Connecting the dots: forecasting and explaining short-term market volatility

Abstract

Market volatility prediction is of significant theoretical and practical importance in the financial market, and the news is a significant source to influence the market. By using deep learning networks, we can forecast the volatility based on the news; meanwhile, how to explain the deep neural network is a prevalent topic, especially the attention mechanism in the NLP field. Current studies mainly focus on unveiling the principles behind attention mechanisms without considering generating human-readable explanations. In this work, we attempt to generate a human-readable explanation about the evidence that led to the prediction. To achieve our goal, we propose news-powered neural models to forecast short-term volatility and present a soft-constrained dynamic beam allocation algorithm to control the state-of-the-art language model (GPT-2) to generate fluent and informative explanations.