Relevance-Based Importance in Prediction Analysis

This document is a research article that introduces a new measure of variable importance called relevance-based importance (RBI). The authors, Megan Czasonis, Mark Kritzman, and David Turkington, explain that traditional measures like the t-statistic and Shapley value have limitations when interpreting variable importance in predictive modeling. The t-statistic can be challenging to interpret under conditions of collinearity, while the Shapley value does not account for individual predictions' reliability. The article outlines how RBI overcomes these limitations by providing a robust measure that considers both average reliability across predictions and individual contributions to reliability. The authors detail the mathematical foundations of RBI and demonstrate its advantages through simulations. They also discuss the relevance-based prediction method, which forms the basis for RBI, and present an empirical analysis of RBI applied to stock market volatility predictions. The article concludes with a summary of the findings and implications for predictive modeling.