基于LSTM的多维特征融合量化交易策略实现
功能说明与风险警示
本策略通过长短期记忆网络（LSTM）模型整合价格序列、成交量动态及技术指标特征，构建时序预测模型用于金融市场方向判断。核心功能包含：1) 多源数据标准化处理；2) 技术指标衍生计算；3) 时序特征工程；4) LSTM网络参数优化。该策略存在过拟合风险、滞后效应及黑箱模型可解释性不足等问题，实际部署需配合严格的风险管理机制。
数据预处理与特征工程
基础数据规范化
importnumpyasnpimportpandasaspdfromsklearn.preprocessingimportMinMaxScalerdefpreprocess_data(df,feature_cols):"""实现多维度数据的归一化处理"""scalers={col:MinMaxScaler(feature_range=(0,1))forcolinfeature_cols}processed=df.copy()forcolinfeature_cols:processed[col]=scalers[col].fit_transform(processed[col].values.reshape(-1,1)).flatten()returnprocessed,scalers技术指标矩阵构建
defcalculate_technical_indicators(df):"""生成包含趋势和波动率特征的技术指标集"""# 移动平均线系统df['MA5']=df['close'].rolling(window=5).mean()df['MA20']=df['close'].rolling(window=20).mean()# 动量指标df['RSI']=compute_rsi(df['close'])df['MACD'],df['MACD_signal']=compute_macd(df['close'])# 波动率度量df['ATR']=compute_atr(df['high'],df['low'],df['close'])# 成交量加权指标df['VWAP']=(df['volume']*df['close']).cumsum()/df['volume'].cumsum()returndf.dropna()LSTM模型架构设计
三维输入张量构造
defcreate_sequences(data,target_col,sequence_length):"""将时间序列转换为监督学习格式"""X,y=[],[]foriinrange(len(data)-sequence_length):# 提取窗口内的价格、成交量和技术指标seq=data.iloc[i:(i+sequence_length)]X.append(seq.values)# 下一时刻的涨跌标记y.append(1ifdata.iloc[i+sequence_length][target_col]>data.iloc[i+sequence_length-1][target_col]else0)returnnp.array(X),np.array(y)混合注意力机制层
fromtensorflow.keras.layersimportAttention,LSTM,Dense,Dropoutdefbuild_lstm_model(input_shape,num_features):"""构建带注意力机制的LSTM网络"""inputs=tf.keras.Input(shape=input_shape)# 第一层LSTM捕获时序依赖x=LSTM(64,return_sequences=True)(inputs)x=Dropout(0.3)(x)# 第二层LSTM提取高层特征x=LSTM(32,return_sequences=True)(x)x=Attention()([x,x])# 自注意力机制# 全连接层进行特征融合x=Dense(16,activation='relu')(x)outputs=Dense(1,activation='sigmoid')(x)returntf.keras.Model(inputs=inputs,outputs=outputs)训练流程与验证体系
时空交叉验证方案
deftemporal_cv_split(data,n_splits=5):"""按时间顺序划分训练测试集"""split_points=np.linspace(0,len(data),n_splits+1,dtype=int)[1:-1]folds=[]foriinrange(n_splits):train_end=split_points[i]test_end=split_points[i+1]ifi<n_splits-1elselen(data)folds.append((data.iloc[:train_end],data.iloc[train_end:test_end]))returnfolds类别平衡处理
class_weights=compute_class_weight('balanced',classes=np.unique(y_train),y=y_train)history=model.fit(X_train,y_train,validation_data=(X_val,y_val),class_weight=dict(enumerate(class_weights)),batch_size=32,epochs=100,callbacks=[EarlyStopping(patience=10,restore_best_weights=True)])实盘部署关键要素
在线特征更新管道
classRealtimeFeatureUpdater:"""实时维护特征矩阵的增量更新组件"""def__init__(self,window_size=60):self.window_size=window_size self.feature_buffer=deque(maxlen=window_size)defupdate_features(self,new_tick):"""接收新行情数据并更新特征缓冲区"""current_features=self._compute_current_features(new_tick)self.feature_buffer.append(current_features)iflen(self.feature_buffer)==self.window_size:returnnp.array(self.feature_buffer)returnNonedef_compute_current_features(self,tick):"""实时计算各项技术指标"""# 实现包括：指数平滑移动平均线(EWMA)# 真实波幅(ATR)递推计算等pass置信度校准模块
defcalibrate_confidence(predictions,actuals,method='isotonic'):"""使用等渗回归校准预测置信度"""fromsklearn.isotonicimportIsotonicRegression ir=IsotonicRegression(increasing=True)calibrated_probs=ir.fit_transform(predictions,actuals)# 绘制可靠性曲线plt.figure(figsize=(8,6))binned_truth,bin_edges,_=histogram(actuals,predictions,bins=10)plot_reliability_diagram(binned_truth,bin_edges)returncalibrated_probs策略有效性验证
样本外测试框架
defbacktest_strategy(model,test_data,initial_capital=100000):"""模拟交易执行过程评估策略表现"""portfolio_value=[initial_capital]position=0foriinrange(len(test_data)-SEQUENCE_LENGTH):# 获取当前预测信号pred_signal=model.predict(test_data[i:i+SEQUENCE_LENGTH])[0][0]# 根据信号调整仓位ifpred_signal>0.5andposition==0:position=portfolio_value[-1]/test_data['close'].iloc[i+SEQUENCE_LENGTH]portfolio_value.append(portfolio_value[-1]-test_data['close'].iloc[i+SEQUENCE_LENGTH]*position)elifpred_signal<=0.5andposition>0:portfolio_value.append(portfolio_value[-1]+test_data['close'].iloc[i+SEQUENCE_LENGTH]*position)position=0else:portfolio_value.append(portfolio_value[-1])# 计算绩效指标returns=np.diff(portfolio_value)/portfolio_value[:-1]sharpe_ratio=np.mean(returns)/(np.std(returns)+1e-8)*np.sqrt(252)max_drawdown=compute_max_drawdown(portfolio_value)return{'final_value':portfolio_value[-1],'sharpe_ratio':sharpe_ratio,'max_drawdown':max_drawdown,'win_rate':np.mean(returns>0)}