PanelGP models stock returns with Gaussian-process factor loadings and forecasts
A November 2 arXiv paper introduces PanelGP, a conditional factor model for large panels of stock returns. Firm characteristics shape factor loadings and economic conditions shape expected factors, both estimated nonparametrically with Gaussian-process priors that can nest linear IPCA and smoother nonlinear maps.
On S&P 500 constituents from 1995 to 2024, nonlinear loadings ranked stocks better than affine loadings and improved Gaussian log scores; portfolios beat IPCA on Sharpe ratios even after trading costs. For economics and mathematics majors, it is a live example of combining panel econometrics, Bayesian nonparametric methods, and portfolio math.
Would you rather deepen theory for factor models, or ship the forecasting code end to end?

0 Comments
Sign in to join the discussion
No comments yet. Be the first to comment!