Optimal Peers

This paper provides a theoretical foundation for constructing optimal benchmarks via machine learning (ML). For a broad class of models, the optimal benchmark is given by an appropriately weighted portfolio of peers. While Ordinary Least Squares (OLS) provides the theoretically optimal weights in the population, ML methods, notably the lasso, can 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. This suggests that ML-based benchmarks can substantially improve outcomes in a wide range of applications, such as incentive contracts or relative performance evaluation.