chore(master): release 1.3.3 (#30)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

.release-please-manifest.json

@@ -1,3 +1,3 @@
 {
-  ".": "1.3.0"
+  ".": "1.3.3"
 }

CHANGELOG.md

@@ -1,5 +1,32 @@
 # Changelog
 
+## [1.3.3](https://github.com/TPTBusiness/Predix/compare/v1.3.2...v1.3.3) (2026-04-25)
+
+
+### Bug Fixes
+
+* **backtest:** replace broken MC permutation test with binomial win-rate test ([c38d894](https://github.com/TPTBusiness/Predix/commit/c38d89478f586825bfca5715a96ca70ccd8791a3))
+* **factors:** detect and correct look-ahead bias in daily-constant factors ([eb490a4](https://github.com/TPTBusiness/Predix/commit/eb490a461b66cbd815ae53ac5205115754712432))
+* **factors:** extend look-ahead rules to session factors and add intraday-factor guidance ([c24c100](https://github.com/TPTBusiness/Predix/commit/c24c100442d6487686c0578de0b32d240fcbf215))
+* **loop:** compress old experiment history in proposal prompt to reduce context size ([4bf90a9](https://github.com/TPTBusiness/Predix/commit/4bf90a905ba8b2aba2a818191c19998088cccaaf))
+* **strategies:** guard against None IC in acceptance check, disable slow wf_rolling ([2197f52](https://github.com/TPTBusiness/Predix/commit/2197f52150a50ef38d9e70991d7e48c8c30caec4))
+* **strategies:** handle None ic/sharpe/dd in rejected strategy log output ([ad2ad3a](https://github.com/TPTBusiness/Predix/commit/ad2ad3ab3360ea75ed3bbc90c12098b9c5cc0114))
+
+## [1.3.2](https://github.com/TPTBusiness/Predix/compare/v1.3.1...v1.3.2) (2026-04-23)
+
+
+### Bug Fixes
+
+* **strategies:** guard against None IC in acceptance check, disable slow wf_rolling ([2197f52](https://github.com/TPTBusiness/Predix/commit/2197f52150a50ef38d9e70991d7e48c8c30caec4))
+* **strategies:** handle None ic/sharpe/dd in rejected strategy log output ([ad2ad3a](https://github.com/TPTBusiness/Predix/commit/ad2ad3ab3360ea75ed3bbc90c12098b9c5cc0114))
+
+## [1.3.1](https://github.com/TPTBusiness/Predix/compare/v1.3.0...v1.3.1) (2026-04-21)
+
+
+### Bug Fixes
+
+* **deps:** bump python-dotenv to >=1.2.2 (CVE symlink overwrite) ([126ae7d](https://github.com/TPTBusiness/Predix/commit/126ae7d5fb556b677d09d10221862a0d648d697a))
+
 ## [1.3.0](https://github.com/TPTBusiness/Predix/compare/v1.2.2...v1.3.0) (2026-04-21)
 
 

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
+    Runs a one-sided binomial test on trade-level win rate.
-    the fraction of runs whose total return is >= the real total return.
 
-    p < 0.05 → strategy has a statistically significant edge (real return
+    Tests H0: win_rate = 0.5 (random trading) against H1: win_rate > 0.5.
-    beats 95% of random sequences with the same set of trades).
+    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())
+    # Binomial test: is the win rate significantly above 50%?
-    rng = np.random.default_rng(seed)
+    # p = probability of observing >= n_wins out of n_trades under null (win_rate=0.5).
-    beat = 0
+    # Low p → strategy has a significant positive edge per trade.
-    for _ in range(n_permutations):
+    from scipy.stats import binomtest
-        perm = rng.permutation(trades)
+    n_wins = int((trades > 0).sum())
-        if perm.sum() >= real_total:
+    n_total = len(trades)
-            beat += 1
+    result = binomtest(n_wins, n_total, p=0.5, alternative="greater")
-    return beat / n_permutations
+    return float(result.pvalue)
 
 
 def walk_forward_rolling(

rdagent/components/coder/factor_coder/prompts.yaml

@@ -53,7 +53,7 @@ evolving_strategy_factor_implementation_v1_system: |-
   - ALWAYS use `min_periods=N` where N equals the window size in rolling calculations (e.g., `.rolling(20, min_periods=20)`)
   - ALWAYS handle infinite values after division: `.replace([np.inf, -np.inf], np.nan)` before saving results
   - ALWAYS use `groupby(level=1)` or `groupby('instrument')` before rolling operations on MultiIndex dataframes
-  - Process the COMPLETE date range (2020-2026), do NOT filter by date
+  - Process the COMPLETE date range available in the HDF5 file (do NOT filter by date — the file may contain 2024 debug data or full 2020-2026 data)
   - Use `groupby().transform()` instead of `groupby().apply()` for single-column assignments
 
   Notice that you should not add any other text before or after the json format.

rdagent/scenarios/qlib/developer/factor_runner.py

@@ -29,6 +29,83 @@ from rdagent.scenarios.qlib.experiment.model_experiment import QlibModelExperime
 DIRNAME = Path(__file__).absolute().resolve().parent
 DIRNAME_local = Path.cwd()
 
+
+def _shift_daily_constant_factor_if_needed(factor_col: "pd.Series", factor_name: str) -> "pd.Series":
+    """Detect and fix look-ahead bias in daily-constant factors.
+
+    A factor is "daily-constant" when every minute bar within the same calendar
+    day carries an identical value. This happens when LLM code computes a daily
+    aggregate (e.g. today's log return) and forward-fills it across all intraday
+    bars without shifting — meaning the end-of-day value is visible at 00:00.
+
+    Fix: shift by one trading day so that the value assigned to day T is the
+    aggregate computed from day T-1, eliminating the forward-looking information.
+    """
+    import numpy as np
+
+    try:
+        notnull = factor_col.dropna()
+        if len(notnull) < 200:
+            return factor_col
+
+        datetimes = notnull.index.get_level_values("datetime")
+        dates = datetimes.normalize()
+
+        # Sample up to 50 random days and check intra-day uniqueness
+        unique_dates = pd.Series(dates.unique())
+        sample_dates = unique_dates.sample(min(50, len(unique_dates)), random_state=42)
+
+        daily_unique_counts = []
+        for d in sample_dates:
+            mask = dates == d
+            vals = notnull.values[mask]
+            if len(vals) > 1:
+                daily_unique_counts.append(len(np.unique(vals[~np.isnan(vals)])))
+
+        if not daily_unique_counts:
+            return factor_col
+
+        # If >90% of sampled days have exactly 1 unique value → daily-constant
+        fraction_constant = sum(1 for c in daily_unique_counts if c == 1) / len(daily_unique_counts)
+        if fraction_constant < 0.90:
+            return factor_col  # Intraday factor — no shift needed
+
+        logger.warning(
+            f"[LookAheadFix] Factor '{factor_name}' is daily-constant "
+            f"({fraction_constant:.0%} of days). Applying 1-day shift to remove look-ahead bias."
+        )
+
+        # Shift: for each instrument, map daily values forward by 1 trading day
+        instruments = factor_col.index.get_level_values("instrument").unique()
+        shifted_parts = []
+        for inst in instruments:
+            inst_series = factor_col.xs(inst, level="instrument")
+            # Get one value per calendar day (the first non-null bar)
+            inst_dt = inst_series.index.normalize()
+            daily_vals = inst_series.groupby(inst_dt).first()
+            # Shift by 1 day
+            daily_vals_shifted = daily_vals.shift(1)
+            # Forward-fill back to minute bars
+            minute_idx = inst_series.index
+            minute_dates = minute_idx.normalize()
+            shifted_minute = minute_dates.map(daily_vals_shifted)
+            shifted_s = pd.Series(
+                shifted_minute.values,
+                index=pd.MultiIndex.from_arrays(
+                    [inst_series.index, [inst] * len(inst_series)],
+                    names=["datetime", "instrument"],
+                ),
+                name=factor_col.name,
+            )
+            shifted_parts.append(shifted_s)
+
+        return pd.concat(shifted_parts).sort_index()
+
+    except Exception as e:
+        logger.debug(f"[LookAheadFix] Could not apply daily shift for '{factor_name}': {e}")
+        return factor_col
+
+
 # TODO: supporting multiprocessing and keep previous results
 
 
@@ -391,8 +468,19 @@ class QlibFactorRunner(CachedRunner[QlibFactorExperiment]):
         import numpy as np
 
         try:
-            # Get workspace path
+            # Get workspace path — factor code and result.h5 live in sub_workspace_list[0],
-            workspace_path = exp.experiment_workspace.workspace_path
+            # not in experiment_workspace (which is the Qlib template workspace).
+            workspace_path = None
+            if exp.sub_workspace_list:
+                for ws in exp.sub_workspace_list:
+                    if ws is not None and hasattr(ws, 'workspace_path'):
+                        candidate = ws.workspace_path / "result.h5"
+                        if candidate.exists():
+                            workspace_path = ws.workspace_path
+                            break
+            if workspace_path is None:
+                # Fallback to experiment_workspace
+                workspace_path = exp.experiment_workspace.workspace_path
             if workspace_path is None:
                 return None
 
@@ -409,6 +497,12 @@ class QlibFactorRunner(CachedRunner[QlibFactorExperiment]):
             factor_col = factor_values.iloc[:, 0]
             factor_name = factor_values.columns[0]
 
+            # Detect and fix look-ahead bias in daily-constant factors.
+            # If a factor has the same value for all minute bars within each calendar day
+            # it was computed from same-day data (e.g. today's close return at 00:00).
+            # Fix: shift by 1 trading day so value at day T = aggregate of day T-1.
+            factor_col = _shift_daily_constant_factor_if_needed(factor_col, factor_name)
+
             # Load source data for forward returns
             data_path = (
                 Path(__file__).parent.parent.parent.parent.parent
@@ -587,10 +681,12 @@ class QlibFactorRunner(CachedRunner[QlibFactorExperiment]):
             from pathlib import Path
             from rdagent.components.backtesting import ResultsDatabase
 
-            # Get factor name from hypothesis
+            # Get factor name: prefer hypothesis, fallback to result Series 'factor_name' key
             factor_name = "unknown"
             if hasattr(exp, 'hypothesis') and exp.hypothesis is not None:
                 factor_name = getattr(exp.hypothesis, 'hypothesis', 'unknown')
+            if factor_name == 'unknown' and isinstance(result, pd.Series) and 'factor_name' in result.index:
+                factor_name = str(result['factor_name'])
 
             # Check if already rejected by protection
             if getattr(exp, 'rejected_by_protection', False):
@@ -824,41 +920,74 @@ class QlibFactorRunner(CachedRunner[QlibFactorExperiment]):
         """
         Save factor time-series values as parquet for strategy building.
 
-        This is essential for walk-forward validation and strategy combination.
+        Reruns the factor code on the FULL 6-year dataset so the parquet covers
-
+        the complete backtest range (not just the debug 2024 subset).
-        Parameters
-        ----------
-        factor_name : str
-            Name of the factor
-        exp : QlibFactorExperiment
-            The experiment with factor values
         """
         import os as _os
+        import subprocess
+        import shutil
+        import tempfile
 
         try:
-            # Get workspace path
+            # factor.py lives in sub_workspace_list[0], not experiment_workspace
-            workspace_path = exp.experiment_workspace.workspace_path
+            workspace_path = None
+            if exp.sub_workspace_list:
+                for ws in exp.sub_workspace_list:
+                    if ws is not None and hasattr(ws, 'workspace_path'):
+                        fp = ws.workspace_path / "factor.py"
+                        if fp.exists():
+                            workspace_path = ws.workspace_path
+                            break
+            if workspace_path is None:
+                workspace_path = exp.experiment_workspace.workspace_path
             if workspace_path is None:
                 return
 
-            result_h5 = workspace_path / "result.h5"
+            factor_py = workspace_path / "factor.py"
-            if not result_h5.exists():
+            if not factor_py.exists():
                 return
 
-            # Read factor values
+            project_root = Path(__file__).parent.parent.parent.parent.parent
+            full_data = (
+                project_root
+                / "git_ignore_folder"
+                / "factor_implementation_source_data"
+                / "intraday_pv.h5"
+            )
+            if not full_data.exists():
+                return
+
+            # Run factor code on full data in a temp workspace
             import pandas as pd
-            df = pd.read_hdf(str(result_h5), key="data")
+            with tempfile.TemporaryDirectory(prefix="predix_fullval_") as tmp_dir:
+                tmp = Path(tmp_dir)
+                shutil.copy(str(factor_py), str(tmp / "factor.py"))
+                shutil.copy(str(full_data), str(tmp / "intraday_pv.h5"))
+
+                ret = subprocess.run(
+                    ["python", "factor.py"],
+                    cwd=str(tmp),
+                    capture_output=True,
+                    timeout=300,
+                )
+                if ret.returncode != 0:
+                    # Fall back to debug-data result if full-data run fails
+                    result_h5 = workspace_path / "result.h5"
+                    if not result_h5.exists():
+                        return
+                    df = pd.read_hdf(str(result_h5), key="data")
+                else:
+                    result_h5_full = tmp / "result.h5"
+                    if not result_h5_full.exists():
+                        return
+                    df = pd.read_hdf(str(result_h5_full), key="data")
+
             if df is None or df.empty:
                 return
 
-            # Get the factor series (first column)
             series = df.iloc[:, 0]
             series.name = factor_name
 
-            # Save to results/factors/values/
-            project_root = Path(__file__).parent.parent.parent.parent.parent
-
-            # Parallel run isolation
             parallel_run_id = _os.getenv("PARALLEL_RUN_ID", "0")
             if parallel_run_id != "0":
                 values_dir = project_root / "results" / "runs" / f"run{parallel_run_id}" / "factors" / "values"
@@ -866,16 +995,11 @@ class QlibFactorRunner(CachedRunner[QlibFactorExperiment]):
                 values_dir = project_root / "results" / "factors" / "values"
 
             values_dir.mkdir(parents=True, exist_ok=True)
-
-            # Safe filename
             safe_name = factor_name.replace("/", "_").replace("\\", "_").replace(" ", "_")[:100]
             parquet_path = values_dir / f"{safe_name}.parquet"
+            series.to_frame().to_parquet(str(parquet_path))
 
-            # Save as parquet (with datetime index)
+        except Exception:
-            series.to_parquet(str(parquet_path))
-
-        except Exception as e:
-            # Don't let factor value saving break the main workflow
             pass
 
     def _log_result_warnings(self, factor_name: str, result, metrics: dict) -> None:

rdagent/scenarios/qlib/experiment/factor_data_template/README.md

@@ -23,14 +23,25 @@ $low: low price at 1-minute bar.
 $volume: volume at 1-minute bar (tick volume for FX).
 
 ## Important Notes for 1min Data
-- 96 bars = 1 trading day (24 hours for FX)
+- 1 bar = 1 minute (confirmed)
 - 16 bars = 16 minutes
-- 4 bars = 4 minutes
+- 60 bars = 1 hour
-- 1 bar = 1 minute
+- ~1440 bars = 1 full trading day (FX trades nearly 24h, Mon 00:00 - Fri 22:00 UTC approx.)
+- Typical bars per calendar day: ~1200-1440 (varies by weekday, holidays have fewer)
+- Do NOT assume 96 bars/day — the actual count depends on the date
 - Data range: 2020-01-01 to 2026-03-20
 - Instrument: EURUSD
 - Timezone: UTC
 
+## IMPORTANT: Bars per Day Correction
+The dataset has approximately 1440 bars per full trading day (1 bar = 1 minute, ~24h of FX trading).
+Some older documentation incorrectly stated "96 bars = 1 day" — this is WRONG. Always use:
+- 60 bars = 1 hour
+- 480 bars = 8 hours (London session 08:00-16:00 UTC)
+- 180 bars = 3 hours (London/NY overlap 13:00-16:00 UTC)
+Use datetime hour filtering (e.g., `df[df.index.get_level_values('datetime').hour.between(8, 15)]`)
+to select session bars — do NOT use bar-count offsets to define sessions.
+
 ## Session Times (UTC)
 - Asian: 00:00-08:00 UTC (low volatility)
 - London: 08:00-16:00 UTC (high volatility)

rdagent/scenarios/qlib/experiment/prompts.yaml

@@ -104,7 +104,7 @@ qlib_factor_strategy: |-
      result_df.columns = ['daily_volume_price_divergence']
      ```
 
-  4. **Process ALL data — do not filter dates**: The source HDF5 contains data from 2020-01-01 to 2026-03-20. Do NOT filter to a single year. If your output has only 314 entries (one year of daily data), the factor will be rejected. Expected output: ~1500+ daily entries for 2020-2026.
+  4. **Process ALL data — do not filter dates**: The source HDF5 contains data from 2020-01-01 to 2026-03-20 (development runs may use a 2024-only debug dataset with ~300 entries, which is acceptable). Do NOT filter to a single year in your code. Write your code to process whatever date range is available in the HDF5 file — do not hardcode date filters. Expected output for production data: ~1500+ daily entries for 2020-2026. Expected output for debug data: ~300 daily entries for 2024. Both are valid.
 
   5. **Use `transform()` instead of `apply()` for per-group calculations**: `transform()` preserves the original index while `apply()` may reduce the number of rows unexpectedly:
      ```python
@@ -121,6 +121,35 @@ qlib_factor_strategy: |-
      assert result_df.index.names == ['datetime', 'instrument'], f"Index names must be ['datetime', 'instrument'], got {result_df.index.names}"
      ```
 
+  7. **NEVER use same-day aggregations as the factor value — always shift by 1 day**: If your factor computes a daily aggregate (e.g. daily close return, daily OHLC range, daily volume), that aggregate is only known at end-of-day. Using it at the start of the same day is look-ahead bias. You MUST shift the daily aggregate by 1 day before forward-filling to minute bars:
+     ```python
+     # WRONG: look-ahead bias! Today's close return is not known at 00:00
+     daily_ret = df['$close'].groupby(level='instrument').resample('1D', level='datetime').last().pct_change()
+     result_df['my_factor'] = daily_ret.groupby(level='instrument').transform(lambda x: x.reindex(df.index.get_level_values('datetime'), method='ffill'))
+
+     # CORRECT: shift by 1 trading day so factor value at day T = aggregate of day T-1
+     daily_close = df.groupby([df.index.get_level_values('datetime').normalize(), df.index.get_level_values('instrument')])['$close'].last()
+     daily_close.index.names = ['date', 'instrument']
+     daily_ret = daily_close.groupby(level='instrument').pct_change().shift(1)  # <-- shift(1) is MANDATORY
+     # then map back to minute bars via ffill
+     ```
+     This rule applies to ALL daily aggregations: returns, OHLC stats, volume, momentum, slopes, etc.
+     **Session-based aggregations (London, NY, Asian session returns) are also daily aggregations** — the London
+     session (08:00-16:00 UTC) ends at 16:00, so its return must be shifted by 1 day before use.
+     Intraday rolling factors (e.g. 30-min rolling std computed at bar t using only bars t-N..t-1) do NOT need this shift.
+
+  8. **PREFER pure intraday rolling factors**: Factors that use only a trailing window of recent bars (e.g.
+     rolling(30).mean() of returns, RSI(14), Bollinger Band z-score) have NO look-ahead risk and vary every
+     minute. These are the best candidates for short-horizon (60-180 bar) prediction. Examples:
+     - Rolling 15-min / 30-min / 60-min return momentum (15, 30, 60 bars respectively)
+     - Rolling volatility (std of returns over 20-60 bars)
+     - Distance of close from N-bar moving average (z-score)
+     - RSI or similar oscillators computed on 1-min bars
+     - VWAP deviation (requires volume — use $volume column)
+     Always use `.shift(1)` on the lagged window (e.g. `rolling(N).mean().shift(1)`) to avoid using the
+     current bar's own price in its own feature value.
+     NOTE: 1 bar = 1 minute. The data has ~1440 bars per full trading day. Do NOT use 96 as a day proxy.
+
 qlib_factor_output_format: |-
   Your output should be a pandas dataframe similar to the following example information:
   <class 'pandas.core.frame.DataFrame'>

rdagent/scenarios/qlib/proposal/quant_proposal.py

@@ -152,9 +152,41 @@ class QlibQuantHypothesisGen(FactorAndModelHypothesisGen):
                         factor_inserted = True
             if len(specific_trace.hist) > 0:
                 specific_trace.hist.reverse()
-                hypothesis_and_feedback = T("scenarios.qlib.prompts:hypothesis_and_feedback").r(
+                # Keep only the 2 most recent experiments in full detail; compress older ones
-                    trace=specific_trace,
+                # to brief bullet points to stay within the LLM context window.
-                )
+                FULL_DETAIL_COUNT = 2
+                old_hist = specific_trace.hist[:-FULL_DETAIL_COUNT] if len(specific_trace.hist) > FULL_DETAIL_COUNT else []
+                recent_hist = specific_trace.hist[-FULL_DETAIL_COUNT:] if len(specific_trace.hist) > FULL_DETAIL_COUNT else specific_trace.hist
+
+                parts = []
+                if old_hist:
+                    summary_lines = ["## Earlier experiments (summarized):"]
+                    for exp, fb in old_hist:
+                        factor_names = []
+                        for task in exp.sub_tasks:
+                            if task is not None and hasattr(task, "factor_name"):
+                                factor_names.append(task.factor_name)
+                            elif task is not None and hasattr(task, "model_type"):
+                                factor_names.append(getattr(task, "model_type", "model"))
+                        names_str = ", ".join(factor_names) if factor_names else "unknown"
+                        ic_str = ""
+                        try:
+                            if exp.result is not None:
+                                ic_val = exp.result.loc["IC"] if "IC" in exp.result.index else ""
+                                ic_str = f" IC={ic_val:.4f}" if ic_val != "" else ""
+                        except Exception:
+                            pass
+                        decision_str = "PASS" if fb.decision else "FAIL"
+                        obs_short = (fb.observations or "")[:120].replace("\n", " ")
+                        summary_lines.append(f"- [{decision_str}]{ic_str} {names_str}: {obs_short}")
+                    parts.append("\n".join(summary_lines))
+
+                if recent_hist:
+                    recent_trace = Trace(specific_trace.scen)
+                    recent_trace.hist = recent_hist
+                    parts.append(T("scenarios.qlib.prompts:hypothesis_and_feedback").r(trace=recent_trace))
+
+                hypothesis_and_feedback = "\n\n".join(parts)
             else:
                 hypothesis_and_feedback = "No previous hypothesis and feedback available."
 

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:

scripts/predix_gen_strategies_real_bt.py

@@ -250,7 +250,7 @@ Hard requirements:
 - NO global mean/std — always use rolling(window).mean() with shift(1) to avoid look-ahead bias"""
 
         else:
-            system_prompt = f"""You are a quantitative trading expert specializing in EUR/USD intraday strategies.
+            system_prompt = f"""You are a quantitative trading expert specializing in EUR/USD daily swing strategies.
 
 CRITICAL RULES for {STYLE_DESC} (forward horizon: {FORWARD_BARS} bars = ~{FORWARD_BARS/60:.1f} hours):
 1. ONLY use the factors listed below - no others!
@@ -258,15 +258,27 @@ CRITICAL RULES for {STYLE_DESC} (forward horizon: {FORWARD_BARS} bars = ~{FORWAR
 3. Create a pandas Series called 'signal' with values: 1 (long), -1 (short), 0 (neutral)
 4. signal.index MUST match close.index
 5. signal.name must be 'signal'
+6. IMPORTANT: factors are DAILY values broadcast to every 1-minute bar — they change once per day.
+   Use daily-level logic: compare today's factor value to a rolling daily mean (window 5-20 DAYS).
+   To get daily rolling mean: group by date, take first value per day, compute rolling, then reindex back.
+   Example: dates = factors[col].index.get_level_values('datetime').normalize()
+            daily_vals = factors[col].groupby(dates).first()
+            daily_mean = daily_vals.rolling(10).mean().shift(1)
+            daily_signal = (daily_vals > daily_mean).astype(int) * 2 - 1
+            signal = daily_signal.reindex(dates).values  (broadcast back to minute bars)
+7. The signal should change roughly once per day — this produces ~250-500 trades over 6 years.
+8. Keep conditions SIMPLE: one factor above/below its N-day rolling average. Avoid combining 3+ conditions.
 
 Output ONLY valid JSON with these fields:
 {{"strategy_name": "short_name", "factor_names": ["f1", "f2"], "description": "one sentence", "code": "python code"}}"""
 
-            user_prompt = f"""Create a EUR/USD trading strategy using these factors:
+            user_prompt = f"""Create a EUR/USD SWING trading strategy (hold ~{FORWARD_BARS/60:.0f} hours) using these factors:
 
 {factor_list}
 
-{f'Previous feedback: {feedback}' if feedback else 'First attempt - be creative!'}"""
+{f'Previous feedback: {feedback}' if feedback else 'First attempt - be creative!'}
+
+Use daily-level signal logic (factor above/below rolling daily mean). Signal changes once per day."""
         
         api = APIBackend()
         response = api.build_messages_and_create_chat_completion(
@@ -371,8 +383,8 @@ signal.fillna(0).to_pickle('signal.pkl')
         txn_cost_bps=TXN_COST_BPS,
         forward_returns=fwd_returns,
         oos_start=OOS_START_DEFAULT,
-        wf_rolling=True,
+        wf_rolling=False,   # too slow on 2M bars — run via rebacktest script instead
-        mc_n_permutations=200,
+        mc_n_permutations=50,
     )
 
 # ============================================================================
@@ -579,7 +591,7 @@ def main(target_count=10):
                 # Check acceptance criteria — OOS must be profitable + statistically significant
                 mc_ok = mc_pvalue is None or mc_pvalue < 0.20  # lenient: top 20% non-random
                 wf_ok = wf_consistency is None or wf_consistency >= 0.5  # ≥50% of WF windows profitable
-                if (abs(ic) > MIN_IC and sharpe > MIN_SHARPE and trades > MIN_TRADES and dd > MAX_DRAWDOWN
+                if (abs(ic or 0) > MIN_IC and sharpe > MIN_SHARPE and trades > MIN_TRADES and dd > MAX_DRAWDOWN
                         and oos_sharpe > 0.0 and oos_monthly > 0.0 and mc_ok and wf_ok):
                     # ACCEPT
                     strategy['real_backtest'] = bt_result
@@ -636,9 +648,10 @@ def main(target_count=10):
                     oos_info = f"OOS_Sharpe={oos_sharpe:+.2f} OOS_Mon={oos_monthly:+.2f}%" if oos_sharpe is not None else ""
                     mc_info  = f" MC_p={mc_pvalue:.2f}" if mc_pvalue is not None else ""
                     wf_info  = f" WF_consistency={wf_consistency:.0%}" if wf_consistency is not None else ""
-                    _log.info(f"REJECTED  IC={ic:.4f}  Sharpe={sharpe:.2f}  Trades={trades}  DD={dd:.1%}  {oos_info}{mc_info}{wf_info}")
+                    _ic = ic or 0; _sh = sharpe or 0; _dd = dd or 0
+                    _log.info(f"REJECTED  IC={_ic:.4f}  Sharpe={_sh:.2f}  Trades={trades}  DD={_dd:.1%}  {oos_info}{mc_info}{wf_info}")
                     feedback_history.append(
-                        f"Failed: IC={ic:.4f}, Sharpe={sharpe:.2f}, Trades={trades}, DD={dd:.1%}, "
+                        f"Failed: IC={_ic:.4f}, Sharpe={_sh:.2f}, Trades={trades}, DD={_dd:.1%}, "
                         f"OOS_Sharpe={oos_sharpe:+.2f}, OOS_Monthly={oos_monthly:+.2f}%"
                         + (f", MC_p={mc_pvalue:.2f}" if mc_pvalue is not None else "")
                         + (f", WF_consistency={wf_consistency:.0%}" if wf_consistency is not None else "")