Forecasting betas with random forests

ABSTRACT

It is common to estimate equity betas for private firms or non-traded assets through a comparable company analysis, and we test if the Random Forest algorithm can provide superior forecasts. In out-of-sample tests from 1992 to 2018, we find that Random Forest forecasts produce substantially lower average errors and mean absolute errors every year.

KEYWORDS:

• Beta
• comparable company
• random forests

JEL CLASSIFICATION:

• G10
• G12

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

¹ Our results do not change if we allow bigger lags (up to 6 months) between market and accounting information or if we do not lag it at all.

² We estimate the Random Forest with the ranger package in the R software. We tune three hyperparameters: the number of trees in the forest, the maximum number of predictors each tree can test and the depth or level of the tree.