)
ema20 = c.ewm(span=20).mean()
kc_up = ema20 + 2 * _atr20
kc_lo = ema20 - 2 * _atr20
out["keltner_pct"] = (c - kc_lo) / (kc_up - kc_lo).replace(0, np.nan) * 100.0
# Donchian channel position (20).
dc_hi = h.rolling(20).max()
dc_lo = l.rolling(20).min()
out["donchian_pct"] = (c - dc_lo) / (dc_hi - dc_lo).replace(0, np.nan) * 100.0
# Stochastic %K (14).
ll14 = l.rolling(14).min()
hh14 = h.rolling(14).max()
out["stoch_k"] = (c - ll14) / (hh14 - ll14).replace(0, np.nan) * 100.0
# CCI (20).
tp_sma = tp.rolling(20).mean()
tp_mad = tp.rolling(20).apply(lambda x: np.mean(np.abs(x - x.mean())), raw=True)
out["cci20"] = (tp - tp_sma) / (0.015 * tp_mad.replace(0, np.nan))
# Money Flow Index (14).
rmf = tp * v
pos_mf = rmf.where(tp.diff() > 0, 0.0).rolling(14).sum()
neg_mf = rmf.where(tp.diff() < 0, 0.0).rolling(14).sum()
mfr = pos_mf / neg_mf.replace(0, np.nan)
out["mfi14"] = 100.0 - 100.0 / (1.0 + mfr)
# Williams %R (14) β mapped to 0..100 (0 = at 14-bar low, 100 = at 14-bar high).
out["williams_r"] = (c - ll14) / (hh14 - ll14).replace(0, np.nan) * 100.0
# Realized (historical) volatility, 20-bar, annualized %.
rets = c.pct_change()
out["hist_vol_20"] = rets.rolling(20).std() * np.sqrt(252) * 100.0
return out
def _mc_features(rets: np.ndarray, horizon: int = 3, n_sims: int = 400) -> tuple:
"""Monte-Carlo bootstrap of forward price paths from the trailing daily returns.
Draws `n_sims` paths of length `horizon` by sampling WITH replacement from the recent
return distribution (non-parametric, captures fat tails / skew unlike a Gaussian), then
measures the forward-excursion distribution the way the labels do:
β’ mc_up_prob = P(best up-excursion over the path > 0) β chance the up-target is reachable
β’ mc_exp_maxup = mean best up-excursion (%) β expected reachable high
β’ mc_exp_maxdn = mean worst down-excursion (%) β expected drawdown
Deterministic (fixed seed) so live inference and the training augmenter agree exactly.
"""
rets = rets[np.isfinite(rets)]
if rets.size < 10:
return (_NAN, _NAN, _NAN)
rng = np.random.default_rng(12345)
draws = rng.choice(rets, size=(n_sims, horizon), replace=True)
paths = np.cumprod(1.0 + draws, axis=1) # cumulative price factor along each path
max_up = (paths.max(axis=1) - 1.0) * 100.0 # best up-excursion per path (%)
min_dn = (paths.min(axis=1) - 1.0) * 100.0 # worst down-excursion per path (%)
return (float(np.mean(max_up > 0.0)) * 100.0, float(np.mean(max_up)), float(np.mean(min_dn)))
def compute_extra_features(c: pd.Series, h: pd.Series, l: pd.Series, v: pd.Series) -> dict:
"""Point-in-time extra-feature dict for the LAST bar of the given series (live path)."""
ser = _extra_feature_series(c, h, l, v)
feat = {k: _last(s, _NAN) for k, s in ser.items()}
rets = c.pct_change().to_numpy()[-63:]
up_prob, exp_up, exp_dn = _mc_features(rets)
feat["mc_up_prob_3d"] = up_prob
feat["mc_exp_maxup_3d"] = exp_up
feat["mc_exp_maxdn_3d"] = exp_dn
return {k: float(feat.get(k, _NAN)) for k in EXTRA_FEATURE_COLS}
def compute_features(
c: pd.Series, h: pd.Series, l: pd.Series, v: pd.Series,
nifty_c: pd.Series | None = None, vix_c: pd.Series | None = None,
date=None,
) -> dict | None:
"""Return the ordered numeric feature dict for one (ticker, date).
c/h/l/v are full daily Close/High/Low/Volume Series (DatetimeIndex).
nifty_c/vix_c are ^NSEI / ^INDIAVIX Close Series (for RS + macro features).
If `date` is given, all series are sliced to `.loc[:date]` (point-in-time).
Returns None if there is too little history (< 26 bars) to compute indicators.
"""
if date is not None:
c = c.loc[:date]
h = h.loc[:date]
l = l.loc[:date]
v = v.loc[:date]
c = c.dropna(); h = h.dropna(); l = l.dropna(); v = v.dropna()
if len(c) < 26:
return None
price = float(c.iloc[-1])
f: dict[str, float] = {}
# ββ Momentum / oscillators ββββββββββββββββββββββββββββββββββββββββββββββ
f["rsi14"] = _last(rsi(c, 14), 50.0)
f["rsi5"] = _last(rsi(c, 5), 50.0)
f["rsi2"] = _last(rsi(c, 2), 50.0)
f["macd_hist"] = _last(macd_h(c), 0.0) if len(c) >= 27 else 0.0
f["adx14"] = _last(adx_s(h, l, c), 20.0) if (len(h) >= 15 and len(l) >= 15) else 20.0
# ββ Trend / EMA (distance in %) βββββββββββββββββββββββββββββββββββββββββ
e20 = float(c.ewm(span=20).mean().iloc[-1])
e50 = float(c.ewm(span=50).mean().iloc[-1]) if len(c) >= 50 else _NAN
e200 = float(c.ewm(span=200).mean().iloc[-1]) if len(c) >= 200 else _NAN
f["price_vs_ema20"] = (price / e20 - 1.0) * 100.0
f["price_vs_ema50"] = (price / e50 - 1.0) * 100.0 if np.isfinite(e50) else _NAN
f["price_vs_ema200"] = (price / e200 - 1.0) * 100.0 if np.isfinite(e200) else _NAN
f["ema_stack"] = _last(ema_stack_score(c), 0.5) if len(c) >= 20 else 0.5
f["supertrend"] = float(1 if _supertrend_dir(c, h, l) > 0 else 0)
# ββ Bollinger position + explicit band-distance (room to band, %) βββββββ
bb_pct = 50.0
bb_lower_dist = _NAN
bb_upper_dist = _NAN
if len(c) >= 20:
sma20 = float(c.rolling(20).mean().iloc[-1])
std20 = float(c.rolling(20).std().iloc[-1])
bb_upper = sma20 + 2 * std20
bb_lower = sma20 - 2 * std20
if bb_upper > bb_lower:
bb_pct = (price - bb_lower) / (bb_upper - bb_lower) * 100.0
bb_lower_dist = (price - bb_lower) / price * 100.0
bb_upper_dist = (bb_upper - price) / price * 100.0
f["bb_pct"] = bb_pct
f["bb_lower_dist"] = bb_lower_dist
f["bb_upper_dist"] = bb_upper_dist
# ββ Relative strength vs Nifty (3M = 63 bars) βββββββββββββββββββββββββββ
rs_3m = 0.0
if nifty_c is not None and len(nifty_c) > 0 and len(c) >= 63:
ni = nifty_c
if date is not None:
ni = ni.loc[:date]
ni = ni.reindex(c.index).ffill()
try:
rs_3m = float((c.iloc[-1] / c.iloc[-63] - 1) - (ni.iloc[-1] / ni.iloc[-63] - 1)) * 100.0
except Exception:
rs_3m = 0.0
f["rs_3m"] = rs_3m
# ββ Volume ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
vol_ratio = 1.0
if len(v) >= 20:
v20 = float(v.rolling(20).mean().iloc[-1])
vol_ratio = float(v.iloc[-1]) / v20 if v20 > 0 else 1.0
f["vol_ratio"] = vol_ratio
obv_z = 0.0
if len(c) >= 20:
ob = obv(c, v)
ob_mu = ob.rolling(20).mean().iloc[-1]
ob_std = ob.rolling(20).std().iloc[-1]
obv_z = float((ob.iloc[-1] - ob_mu) / ob_std) if ob_std and ob_std > 0 else 0.0
f["obv_z"] = float(np.clip(obv_z, -5, 5))
vt = 1.0
if len(v) >= 25:
v20s = v / (v.rolling(20).mean() + 1e-9)
vt = float(v20s.rolling(5).mean().iloc[-1])
f["vol_trend_5d"] = float(np.clip(vt, 0, 5))
# ββ Volatility / range ββββββββββββββββββββββββββββββββββββββββββββββββ
atr14 = _last(atr(h, l, c, 14), _NAN) if (len(h) >= 15 and len(l) >= 15) else _NAN
f["atr_pct"] = (atr14 / price * 100.0) if (np.isfinite(atr14) and price > 0) else _NAN
f["shadow"] = _last(shadow_flag(c, l), 0.0) if len(c) >= 20 else 0.0
# ββ Returns / streak ββββββββββββββββββββββββββββββββββββββββββββββββββββ
f["return_10d"] = (price / float(c.iloc[-10]) - 1) * 100.0 if len(c) >= 10 else _NAN
f["return_20d"] = (price / float(c.iloc[-20]) - 1) * 100.0 if len(c) >= 20 else _NAN
f["return_90d"] = (price / float(c.iloc[-63]) - 1) * 100.0 if len(c) >= 63 else _NAN
f["dist_52w_high"] = (price / float(c.iloc[-252:].max()) - 1) * 100.0 if len(c) >= 252 else _NAN
consec = _consec_days(c)
f["consec_days"] = consec
# ββ Domain trigger flags ββββββββββββββββββββββββββββββββββββββββββββββ
above_ema50 = np.isfinite(e50) and price > e50
above_ema200 = np.isfinite(e200) and price > e200
r10 = f["return_10d"] if np.isfinite(f["return_10d"]) else 0.0
r20 = f["return_20d"] if np.isfinite(f["return_20d"]) else 0.0
f.update(_trigger_flags(f["rsi14"], bb_pct, r10, r20, f["macd_hist"],
above_ema50, above_ema200, consec))
# ββ Macro regime ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
vix_level = _NAN
vix_decl = 0.0
if vix_c is not None and len(vix_c) > 0:
vc = vix_c
if date is not None:
vc = vc.loc[:date]
vc = vc.dropna()
if len(vc) > 0:
vix_level = float(vc.iloc[-1])
if len(vc) >= 10:
vix_ema5 = vc.ewm(span=5).mean()
vix_decl = float(1 if vix_ema5.iloc[-1] < vix_ema5.iloc[-2] else 0)
f["vix_level"] = vix_level
f["vix_decl"] = vix_decl
nifty_ok = 0.0
if nifty_c is not None and len(nifty_c) > 0:
ni = nifty_c
if date is not None:
ni = ni.loc[:date]
ni = ni.dropna()
if len(ni) >= 3:
nema = ni.ewm(span=200).mean().iloc[-1]
nifty_ok = float(1 if float(ni.iloc[-1]) > float(nema) else 0)
f["nifty_ok"] = nifty_ok
# ββ Extra indicator + Monte-Carlo features (ML_EXTRA_FEATURES=1) ββββββββββ
if _USE_EXTRA_FEATS:
try:
f.update(compute_extra_features(c, h, l, v))
except Exception:
f.update({k: _NAN for k in EXTRA_FEATURE_COLS})
# Return in canonical order (any missing key β NaN, tolerated by HistGBM).
return {k: float(f.get(k, _NAN)) for k in FEATURE_COLUMNS}
def features_to_row(feat: dict) -> list[float]:
"""Feature dict β ordered list matching FEATURE_COLUMNS (for model input)."""
return [feat.get(k, _NAN) for k in FEATURE_COLUMNS]
|