feat(training): 新增 TabM 排序学习模型支持并优化训练流程

- 新增 TabMRankModel、TabMRankTask 及配套损失函数与配置
- 将 DataQualityAnalyzer 从 experiment 迁移至 training 模块
- 调整数据处理器移除过度的 NaN/null 硬填充逻辑
- 优化 RankTask 评估指标使用分位数标签替代原始收益率
- 更新实验脚本处理器顺序与模型超参数配置

src/experiment/learn_to_rank.py

@@ -54,27 +54,10 @@ N_QUANTILES = 20
 
 # 排除的因子列表
 EXCLUDED_FACTORS = [
-    'active_market_cap',
-    'close_vwap_deviation',
-    'sharpe_ratio_20',
-    'upper_shadow_ratio',
-    'volume_ratio_5_20',
-    'GTJA_alpha090',
-    'GTJA_alpha084',
-    'GTJA_alpha066',
-    'GTJA_alpha150',
-    'GTJA_alpha148',
-    'GTJA_alpha106',
-    'GTJA_alpha109',
-    'GTJA_alpha108',
-    'GTJA_alpha176',
-    'GTJA_alpha169',
-    'GTJA_alpha156',
-    'chip_dispersion_70',
-    'winner_rate_cs_rank',
-    'atr_price_impact',
-    'low_vol_days_20',
-    'liquidity_shock_momentum',
+    # 'debt_to_equity',
+    # 'GTJA_alpha016',
+    # 'GTJA_alpha141',
+
 ]
 
 # LambdaRank 模型参数配置
@@ -145,8 +128,8 @@ def main():
     pipeline = DataPipeline(
         factor_manager=factor_manager,
         processor_configs=[
-            (NullFiller, {"strategy": "mean"}),
             (Winsorizer, {"lower": 0.01, "upper": 0.99}),
+            (NullFiller, {"strategy": "mean"}),
             (CrossSectionalStandardScaler, {}),
         ],
         filters=[STFilter(data_router=engine.router)],