Using Machine Learning to Identify Mispricing in European Stock Markets

Boosting Agnostic Fundamental Analysis: Using Machine Learning to Identify Mispricing in European Stock Markets

AuthorsMatthias X. Hanauer; Marina Kononova; Marc Steffen Rapp

Paper linkhttps://papers.ssrn.com/sol3/papers.cfm?abstract_id=3977872

This strategy invests in EU17 stocks (EU15 plus Switzerland and Norway), excluding financial firms, non-common equities, secondary listings, and companies with missing or invalid data. Stocks with market capitalization below USD 10 million are also excluded. Using accounting variables from the previous 48 months—such as total assets, sales, and long-term debt (as defined in Table A)—each variable is standardized by cross-sectional ranking into a [-1, 1] range. Market capitalization is deflated by the total market value to control for shifts in market-wide valuation norms. Random forest and gradient boosting models are trained to estimate each firm’s fair value, and the final estimate is the average of both. The mispricing signal is defined as the difference between the estimated fair value and market capitalization, scaled by market capitalization. Each month, firms are sorted into quintiles based on this signal. The portfolio goes long the most undervalued quintile and short the most overvalued, with monthly rebalancing and value-weighted positions.

The strategy’s profitability stems from two core drivers: mispricing detection and model design. By leveraging fundamental accounting data, it systematically identifies undervalued and overvalued firms, capitalizing on valuation inefficiencies. Moreover, advanced machine learning methods—random forest and gradient boosting—capture complex, nonlinear relationships that traditional linear models fail to detect. The ensemble approach enhances prediction accuracy, making both the conceptual idea of exploiting mispricing and the technological sophistication of the model equally crucial to its success.