What Can Analysts Learn from Artificial Intelligence about Fundamental Analysis

Taking the perspective of an equity investor seeking to maximize risk-adjusted returns through financial statement analysis, we apply a machine learning algorithm to estimate Nissim and Penman’s (2001) structural decomposition framework of profitability. Our approach explicitly takes account of the nonlinearities that precluded Nissim and Penman from estimating their framework. We first forecast profitability and then estimate intrinsic values using different subsets of Nissim and Penman’s framework and different fundamental analysis design choices; we find that trading on these estimates generates substantial risk-adjusted returns. Choices that improve performance include increasingly granular ratio disaggregation and long-horizon forecasts of operating performance. Perhaps surprisingly, we find only weak evidence of benefits from a fundamental analysis that incorporates historical financial statement information beyond the current-period information or focuses only on core items. While taking account of non-linearities improves model performance for all firms, the effect is strongest for small, loss-making, technology, and financially distressed firms.