变步长BLSTM集成学习股票预测

Variable Step BLSTM Ensemble Learning for Stock Prediction

提出采用变步长双向长短期记忆网络(BLSTM)集成学习方法学习历史数据中股票价格变动的规律.针对股票涨跌变化的预测改进均方误差(MSE)损失函数,采用简易的模拟交易盈利评价指标以更好地度量预测模型在金融市场中的期望表现.通过前10～50步长的数据训练BLSTM,预测下1min各股票的涨跌变化.实验结果验证了不同数据预处理下,改进损失函数的有效性及变步长集成方法相对于单一网络的有效性.

We present a bi-directional long short-term memory(BLSTM)ensemble learning method with variable step to learn regular pattern of stock price fluctuate from history data.Improved the mean-square error(MSE)loss function for stock fluctuation prediction.This paper use simple simulated trading strategy as evaluating indicator to better evaluate the model performance in financial markets.Use the step between 10 and 50 to train BLSTM to forecast the rise and fall of the stock in next minute.The experimental results verified the effectiveness of the BLSTM ensemble learning method under different data preprocessing and variable step ensemble is more effective than any single network.