Enhancing data envelopment analysis: a comprehensive calibration with random forests

Abstract This paper proposes an algorithm integrating random forests (RF) and data envelopment analysis (DEA), termed RF-DEA, to address several methodological challenges commonly encountered in DEA applications, e.g., sensitivity to outliers and the ‘curse of dimensionality’. We show that by leveraging the strengths of RF, the DEA estimator with a calibrated frontier and input/output vectors can effectively mitigate the influence of outliers, identify redundant variables, and account for the influence of environmental variables. Evidence from Monte Carlo simulations demonstrates that these properties of RF-DEA yield substantial improvements in estimation accuracy compared to the conventional DEA. Through an empirical application to global production efficiency, we illustrate that this easy-to-implement method constitutes a viable alternative DEA estimator, capable of disentangling some practical challenges in efficiency analysis.