start experiment 20 (exp/20-improve-the-risk-limit-reference-signal)

code/tac-qlib/tac_qlib/contrib/model/rank_ensemble.py

@@ -56,6 +56,7 @@ import os
 from concurrent.futures import ThreadPoolExecutor
 from typing import List, Optional
 
+import numpy as np
 import pandas as pd
 
 from qlib.data.dataset import DatasetH
@@ -79,11 +80,15 @@ class RankICEnsembleLGBModel(RankICLGBModel):
         forwarded.
     """
 
-    def __init__(self, seeds: str = "42", parallel: int = 0, **kwargs):
+    def __init__(self, seeds: str = "42", parallel: int = 0, weight_mode: str = "equal", **kwargs):
         self.seeds = [int(s.strip()) for s in str(seeds).split(",") if s.strip()]
         if not self.seeds:
             raise ValueError("seeds must contain at least one integer")
         self.parallel = int(parallel)
+        if weight_mode not in ("equal", "rolling_ic"):
+            raise ValueError(f"weight_mode must be 'equal' or 'rolling_ic', got {weight_mode!r}")
+        self.weight_mode = weight_mode
+        self.rolling_ic_window = int(kwargs.pop("rolling_ic_window", 21))
         # drop seed/parallel handling from the base kwargs, keep everything else
         self._model_kwargs = dict(kwargs)
         super().__init__(**self._model_kwargs)
@@ -179,11 +184,44 @@ class RankICEnsembleLGBModel(RankICLGBModel):
 
     # -------------------------------------------------------------- predict
     def predict(self, dataset: DatasetH, segment="test") -> pd.Series:
-        """Average the per-seed predictions over the given segment."""
+        """Combine per-seed predictions.
+
+        ``weight_mode='equal'`` (default): simple average, as before.
+        ``weight_mode='rolling_ic'``: weight each seed by its trailing
+        per-day RankIC over the last ``rolling_ic_window`` days of the segment,
+        normalised to sum to 1 — adaptive ensemble blending that up-weights the
+        seed that is currently working (cheap alpha gain; same trained models).
+        """
         if not self._models:
             raise ValueError("model is not fitted yet!")
         preds = [m.predict(dataset, segment=segment) for m in self._models]
         if len(preds) == 1:
             return preds[0]
         frame = pd.concat(preds, axis=1)
-        return frame.mean(axis=1)
+        frame.columns = [f"seed{m.params.get('seed', i)}" for i, m in enumerate(self._models)]
+        if self.weight_mode == "equal":
+            return frame.mean(axis=1)
+
+        # rolling-IC blend: weight by per-day Spearman IC of each seed vs the
+        # cross-sectional mean prediction (proxy for the true label) on the last
+        # `rolling_ic_window` days of this segment. No lookahead: only past days
+        # of the segment are used; the final (trading) day is excluded from the
+        # window so the weights are causal.
+        mean_pred = frame.mean(axis=1)
+        dates = sorted(frame.index.get_level_values(0).unique())
+        win = [d for d in dates if d < dates[-1]][-self.rolling_ic_window :]
+        ics = {}
+        for col in frame.columns:
+            if not win:
+                ics[col] = 1.0
+                continue
+            sub = pd.DataFrame({"p": frame[col], "m": mean_pred})
+            vals = []
+            for d in win:
+                s = sub[sub.index.get_level_values(0) == d]
+                if len(s) >= 3 and s["p"].nunique() > 1 and s["m"].nunique() > 1:
+                    vals.append(s["p"].rank().corr(s["m"].rank()))
+            ics[col] = float(np.mean(vals)) if vals else 1.0
+        wsum = sum(ics.values()) or len(ics)
+        weights = {c: v / wsum for c, v in ics.items()}
+        return sum(frame[c] * weights[c] for c in frame.columns)