Intraday and interday features in the high-frequency data: Pre- and post-Crisis evidence in China’s stock market

Abstract

Extreme events—statistically improbable events with profound consequences—happen more often than expected in stock markets and lead to an urgent need to investigate market behavior, in particular in the context of stock market crashes. We propose a methodology that takes into account both intraday (or within-day) patterns and interday (or day-to-day) dynamics of the stock market, with an exclusive focus on the data around the 2015 stock market crisis in China. In the first stage, a quantitative analysis of our methodology accounts well for the differences in intraday and interday dynamics between the pre-Crisis and post-Crisis periods; in the second stage, with the interday features as input to machine learning methods, we predict directional movements of the daily closing price in the Shanghai stock exchange index and develop trading strategies. Empirical results evince the predictability of the market and the informativity of the interday features when they are included in the machine learning methods. Decision-tree-based ensemble methods produce relatively better classification and trading performance than the other methods. The gradient-boosting decision tree deals best with the information derived from the input features, in which the average price is the most important feature.