Spatio-temporal Model for Limit Order Books in The Stock Exchange of Thailand

Abstract

We propose a spatio-temporal model for predicting anomaly price movements in the Stock Exchange of Thailand (SET). The model used a deep neural network classification algorithm that has a time series of of limit order books (LOB) as an input. There were three output classes: anomaly price uptrend, anomaly price downtrend, and normal price movements. We performed experiments to compare the efficiency among convolutional neural network model, Long short-term memory model, and the combination of both in order to classify anomaly price movements. The results of the experiment showed that the combination of both convolutional layers and Long short-term memory model had the highest accuracy with 74.55% for predicting abnormal price movements.