What Can Analysts Learn from Artificial Intelligence about Fundamental Analysis?

Abstract

We use machine learning to estimate Nissim and Penman’s (2001) (NP) structural framework that decomposes profitability into four levels of increasingly disaggregated profitability drivers. Our analysis has two distinct features: first, we apply machine learning to accommodate the non-linearities that precluded NP from estimating their framework and second, we analyze the financial statement design choices in NP to provide insights for the teaching and practice of fundamental analysis. We find that out-of-sample profitability forecasts obtained by applying machine learning to NP’s framework are more accurate than those from benchmark models, and that investing strategies based on intrinsic values generated from our profitability forecasts yield risk-adjusted returns. With respect to insights for fundamental analysis, we find that focusing on operating activities, core items and five-year-horizon forecasts improves performance while using a long time series of past information impairs performance. We find mixed evidence of benefits from increasingly granular disaggregation of profitability.

Keywords: Financial Statement Analysis, Machine Learning, Earnings Forecasting

JEL Classification: C53, G10, M41