Deep learning models for high-frequency financial data

Abstract

The limit order book of a financial instrument represents its supply and demand at each point in time. The limit order book data can be used to predict the future price of the financial instrument. We develop deep learning models to capture the high dimensional data distributions (on R^d) of the limit order data. These models exploit the underlying structure of this complex data. We develop a uniform data grid model for limit order book data to achieve state-of-the-art accuracy for predicting price changes in a stock. We also develop a novel way to use non-uniform events from the limit order book data to train a non-uniform grid data model. This model substantially and consistently outperforms our uniform data grid model. Both the models have been trained and tested over a wide range of periods spanning multiple years for many stocks. The out-of-sample predictions are stable across time for both the models as shown by tests for multiple stocks. Given the huge size of the dataset we use a cluster of CPUs and GPUs to perform our experiments.