Predicting Analyst Performance: A Machine Learning Fund-of-Analysts Framework

Inspired by Alpha in Analysts (2025), this paper investigates whether the cross-sectional and time-series heterogeneity in sell-side analyst skill can be exploited using modern machine learning techniques. We treat each analyst as a portfolio manager and construct analyst-level portfolios based on their 12-month target price recommendations, translating implied returns into systematic long–short positions. The objective is to predict one-month-ahead 12-months cumulative returns of these analyst portfolios using a rich set of analyst and portfolio-level characteristics. We extend the original framework along three key dimensions. First, we expand the information set from six to twenty-four predictive features capturing analyst behavior, portfolio composition, and historical performance. Second, we move from rolling ordinary least squares to an expanding walkforward cross-validation framework, allowing for realistic real-time prediction and model selection. Third, we compare linear and non-linear models by evaluating 8 forecasting approaches: OLS benchmarks with Ridge, Lasso, and Elastic Net regularization, as well as Random Forests, XGBoost, LightGBM, and multilayer perceptrons. Our empirical analysis covers 1,474 unique analysts and 105,866 analyst-month observations. We find that, although machine learning provides only modest improvements in point forecasting accuracy, it delivers large gains in cross-sectional ranking ability, a key input for analyst- selection and long–short investment strategies. However, for now, these gains difficulty translate into economically meaningful improvements in a dynamic “fund-of-analysts” allocation strategy that weights analysts based on predicted performance.