DeepLOB: Deep Convolutional Neural Networks for Limit Order BooksZihao Zhang 1,2 , Stefan Zohren 1,2 , Stephen Roberts 1,2,31. Machine Learning Research Group, 2. Oxford-Man Institute of Quantitative Finance,Department of Engineering Science, University of Oxford & 3. Mind Foundry Ltd.{zihao, zohren, sjrob}@robots.ox.ac.ukAbstractWe develop a large-scale deep learning model to predict price movements from limit order book(LOB) data of cash equities. The architecture utilises convolutional f i lters to capture the spatialstructure of the limit order books as well as LSTM modules to capture longer time dependencies.The model is trained using electronic market quotes from the London Stock Exchange. Our modeldelivers a remarkably stable out-of-sample prediction accuracy for a variety of instruments and out-performs existing methods such as Support Vector Machines, standard Multilayer Perceptrons, aswell as other previously proposed convolutional neural network (CNN) architectures. The resultsobtained lead to good prof i ts in a simple trading simulation, especially when compared with thebaseline models. Importantly, our model translates well to instruments which were not part of thetraining set, indicating the model’s ability to extract universal features. In order to better understandthese features and to go beyond a “black box” model, we perform a sensitivity analysis to understandthe rationale behind the model predictions and reveal the components of LOBs that are most rele-vant. The ability to extract robust features which translate well to other instruments is an importantproperty of our model which has many other applications.Contents1 Introduction 22 Background and Related Work 33 Data, Normalisation and Labelling 43.1 Limit Order Books . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43.2 Input Data and Normalisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53.3 Labelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 Model Architecture 74.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74.2 Details of Each Component . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 Experimental Results 105.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105.2 Statistical Accuracy of the Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115.3 Performance of the Model in a Simple Trading Simulation . . . . . . . . . . . . . . . . 135.4 Transfer Learning: Applying the Trained Model to New Instruments . . . . . . . . . . . 155.5 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 Conclusion 191arXiv:1808.03668v1 [q-fin.CP] 10 Aug 2018
