Optimal Peers

Abstract

This paper introduces a regression-based methodology for constructing optimal benchmarks, employing machine learning (ML) techniques to handle high-dimensional sets of regressors. The regression-based approach is implied by theory: For a broad class of models, the OLS coefficients from a regression of a focal firm's outcome on the outcomes of all other firms in the economy yield the theoretically optimal portfolio weights for benchmark construction. In practice, machine learning methods, notably the lasso, provide a robust, implementable solution. In an application to a large sample of U.S. public firms, ML-based benchmarks strongly outperform traditional industry benchmarks in out-of-sample explanatory power. These findings suggest that ML-based benchmarks can significantly improve incentive design and relative performance evaluation.

Keywords: Peer groups, benchmarking, machine learning, lasso, relative performance evaluation, industry classification. JEL Classification: D2, D4, D8, G1, G3, L1, L2

JEL Classification: D2, D4, D8, G1, G3, L1, L2