Relative Valuation with Machine Learning

Abstract

We use a decision-tree-based machine learning approach to perform relative valuation. Stocks are valued using market-to-book, enterprise-value-to-assets and enterprise-value-to-sales multiples. Our machine learning models reduce median absolute valuation errors by a minimum of 5.6 to 31.4 percentage points relative to traditional models, depending on the multiple used. The identified valuation drivers are consistent with theoretical predictions derived from a discounted cash flow approach. Accounting variables related to profitability, growth, efficiency and financial soundness are important valuation drivers. The valuations produced by machine learning models behave like fundamental values. A value-weighted strategy that buys (sells) undervalued (overvalued) stocks generates highly significant abnormal returns. When we use models trained on listed firms to value IPOs, machine learning models outperform traditional models in valuation accuracy and are better at identifying overpriced IPOs.