29/10/2026 - 12:45 - 14:00 - DOYEN 22
(Université du Québec)
will give a presentation on
Abstract:
For interpretability, sparsity is key and linearity is preferable. Yet neither is common in machine-learning-based macroeconomic forecasting, where large predictor sets are fed into nonlinear, nonparametric models. We show that any supervised learning algorithm whose predictions are linear in the target variable—regardless of nonlinearities in the features—admits a linear-factor representation. We call the resulting components the model’s implied factors. Studying these factors provides a new route to uncover low-dimensional structure, even when the model was not built with such structure in mind. Empirically, two or three implied factors suffice to reproduce out-of-sample forecasts of U.S. GDP, inflation, and unemployment. Across models, the leading direction is common and these factors cluster into a few common signals, while higher-order factors differ by family, so cross-model differences are quantitative rather than conceptual. Used as conditioning controls in local projections, implied factors from a random forest deliver correctly signed inflation responses where standard PCA controls leave the price puzzle intact.
More information about the speaker