Three Machine Learning Solutions to the Bias-Variance Dilemma (Seminar Slides)

Abstract

Classical statistics (e.g., Econometrics) relies on assumptions that are often unrealistic in finance. Two critical assumptions are that the researcher has perfect knowledge about the model’s specification, and that the researcher knows all the variables involved in a phenomenon (including all interaction effects). When those assumptions are incorrect, classical estimators are not guaranteed to be unbiased, or to be the most efficient among the unbiased, leading to poor performance.

In this presentation we explore why machine learning algorithms are generally more appropriate for financial datasets, how they outperform classical estimators, and how they solve the bias-variance dilemma.

Keywords: Bias, variance, MVUE, BLUE, econometrics, machine learning, ensemble, cross-validation, regularization

JEL Classification: G0, G1, G2, G15, G24, E44