What Can Analysts Learn from Artificial Intelligence about Fundamental Analysis?

Abstract

The accounting-based valuation model developed by Nissim and Penman (2001) presents both increasingly disaggregated equity-value drivers (financial statement ratios) in a structured hierarchy and descriptive evidence on the over-time and cross-sectional properties of these ratios. Nissim and Penman do not use their model to forecast residual income, either as an end-objective or as an input to valuation, primarily because of estimation difficulties arising from the highly non-linear structure of their framework. We revisit their theory, taking the perspective of an equity investor seeking to maximize risk-adjusted returns through fundamental analysis. We address the estimation difficulty described by Nissim and Penman using a machine learning algorithm, Deep Learning, designed to approximate arbitrarily-complex higher-order interactive relationships. We derive firm-level profitability forecasts using their structured hierarchy, use these forecasts to estimate intrinsic values and compare the estimated values to price. We find that risk-adjusted returns to a trading strategy based on buying (selling) firms with high (low) value-to-price ratios are substantial, and larger when the underlying valuation model is based on greater disaggregation and long-horizon forecasts of operating activities. We find only weak evidence that incorporating historical information beyond the information in the current period financial statements or focusing on core items is beneficial, and that taking account of non-linearities is especially important in valuing small, loss-making, technology and financially distressed firms.