The Empirical Validation of Machine Learning

In this paper, a complementary method to classical statistical modeling such as ordinary least squares, and instrumental variables estimation, is introduced. This method, double/debiased machine learning, is applied in the presence of a high-dimensional nuisance parameter set, which possibly interferes with the estimator for the treatment effect. The following techniques are introduced and thoroughly examined: lasso, random forests, neural nets, boosted trees, and an ensemble method. Thereafter, estimates obtained by these techniques are compared to those obtained by traditional OLS and IV estimation. Specifically, research results obtained from Nunn and Wantchekon (2011) and Acemoglu et al. (2005) are used to empirically examine this nonparametric estimation. The hypothesis that DML performs better out-of-sample estimations is supported by Nunn and Wantchekon (2011), which therefore questions the reliability of classical statistical modeling. This hypothesis can not be confirmed nor refuted by Acemoglu et al. (2005).