A Transparent Alternative to Neural Networks with an Application to Predicting Volatility

Vol. 23, No. 3, 2025
By Megan Czasonis, Mark Kritzman and David Turkington

Many prediction tasks in economics and finance involve complicated relationships that lie beyond the reach of linear regression analysis. Neural networks can capture these complex dynamics, but they are notoriously opaque and difficult to implement. We show that an alternative model-free prediction method, called relevance-based prediction, captures many of the same complex dynamics as neural networks, but with transparency into how each observation contributes to each prediction and how each predictive variable contributes to the reliability of each prediction. We describe both prediction methods and compare their respective efficacy for predicting stock market volatility.