Extended Abstract: Neural Networks for Limit Order Books

Abstract

We design and test neural networks for modeling the dynamics of the limit order book. In addition to testing traditional neural networks originally designed for classification, we develop a new neural network architecture for modeling spatial distributions (i.e., distributions on $\mathbb{R}^d$) which takes advantage of local spatial structure. Model performance is tested on 140 S\&P 500 and NASDAQ-100 stocks. The neural networks are trained using information from deep into the limit order book (i.e., many levels beyond the best bid and best ask). Techniques from deep learning such as dropout are employed to improve performance. Due to the computational challenges associated with the large amount of data, the neural networks are trained using GPU parallel computing. The neural networks are shown to outperform simpler models such as the naive empirical model and logistic regression, and the new neural network for spatial distributions outperforms the standard neural network.