Add 6 features: horizon returns + overnight/intraday + dist-from-high (#36)

Predictor ROADMAP P2 diagnostic — test whether 5d forward is reversal
regime vs momentum regime, and whether splitting close-to-close returns
into overnight/intraday components improves signal.

New features (all technical group):

  return_60d, return_120d
    Longer-horizon momentum. Neutral name — meta ridge coefficient sign
    determines regime. At 5d forward, short-horizon returns load negative
    (reversal). If 60d/120d load positive, momentum persists at longer
    lookback — a well-documented pattern (Jegadeesh/Titman 1993).

  overnight_return_5d, intraday_return_5d
    5d sum of (Open_t / Close_{t-1} - 1) vs (Close_t / Open_t - 1).
    Lou/Polk/Skouras 2019 "A Tug of War" found overnight persists
    positive (earnings, news, macro) while intraday is noisier and often
    negative (microstructure, flow). Total momentum_5d ≈ overnight_5d +
    intraday_5d. Decomposing lets the model learn different dynamics.
    NaN when Open column is missing (no silent zero-fill per
    feedback_no_silent_fails).

  dist_from_5d_high, dist_from_20d_high
    Reversal-native signals. Distance from recent peak as fraction:
    (Close - rolling_max(High, N)) / rolling_max(High, N). Always ≤ 0.
    A stock at its 5d high has no short-term reversal room; a stock
    pulled back has more. Conceptually cleaner than past returns for
    reversal signal.

Registry: 6 FeatureEntry rows added under "v3.1 technical additions".
FEATURES list in feature_engineer.py goes from 53 → 59. dropna still
correct — rows missing any required feature are dropped.

## Test plan
- [x] Synthetic OHLCV smoke: all 6 features compute, values in
      sensible ranges (dist_from_5d_high ≤ 0 always, overnight/intraday
      small magnitudes, return_60d/120d larger).
- [x] Full suite: 43 passed.
- [ ] After merge: re-run alpha-engine-data backfill to populate
      historical rows in ArcticDB with the new columns.
- [ ] Predictor PR B2 (follow-up) adds features to MOMENTUM_FEATURES
      list + 21d forward IC diagnostic.

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: cipher813

features/feature_engineer.py

@@ -120,6 +120,17 @@
     "gross_margin",
     "roe",
     "current_ratio",
+    # v3.1 additions — longer-horizon + overnight/intraday decomposition +
+    # reversal-native signals. Predictor ROADMAP P2: collapse FLAT +
+    # test whether 5d is reversal or momentum regime. 2026-04-15: neutral
+    # names chosen — meta ridge coefficient sign determines whether the
+    # feature behaves as reversal (positive coef) or momentum (negative).
+    "return_60d",
+    "return_120d",
+    "overnight_return_5d",
+    "intraday_return_5d",
+    "dist_from_5d_high",
+    "dist_from_20d_high",
 ]
 
 MIN_ROWS_FOR_FEATURES = 265  # 252 warmup + buffer
@@ -272,6 +283,55 @@ def compute_features(
     _mom_short = _FC["momentum_short"]
     df["momentum_5d"] = (close / close.shift(_mom_short)) - 1.0
 
+    # ── v3.1: Longer-horizon returns ──────────────────────────────────────────
+    # ROADMAP P2 diagnostic — test whether 5d is the right label horizon.
+    # Neutral naming: meta ridge coefficient sign determines reversal vs
+    # momentum regime. Positive coef → reversal (high past returns predict
+    # negative future returns). Negative coef → momentum persists at this
+    # horizon.
+    df["return_60d"] = (close / close.shift(60)) - 1.0
+    df["return_120d"] = (close / close.shift(120)) - 1.0
+
+    # ── v3.1: Overnight / intraday decomposition ──────────────────────────────
+    # Lou/Polk/Skouras 2019 "A Tug of War": overnight returns
+    # (Open_t vs Close_{t-1}) have been historically persistent and positive
+    # (earnings, news, macro), while intraday returns (Close_t vs Open_t)
+    # have been noisier and often negative (microstructure, flow). Total
+    # 5d return = overnight_5d + intraday_5d (approximately — compounding
+    # differences are small at 5d horizons and this additive sum is the
+    # form used in the source literature).
+    if "Open" in df.columns:
+        open_ = df["Open"].astype(float)
+        overnight_daily = (open_ / close.shift(1)) - 1.0
+        intraday_daily = (close / open_) - 1.0
+        df["overnight_return_5d"] = overnight_daily.rolling(
+            window=_mom_short, min_periods=_mom_short,
+        ).sum()
+        df["intraday_return_5d"] = intraday_daily.rolling(
+            window=_mom_short, min_periods=_mom_short,
+        ).sum()
+    else:
+        # Without Open, these features are undefined — NaN propagates and
+        # dropna will exclude the ticker. No silent zero-fill (per
+        # feedback_no_silent_fails).
+        df["overnight_return_5d"] = float("nan")
+        df["intraday_return_5d"] = float("nan")
+
+    # ── v3.1: Distance from recent highs (reversal-native) ────────────────────
+    # Distance from recent peak is a cleaner reversal signal than past
+    # returns: a stock at its 5d high has nowhere to "continue" in the
+    # short-term reversal regime, while a stock pulled back from its 5d
+    # high has more room to mean-revert. Negative values always (close
+    # cannot exceed max by definition). Closer to zero = near high.
+    if "High" in df.columns:
+        high_col = df["High"].astype(float)
+    else:
+        high_col = close
+    rolling_max_5 = high_col.rolling(window=5, min_periods=5).max()
+    rolling_max_20 = high_col.rolling(window=20, min_periods=20).max()
+    df["dist_from_5d_high"] = (close - rolling_max_5) / rolling_max_5
+    df["dist_from_20d_high"] = (close - rolling_max_20) / rolling_max_20
+
     # ── Relative volume ratio ──────────────────────────────────────────────────
     rolling_mean_vol_20 = volume.rolling(window=_vol_slow, min_periods=_vol_slow).mean()
     df["rel_volume_ratio"] = volume / rolling_mean_vol_20.replace(0, float("nan"))

features/registry.py

@@ -85,6 +85,14 @@ class FeatureEntry:
     FeatureEntry("iv_rank", "alternative", "IV percentile rank (0-1)", source="yfinance", refresh="weekly"),
     FeatureEntry("iv_vs_rv", "alternative", "Implied vol / realized vol ratio", source="yfinance", refresh="weekly"),
 
+    # ── v3.1 technical additions — horizon + decomposition + reversal-native ──
+    FeatureEntry("return_60d", "technical", "60-day price return (Close_t / Close_{t-60} - 1)", source="yfinance", refresh="daily"),
+    FeatureEntry("return_120d", "technical", "120-day price return (Close_t / Close_{t-120} - 1)", source="yfinance", refresh="daily"),
+    FeatureEntry("overnight_return_5d", "technical", "5d sum of overnight returns (Open_t vs Close_{t-1})", source="yfinance", refresh="daily"),
+    FeatureEntry("intraday_return_5d", "technical", "5d sum of intraday returns (Close_t vs Open_t)", source="yfinance", refresh="daily"),
+    FeatureEntry("dist_from_5d_high", "technical", "(Close - 5d rolling max High) / 5d rolling max High", source="yfinance", refresh="daily"),
+    FeatureEntry("dist_from_20d_high", "technical", "(Close - 20d rolling max High) / 20d rolling max High", source="yfinance", refresh="daily"),
+
     # ── Fundamental (8) — quarterly financials ────────────────────────────────
     FeatureEntry("pe_ratio", "fundamental", "Trailing P/E ratio, normalized (PE / 30)", source="fmp", refresh="quarterly"),
     FeatureEntry("pb_ratio", "fundamental", "Price-to-book ratio, normalized (PB / 5)", source="fmp", refresh="quarterly"),