Measuring Firm Quality Using Machine Learning

Firm quality is a foundational construct in the fundamental analysis literature. Asness et al. (2019), a recent representative example of this literature, measures firm quality based on 19 fundamental ratios guided by valuation theory (referred to as Asness’ Q score). We examine whether it is possible to leverage the power of machine learning to construct a better measure of firm quality using the same 19 fundamental signals. We show that an advanced machine learning model called XGBoost based on the 19 ratios can outperform a linear OLS regression model based on Asness’ Q score (our benchmark) by 27%. However, we fail to find economically significant evidence that adding more raw accounting data items identified by the prior literature or financial statements or using alternative ratios derived from the DuPont decomposition can yield stronger prediction models. We show that our measure of firm quality based on XGBoost and the 19 ratios can better explain contemporaneous stock prices than Asness’ Q score. In addition, a value investing trading strategy using our XGBoost model outperforms the same trading strategy based on Asness’ Q score by an economically significant margin