Exploiting intra-day patterns for market shock prediction: A machine learning approach

Discovering hidden patterns under unexpected market shocks is a significant and challenging problem, which continually attracts attention from research communities of mathematics, economics, and data science. Classic financial pricing models present unsatisfactory prediction accuracy when applied to real-world data due to limited capacity in depicting complex market movements. In this paper, we develop a machine learning approach, called ARMA-GARCH-NN, to capture intra-day patterns for stock market shock forecasting. Specifically, we integrate classical financial pricing models with artificial neural networks, with explicitly designed feature selection and cross-validation methods. We conduct empirical studies on high-frequency data of the U.S. stock market for evaluation. Our results provide initial evidence of the predictability of market shocks. Additionally, we confirm the effectiveness of ARMA-GARCH-NN by recognizing patterns in massive stock data without strong assumptions on distribution. Our method can serve as a portable methodology that integrates the advantages of traditional financial models and data-driven methods to reveal hidden patterns in large-scale financial data.