Movement forecasting of financial time series based on adaptive LSTM-BN network

Long-short term memory (LSTM) network is one of the state-of-the-art models to forecast the movement of financial time series (FTS). However, existing LSTM networks do not perform well in the long-term forecasting FTS with sharp change points, which significantly influences the accumulated returns. This paper proposes a novel long-term forecasting method of FTS movement based on a modified adaptive LSTM model. The adaptive network mainly consists of two LSTM layers followed by a pair of batch normalization (BN) layers, a dropout layer and a binary classifier. In order to capture the important profit points, we propose to use an adaptive cross-entropy loss function that enhances the prediction capacity on the sharp changes and deemphasizes the slight oscillations. Then, we perform the forecasting on multiple independent networks and vote on their output data to obtain stable forecasting result. Considering the temporal correlation of FTS, an inherited training strategy is introduced to accelerate the retraining procedure when performing the long-term forecasting task. The proposed methods are assessed and verified by the numerical experiments on the stock index datasets, including "Standard's & Poor's 500 Index", "China Securities Index 300" and "Shanghai Stock Exchange 180". A substantial improvement of forecasting performance is proved. Moreover, the proposed hybrid forecasting framework can be generalized to different FTS datasets and deep learning models.