Upload runtime.py

runtime.py

@@ -182,6 +182,31 @@ def previous_trading_day(start: date) -> date:
     return pd.Timestamp(schedule.index[-1]).date()
 
 
+def last_n_trading_sessions(end_day: date, count: int) -> list[date]:
+    """Return the last ``count`` NSE sessions ending on (or before) ``end_day``."""
+    sessions: list[date] = []
+    cursor = end_day
+    guard = 0
+    while len(sessions) < count and guard < count * 12:
+        guard += 1
+        if is_trading_day(cursor):
+            sessions.append(cursor)
+        if len(sessions) >= count:
+            break
+        cursor = previous_trading_day(cursor - timedelta(days=1))
+    sessions.reverse()
+    return sessions
+
+
+def _track_record_end_session(now: datetime | None = None) -> date:
+    """Latest session the track record should score (today after the close refresh window)."""
+    now = now or datetime.now(IST)
+    today = now.date()
+    if is_trading_day(today) and now.time() >= CLOSE_REFRESH_READY:
+        return today
+    return expected_completed_daily_date(now)
+
+
 class ProbabilityBlend:
     def __init__(self, models: list[Any], weights: np.ndarray):
         self.models = models
@@ -1296,7 +1321,6 @@ def _rolling_tomorrow_prediction(
 
 
 def build_prediction_track_record(
-    daily: pd.DataFrame,
     sessions: int = 10,
 ) -> list[dict[str, Any]]:
     """Rolling last-N Tomorrow simulation: predict each session, score vs prior close."""
@@ -1306,14 +1330,20 @@ def build_prediction_track_record(
     fallback_prob = float(summary.get("latest_forecast_prob_up", 0.49900560447008563))
     forecaster_by_target = _load_forecaster_predictions_by_target()
 
-    daily_rows = daily.copy()
+    end_session = _track_record_end_session()
+    latest_daily = latest_parquet_date(NIFTY_1D_PATH)
+    if latest_daily is None or latest_daily < end_session:
+        try:
+            refresh_daily_data()
+        except Exception:
+            pass
+
+    daily_rows = pd.read_parquet(NIFTY_1D_PATH)
     daily_rows["date"] = pd.to_datetime(daily_rows["date"], errors="coerce").dt.normalize()
     daily_rows = daily_rows.dropna(subset=["date"]).sort_values("date")
     daily_rows = daily_rows[
         daily_rows["close"].map(lambda value: np.isfinite(float(value)) if pd.notna(value) else False)
     ].copy()
-    completed_day = expected_completed_daily_date()
-    daily_rows = daily_rows[daily_rows["date"].dt.date <= completed_day]
     if daily_rows.empty:
         return []
 
@@ -1323,10 +1353,13 @@ def build_prediction_track_record(
         if pd.notna(row["close"]) and np.isfinite(float(row["close"]))
     }
 
+    session_dates = last_n_trading_sessions(end_session, sessions)
+
     records: list[dict[str, Any]] = []
-    for _, row in daily_rows.tail(sessions).iterrows():
-        target_day = row["date"].date()
-        day_close = float(row["close"])
+    for target_day in session_dates:
+        day_close = closes_by_date.get(target_day)
+        if day_close is None:
+            continue
         actual_move, actual_direction = _tomorrow_actual_outcome(target_day, day_close, closes_by_date)
         if actual_direction is None:
             continue
@@ -1452,7 +1485,7 @@ def _dashboard_payload_cached(key: tuple[tuple[str, int | None, int | None], ...
         "total_test_days": int(tomorrow_summary.get("n_test") or len(tomorrow_test) or 0),
         "models": model_metrics,
     }
-    track_record = build_prediction_track_record(daily, sessions=10)
+    track_record = build_prediction_track_record(sessions=10)
     return {
         "latest": t5_latest,
         "tomorrow_latest": tomorrow_latest,