Multi-Step Stock Closing Price Forecasting Using CNN-LSTM with Time Location

This paper proposed a novel and effective deep model based on convolutional neural network (CNN) and long-short-term memory (LSTM) with time location for multi-step closing price forecasting. Unlike the traditional CNN-LSTM, the proposed method did not adopt one cascade structure; instead, it used one parallel CNN and LSTM structure to extract rich hidden patterns for closing price forecasting. In which CNN is used to extract their global temporal hidden features; LSTM is used to extract the features that have long-term dependencies. CNN and LSTM extracted features are concatenated as final fusion features for forecasting. The experimental results proved its effectiveness on three real data sets.