Two-stage attentional temporal convolution and LSTM model for financial data forecasting

Abstract

Financial time series usually consist of multiple time series, and financial time series data forecasting models use the historical data plays of multiple driving series to predict the future values of the target series. In recent years, attention-based Long and Short-Term Memory (LSTM) neural networks and Temporal Convolutional Networks (TCN) have been widely used in time series forecasting. In this paper, we propose a two-stage attention-based TCN and LSTM hybrid forecasting model, in order to better obtain the spatial correlation of driving sequences, we used causal self-attention to obtain the spatial attention weights of driving sequences, then use TCN to extract the short-term features of the series in the first stage, in the second stage, adding the temporal attention module computes the sequence adaptively assigning weights to the input sequence for the current and historical moments, and finally use LSTM to capture the long-term dependence of the time-series data. We used the NASDAQ 100 stock dataset and the financial time series of CSI 300 companies to measure the performance of the proposed model in financial data forecasting.