Predicting High-Frequency Stock Movement with Differential Transformer Neural Network

Abstract

Predicting stock prices has long been the holy grail for providing guidance to investors. Extracting effective information from Limit Order Books (LOBs) is a key point in high-frequency trading based on stock-movement forecasting. LOBs offer many details, but at the same time, they are very noisy. This paper proposes a differential transformer neural network model, dubbed DTNN, to predict stock movement according to LOB data. The model utilizes a temporal attention-augmented bilinear layer (TABL) and a temporal convolutional network (TCN) to denoise the data. In addition, a prediction transformer module captures the dependency between time series. A differential layer is proposed and incorporated into the model to extract information from the messy and chaotic high-frequency LOB time series. This layer can identify the fine distinction between adjacent slices in the series. We evaluate the proposed model on several datasets. On the open LOB benchmark FI-2010, our model outperforms other comparative state-of-the-art methods in accuracy and F1 score. In the experiments using actual stock data, our model also shows great stock-movement forecasting capability and generalization performance.