upd: 更新了数据库下载组件获取未来成分股的 bug，修改了 IC threshold

Author: Parsnip77

Instruction.md

@@ -361,7 +361,7 @@
   | 1 | 载入 `prices.parquet` 与 `factors_clean.parquet` |
   | 2 | 调用 `calc_forward_return(prices_df, d=5)` 生成 target |
   | 3 | 遍历每个 alpha 列，依次计算 IC 时间序列、IC metrics，展示图表 |
-  | 4 | 筛选满足 `abs(IC mean) > 0.015` 且 `abs(ICIR) > 0.15` 的因子并输出列表 |
+  | 4 | 筛选满足 `abs(IC mean) > 0.01` 且 `abs(ICIR) > 0.15` 的因子并输出列表 |
 
 - **使用**：
   ```bash
@@ -374,7 +374,7 @@
   | 变量 | 默认值 | 说明 |
   |------|--------|------|
   | `FORWARD_DAYS` | `5` | 未来收益率天数 |
-  | `IC_MEAN_THRESHOLD` | `0.015` | IC 均值绝对值阈值 |
+  | `IC_MEAN_THRESHOLD` | `0.01` | IC 均值绝对值阈值 |
   | `ICIR_THRESHOLD` | `0.15` | ICIR 绝对值阈值 |
   | `SHOW_PLOTS` | `False` | 是否交互展示 IC 图表 |
 

analyze_main.py

@@ -33,7 +33,7 @@
 DATA_DIR  = ROOT / "data"
 PLOTS_DIR = ROOT / "plots"
 FORWARD_DAYS = 5          # d-day forward return
-IC_MEAN_THRESHOLD = 0.015  # minimum |IC mean| to keep a factor
+IC_MEAN_THRESHOLD = 0.01  # minimum |IC mean| to keep a factor
 ICIR_THRESHOLD = 0.15     # minimum |ICIR| to keep a factor
 SHOW_PLOTS = False         # set False to suppress interactive charts
 COMBINE_WINDOW = 3       # rolling OLS training window (trading days)
@@ -214,6 +214,21 @@ def main() -> None:
         _ok(f"Synthetic factor : {synth_df.shape[0]:>7,} rows  "
             f"(dates: {synth_df['trade_date'].nunique()})")
 
+        # Apply a 3-day rolling mean per stock to smooth the signal and
+        # reduce day-to-day portfolio turnover.  Rows with fewer than 3
+        # history points (first 2 days per stock) are dropped so the
+        # backtester only operates on fully-smoothed scores.
+        '''
+        synth_df = synth_df.sort_values(["ts_code", "trade_date"]).reset_index(drop=True)
+        synth_df["synthetic_factor"] = (
+            synth_df.groupby("ts_code")["synthetic_factor"]
+            .transform(lambda s: s.rolling(window=3, min_periods=3).mean())
+        )
+        synth_df = synth_df.dropna(subset=["synthetic_factor"]).reset_index(drop=True)
+        _info(f"  After 3-day smoothing : {synth_df.shape[0]:,} rows  "
+              f"(dates: {synth_df['trade_date'].nunique()})")
+        '''
+
         bt_synth = LayeredBacktester(synth_df, target_df, forward_days=FORWARD_DAYS, plots_dir=PLOTS_DIR)
         perf_synth = bt_synth.run_backtest()
         _info("  Backtest metrics (synthetic factor):")

src/data_loader.py

@@ -131,17 +131,25 @@ def init_db(self) -> None:
     # ------------------------------------------------------------------
 
     def _get_constituents(self) -> List[str]:
-        """Return list of ts_codes for the configured universe index."""
-        end = datetime.now()
-        start = end - timedelta(days=31)
+        """Return the index constituents as of the backtest start date.
+
+        Queries a 31-day window anchored at config.START_DATE so that the
+        universe reflects the index composition at the *beginning* of the
+        study period.  Using today's constituents would introduce look-ahead
+        bias because the CSI 300 continuously replaces weaker stocks with
+        stronger ones, leaking future information into the stock universe.
+        """
+        start_dt = datetime.strptime(config.START_DATE, "%Y%m%d")
+        end_dt = start_dt + timedelta(days=31)
         df = self.pro.index_weight(
             index_code=config.UNIVERSE_INDEX,
-            start_date=start.strftime("%Y%m%d"),
-            end_date=end.strftime("%Y%m%d"),
+            start_date=start_dt.strftime("%Y%m%d"),
+            end_date=end_dt.strftime("%Y%m%d"),
         )
         if df is None or df.empty:
             raise RuntimeError(
-                f"No constituents returned for index '{config.UNIVERSE_INDEX}'. "
+                f"No constituents returned for index '{config.UNIVERSE_INDEX}' "
+                f"around start date {config.START_DATE}. "
                 "Check Tushare permissions (index_weight requires ~2000 points) "
                 "or try '399300.SZ' instead of '000300.SH' in config.py."
             )