In our work, we design a deep neural network that incorporates convolutional layers, Inception Modules and Long Short-Term Memory (LSTM) units to predict future stock price movements in large- scale high-frequency LOB data. The usage of convolutional layers and Inception modules avoids the limitations of handcrafted features and a LSTM layer is used to capture time dependencies among the resulting feature maps. One advantage of our model over previous research is that it has the ability to adapt for many stocks by extracting representative features from highly noisy data. We compared our model, DeepLOB, with a large group of state-of-the-art algorithms including Linear Discriminant Analysis (LDA), Multilinear Time-Series Regression (MTR), Bag-of-Feature (BOF) and Autoencoder (AE), and our model outperforms all existing methods.

We have limit order book data for ten stocks over the entire 2017 year and our model delivers robust out-of-sample prediction accuracy over a test period of three months. In addition, our model generalizes well to stocks not even in the training data (transfer learning), indicating the existence of universal features in the limit order book that modulates stock demand and price. To show the practicability of our model, we designed a simple trading simulation and Figure 1 indicates the normalized cumulative profits of our model, under the assumption of mid-price simulation and no transaction costs, over three months for different prediction horizons (k).

In the subsequent works, we extend our model with Bayesian deep neural networks and quantile regression methods to obtain uncertainty information on predicted outputs. We demonstrate that uncertainty information derived from posterior predictive distributions can be utilised for position sizing, avoiding unnecessary trades and improving profits.

Figure 1: Normalized cumulative profits for test periods for different stocks and prediction horizons (k). Profits are in GBX (=GBP/100).