This month’s seminar was given by Dr Nicholas Westray, Courant Institute of Mathematical Sciences, New York University with a talk on **extracting alpha from the limit order book** using deep learning.

Firstly, Dr Westray explained the many challenges that firms face to generate alpha signals from limit order books such as: the enormous amounts of data generated; the need for specialist infrastructure to store, process and analyse the data; the data is noisy, non-stationary and fat-tailed and also the field is extremely competitive. The current approach would be for Quants to extract features using expert domain knowledge.

Dr Westray then described how neural networks have transformed problems that had previously used hand-crafted approaches and gave a summary of neural network architectures: multi-layer perceptron, recurrent neural networks, long-short term memory models and convolutional neural networks.

Dr Westray gave an outline the problem, which is to predict returns from limit order books data, where returns are stationary, formulated in terms of prices and volumes and the prices are non-stationary. The bid and ask prices were transformed, and from these order flow and order flow imbalance can be then determined which are better suited to determine returns.

Several models were compared: an autoregressive model, multilayer perceptron, a standard LSTM, LSTM feeding a multi-layer perceptron, a multi-layer LSTM and a CNN feeding into a LSTM and tested on 13 months of NASDAQ stock data. The models were fit to each stock, implemented in Python using TensorFlow and Keras, and run on GPUs.

Dr Westray concluded from his results that stationarity of the inputs was critical to getting good outcomes, and since the model results had strong microstructural dependencies that universality also needs to be taken into consideration. He also noted that the simple models gave similar performance as the complicated ones.

Dr Westray concluded his talk with opportunities to extend the approach such as considering Bayesian Deep Nets, volume prediction and considering other venues/asset classes such as foreign exchange or futures.