Stock Market Embedding and Prediction: A Deep Learning Method

摘要

It has always been a challenging issue for people to understand the stock market and make reasonable predictions. For a very long period, investors are trying to manually extract useful features from the financial market which generates an enormous volume of data every single day. People create a lot of technical indicators like MACD, TR, MFI to describe momentum, volume and volatility signals of the financial time series. However, the limitation is evident due to the efficiency of manually feature engineering. With the rapidly growing volume of data, deep neural network shows excellent performance in many research areas like natural language processing, voice recognition, image identification. It provides a new view to dig out potentially useful information automatically. In this paper, we present a novel end-to-end training using an embedding method to automatically extract features and get a summary representation of the daily market. Moreover, we apply the Long Short-Term Memory (LSTM) with attention mechanism to predict daily return ratio of HS300 index. The features extracted by the embedding layer show greater predictive power than manually defined technical signals by 92.42% lower MSE. Moreover, the use of attention mechanism also provides an average enhance of 55.68% in MSE. Our study shows that deep neuron network structure has a strong potential for better understanding market complex behaviors.