Price change prediction of ultra high frequency financial data based on temporal convolutional network

TL;DR

This paper applies a temporal convolutional network with attention mechanisms to predict discrete categories of ultra high frequency stock price changes, outperforming traditional models on a large Chinese stock dataset.

Contribution

It introduces a novel TCN-based framework with attention for modeling large-scale UHF stock price change data, improving prediction accuracy over existing methods.

Findings

TCN with attention outperforms GARCH and LSTM models

The dataset size is nearly 10 million data points

First application of TCN to such large-scale UHF stock data in China

Abstract

Through in-depth analysis of ultra high frequency (UHF) stock price change data, more reasonable discrete dynamic distribution models are constructed in this paper. Firstly, we classify the price changes into several categories. Then, temporal convolutional network (TCN) is utilized to predict the conditional probability for each category. Furthermore, attention mechanism is added into the TCN architecture to model the time-varying distribution for stock price change data. Empirical research on constituent stocks of Chinese Shenzhen Stock Exchange 100 Index (SZSE 100) found that the TCN framework model and the TCN (attention) framework have a better overall performance than GARCH family models and the long short-term memory (LSTM) framework model for the description of the dynamic process of the UHF stock price change sequence. In addition, the scale of the dataset reached nearly 10 million, to the best of our knowledge, there has been no previous attempt to apply TCN to such a large-scale UHF transaction price dataset in Chinese stock market.