freqai同时优化回归和分类参数

freqtrade/templates/freqaiprimer.py

@@ -125,8 +125,8 @@ class FreqaiPrimer(IStrategy):
     }
 
     freqai_info = {
-        "identifier": "test58",
-        "model": "LightGBMRegressor",
+        "identifier": "freqai_primer_mixed",
+        "model": "LightGBMRegressor",  # 默认回归模型
         "feature_parameters": {
             "include_timeframes": ["3m", "15m", "1h"],
             "label_period_candles": 12,
@@ -137,10 +137,27 @@ class FreqaiPrimer(IStrategy):
             "shuffle": False,
         },
         "model_training_parameters": {
-            "n_estimators": 200,
-            "learning_rate": 0.05,
-            "num_leaves": 31,
-            "verbose": -1,
+            "price_value_divergence": {  # 回归模型配置
+                "model": "LightGBMRegressor",
+                "model_params": {
+                    "n_estimators": 200,
+                    "learning_rate": 0.05,
+                    "num_leaves": 31,
+                    "verbose": -1,
+                }
+            },
+            "optimal_first_length": {  # 分类模型配置
+                "model": "LightGBMClassifier",
+                "model_params": {
+                    "n_estimators": 150,
+                    "learning_rate": 0.1,
+                    "num_leaves": 15,
+                    "max_depth": 8,
+                    "min_child_samples": 10,
+                    "class_weight": "balanced",
+                    "verbose": -1,
+                }
+            }
         },
         "fit_live_predictions_candles": 100,
         "live_retrain_candles": 100,
@@ -227,7 +244,29 @@ class FreqaiPrimer(IStrategy):
         dataframe["volume_z_score"] = dataframe["volume_z_score"].replace([np.inf, -np.inf], 0).ffill().fillna(0)
 
         dataframe["&-price_value_divergence"] = dataframe["&-price_value_divergence"].replace([np.inf, -np.inf], 0).ffill().fillna(0)
-
+        
+        # 新增：分类模型优化 first_length
+        # 基于市场状态预测最优的 first_length 分类（2, 4, 6, 8, 10）
+        # 使用技术指标和市场波动率作为特征
+        atr = ta.ATR(dataframe, timeperiod=14)
+        volatility_ratio = atr / dataframe['close']
+        
+        # 定义市场状态特征
+        trend_strength = abs(dataframe['close'] - dataframe['ema200']) / dataframe['ema200']
+        volume_anomaly = abs(dataframe['volume_z_score']) > 2
+        
+        # 基于市场状态确定最优 first_length 类别
+        # 0: 激进(2)  1: 中性(4)  2: 稳健(6)  3: 保守(8)  4: 极保守(10)
+        conditions = [
+            (volatility_ratio > 0.02) & (trend_strength > 0.05),  # 高波动+强趋势 -> 激进
+            (volatility_ratio > 0.015) & (trend_strength > 0.03),  # 中高波动+中强趋势 -> 中性
+            (volatility_ratio > 0.01) | (volume_anomaly),  # 中等波动或成交量异常 -> 稳健
+            (volatility_ratio < 0.008) & (trend_strength < 0.02),  # 低波动+弱趋势 -> 保守
+        ]
+        choices = [0, 1, 2, 3]
+        
+        dataframe["&*-optimal_first_length"] = np.select(conditions, choices, default=4)
+        
         return dataframe
     def is_stochrsi_overbought(self, dataframe: DataFrame, period=10, threshold=85) -> bool:
         """
@@ -1292,21 +1331,28 @@ class FreqaiPrimer(IStrategy):
 
     def detect_trend_status(self, dataframe: DataFrame, metadata: dict) -> str:
         """
-        基于加权分段的trend_score判断趋势状态
+        基于分类模型优化 first_length 的趋势检测
         规则：
-        - 将最近N个trend_score分为3段：first, second, third
-        - 长度比例为：first:second:third = 1:3:5
-        - 例如：2,6,10（固定使用段长2）
-        - 权重：first(权重10)，second(权重7)，third(权重3)
-        - 使用静态段长：first段长度固定为2
+        - 使用 LightGBMClassifier 预测最优 first_length 类别
+        - 类别映射：0:激进(2), 1:中性(4), 2:稳健(6), 3:保守(8), 4:极保守(10)
+        - 根据预测的类别动态调整段长，而非固定使用2
         """
         pair = metadata.get('pair', 'Unknown')
         
         try:
-            # 直接使用静态值2作为第一段长度
-            first_length = 2
+            # 使用分类模型预测最优 first_length
+            if "&*-optimal_first_length" in dataframe.columns:
+                # 获取最新预测值
+                optimal_length_class = int(dataframe["&*-optimal_first_length"].iloc[-1])
+                length_mapping = {0: 2, 1: 4, 2: 6, 3: 8, 4: 10}
+                first_length = length_mapping.get(optimal_length_class, 2)
+                logger.info(f"[{pair}] 使用分类模型优化 first_length: {first_length} (类别: {optimal_length_class})")
+            else:
+                # 回退到默认值
+                first_length = 2
+                logger.warning(f"[{pair}] 分类模型未预测，使用默认 first_length: {first_length}")
             
-            # 根据比例计算第二段和第三段长度
+            # 根据 first_length 动态调整其他段长
             second_length = first_length * 3
             third_length = first_length * 5