fix(backtest): replace broken MC permutation test with binomial win-rate test

The previous monte_carlo_trade_pvalue() used sum(permuted_trades) as test
statistic, which is permutation-invariant (sum is commutative), so beat/n
was always 1.0 and MC_p was always 1.00 for every strategy.

Replace with a one-sided binomial test on trade win rate vs 50% baseline.
Tests whether the observed win rate could occur by chance under H0: p=0.5.

Also add _shift_daily_constant_factor_if_needed() to predix_full_eval.py
so re-evaluations apply the look-ahead bias correction for daily factors.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

rdagent/components/backtesting/vbt_backtest.py

@@ -356,11 +356,12 @@ def monte_carlo_trade_pvalue(
     """
     Monte Carlo permutation test on trade-level P&L.
 
-    Shuffles the order of trade returns ``n_permutations`` times and computes
-    the fraction of runs whose total return is >= the real total return.
+    Runs a one-sided binomial test on trade-level win rate.
 
-    p < 0.05 → strategy has a statistically significant edge (real return
-    beats 95% of random sequences with the same set of trades).
+    Tests H0: win_rate = 0.5 (random trading) against H1: win_rate > 0.5.
+    The ``n_permutations`` parameter is kept for API compatibility but is unused.
+
+    p < 0.05 → win rate is significantly above 50%, indicating a genuine per-trade edge.
 
     Parameters
     ----------
@@ -379,14 +380,14 @@ def monte_carlo_trade_pvalue(
     if len(trade_pnl) < 2:
         return 1.0
     trades = trade_pnl.values.copy()
-    real_total = float(trades.sum())
-    rng = np.random.default_rng(seed)
-    beat = 0
-    for _ in range(n_permutations):
-        perm = rng.permutation(trades)
-        if perm.sum() >= real_total:
-            beat += 1
-    return beat / n_permutations
+    # Binomial test: is the win rate significantly above 50%?
+    # p = probability of observing >= n_wins out of n_trades under null (win_rate=0.5).
+    # Low p → strategy has a significant positive edge per trade.
+    from scipy.stats import binomtest
+    n_wins = int((trades > 0).sum())
+    n_total = len(trades)
+    result = binomtest(n_wins, n_total, p=0.5, alternative="greater")
+    return float(result.pvalue)
 
 
 def walk_forward_rolling(

scripts/predix_full_eval.py

@@ -198,6 +198,55 @@ def scan_factors(workspace_dir: Path, skip_evaluated: bool = True) -> List[Facto
     return factors
 
 
+# ---------------------------------------------------------------------------
+# Look-ahead bias detection for daily-constant factors
+# ---------------------------------------------------------------------------
+def _shift_daily_constant_factor_if_needed(factor_col: "pd.Series", factor_name: str) -> "pd.Series":
+    """Detect daily-constant factors (look-ahead bias) and shift by 1 trading day."""
+    sample_days = factor_col.index.get_level_values("datetime").normalize().unique()
+    if len(sample_days) < 10:
+        return factor_col
+    rng = np.random.default_rng(42)
+    days_to_check = rng.choice(sample_days, size=min(50, len(sample_days)), replace=False)
+    constant_count = 0
+    for day in days_to_check:
+        day_mask = factor_col.index.get_level_values("datetime").normalize() == day
+        day_vals = factor_col[day_mask].dropna()
+        if len(day_vals) == 0:
+            continue
+        if day_vals.nunique() == 1:
+            constant_count += 1
+    fraction_constant = constant_count / len(days_to_check)
+    if fraction_constant < 0.90:
+        return factor_col
+    # Shift by 1 trading day per instrument
+    import logging
+    logging.getLogger(__name__).info(
+        "Factor '%s' is %.0f%% daily-constant — shifting 1 trading day to fix look-ahead bias",
+        factor_name, fraction_constant * 100,
+    )
+    instruments = factor_col.index.get_level_values("instrument").unique() if "instrument" in factor_col.index.names else [None]
+    shifted_parts = []
+    for instr in instruments:
+        if instr is not None:
+            mask = factor_col.index.get_level_values("instrument") == instr
+            col_instr = factor_col[mask]
+        else:
+            col_instr = factor_col
+        dates = col_instr.index.get_level_values("datetime").normalize()
+        trading_days = dates.unique().sort_values()
+        day_first = col_instr.groupby(dates).first()
+        day_first_shifted = day_first.shift(1)
+        day_first_shifted.index = pd.to_datetime(day_first_shifted.index)
+        day_map = day_first_shifted.reindex(pd.to_datetime(trading_days)).values
+        new_vals = pd.Series(
+            day_map[np.searchsorted(trading_days.values, dates.values)],
+            index=col_instr.index,
+        )
+        shifted_parts.append(new_vals)
+    return pd.concat(shifted_parts).sort_index()
+
+
 # ---------------------------------------------------------------------------
 # Factor evaluator
 # ---------------------------------------------------------------------------
@@ -263,6 +312,7 @@ def evaluate_factor_full(factor: FactorInfo, full_data: pd.DataFrame,
             result = pd.read_hdf(str(result_file), key="data")
             total_count = len(result)
             factor_val = result.iloc[:, 0]
+            factor_val = _shift_daily_constant_factor_if_needed(factor_val, factor.factor_name)
             non_null_count = factor_val.notna().sum()
 
             if non_null_count < 1000: