Can Machine Be a Good Stock Picker?: Bridging the Gap between Fundamental Data and Machine Learning

Abstract: We investigate the efficacy of historical accounting data and consensus forecasts for relative valuation of stocks, employing tree-based machine learning methods. We run an XGBoost model for monthly cross-sections of financial and pricing data of US equities from 1984 to 2021. We find that our model is effective for predicting pricing multiples based on non-linear relationships among various financial ratios calculated from historical financial reports, and consensus forecasts contribute to improving prediction errors of valuation. Although some predictors based on analyst forecasts score high in variable importance ranking based on SHAP, overall, they do not become consistently more important than variables based on accounting reports, when analyst forecast data is added to the models. Furthermore, when we use valuation errors as a trading signal for convergence trade, the performance is the best for the trading signals based only on historical accounting data. The convergence trade is successful for small-cap firms, earning sizable abnormal returns with limited portfolio turnover, drawdown and exposure to the Fama French 6 factors. It is suggested that the machine learning method could help to detect cheap and expensive companies within the small-cap universe while avoiding distressed firms.