Machine Valuation

Abstract

We present a machine learning approach to firm valuation that requires only historical accounting data as input. The machine learning model generates a median absolute percentage error of 17.2% in out-of-sample firm value predictions. The model out-performs a sample of final-year finance students (41.3%) and individual analyst forecasts of one-year-ahead firm value (22.4%). We show that subsequent market valuations move towards the model valuation, generating return predictability over horizons of up to five years.

Keywords: valuation; asset pricing; return predictability; machine learning; gradient boosted trees

JEL Classification: G12; G14; C38