From fundamental signals to stock volatility: A machine learning approach

Abstract

Enriched with a large set of accounting-based characteristics, we find that the aggregate fundamental risk, constructed with several machine learning algorithms, predicts stock return volatility. We find that nonlinear models, especially neural networks, outperform linear methods and single characteristics and attribute the improvements in prediction accuracy to their ability to capture nonlinear patterns. All approaches concur that profitability-related characteristics are the dominant predictive indicators. In addition, volatility-managed market portfolios through machine learning improve economic profits. Our study contributes to the body of knowledge on risk management in emerging markets in the age of big data.