feat(data): 添加个股资金流向接口并重构速率限制配置

- 新增 moneyflow 资金流向数据同步模块
- 实现接口级速率限制配置（sync_config.py）
- 更新流动性相关因子定义
- 添加非对称量化损失函数

src/training/components/models/ensemble_quant_loss.py

@@ -84,3 +84,65 @@ class EnsembleQuantLoss(nn.Module):
             total_loss += self.alpha * h_loss + (1.0 - self.alpha) * ic_loss
 
         return total_loss / self.ensemble_size
+
+
+class AsymmetricQuantLoss(nn.Module):
+    """Asymmetric Quant Loss (非对称 Huber + IC)
+
+    保留全截面计算以维持稳定梯度的前提下，对多头关注的错误进行加权惩罚：
+    1. 过度高估烂股票 (买入陷阱) -> 加重惩罚
+    2. 严重低估好股票 (错失金股) -> 加重惩罚
+    """
+
+    def __init__(self, alpha: float = 0.5, ensemble_size: int = 32):
+        super().__init__()
+        self.alpha = alpha
+        self.ensemble_size = ensemble_size
+        # 我们不使用内置的 HuberLoss，而是手动写以支持 element-wise 加权
+        self.delta = 1.0  # Huber threshold (可以设得比较小，比如对于收益率设为 0.05)
+
+    def forward(self, preds: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        batch_size = preds.shape[0]
+        total_loss = 0.0
+
+        if batch_size < 10:
+            # 极小 batch 退回简单 MSE
+            return ((preds - target.unsqueeze(1)) ** 2).mean()
+
+        target_mean = target.mean()
+        target_std = target.std(unbiased=False) + 1e-8
+        target_norm = (target - target_mean) / target_std
+
+        for i in range(self.ensemble_size):
+            pred_i = preds[:, i]
+
+            # --- 1. 非对称 Huber 损失 ---
+            error = pred_i - target
+            abs_error = torch.abs(error)
+
+            # 标准 Huber 计算
+            quadratic = torch.clamp(abs_error, max=self.delta)
+            linear = abs_error - quadratic
+            base_huber = 0.5 * quadratic**2 + self.delta * linear
+
+            # 【核心逻辑】：非对称权重
+            # 如果 target > 0 且 pred < target (错过了涨的好票) -> 权重 1.5
+            # 如果 target < 0 且 pred > 0 (把跌的票预测成了涨的，容易实盘买入踩雷) -> 权重 2.0
+            # 其他情况 (比如烂票预测得更烂) -> 权重 1.0 正常学习
+            weights = torch.ones_like(target)
+            weights[(target > 0) & (error < 0)] = 1.5
+            weights[(target < 0) & (pred_i > 0)] = 2.0
+
+            weighted_huber = (base_huber * weights).mean()
+
+            # --- 2. IC 损失 (维持全截面排序能力) ---
+            pred_mean = pred_i.mean()
+            pred_std = pred_i.std(unbiased=False) + 1e-8
+            pred_norm = (pred_i - pred_mean) / pred_std
+
+            ic = (pred_norm * target_norm).mean()
+            ic_loss = 1.0 - ic
+
+            total_loss += self.alpha * weighted_huber + (1.0 - self.alpha) * ic_loss
+
+        return total_loss / self.ensemble_size

src/training/components/models/tabm_model.py

@@ -19,7 +19,10 @@ from tabm import TabM
 
 from src.training.components.base import BaseModel
 from src.training.components.models.cross_section_sampler import CrossSectionSampler
-from src.training.components.models.ensemble_quant_loss import EnsembleQuantLoss
+from src.training.components.models.ensemble_quant_loss import (
+    EnsembleQuantLoss,
+    AsymmetricQuantLoss,
+)
 from src.training.registry import register_model
 
 
@@ -235,8 +238,8 @@ class TabMModel(BaseModel):
             optimizer, T_max=epochs, eta_min=1e-6
         )
 
-        # 使用 EnsembleQuantLoss 替代 MSE
-        self.criterion = EnsembleQuantLoss(
+        # 使用 AsymmetricQuantLoss (非对称 Huber + IC)
+        self.criterion = AsymmetricQuantLoss(
             alpha=self.params.get("loss_alpha", 0.5), ensemble_size=ensemble_size
         )