Nowcasting Mexican Manufacturing Activity with Traditional and Non-Traditional Indicators: A High-Dimensional Regularization Approach

This paper develops a high-dimensional nowcasting framework for Mexico’s Monthly Industrial Production Index (IMAI) using 302 monthly indicators combining traditional and non-traditional high-frequency information. Under a pseudo-real-time design, we compare penalized regressions with ARMA errors (Ridge, LASSO, Elastic Net, Adaptive LASSO, Adaptive Elastic Net, and Multi-Step Adaptive Elastic Net) and DFM-based alternatives. All models outperform a na¨ıve AR(1) benchmark. The best result is achieved by the Multi-Step Adaptive Elastic Net (MSAENET) with a relative RMSFE of 0.273. Non-traditional indicators improve performance mainly under adaptive penalties, with Diebold–Mariano evidence favoring the richer information set. DFM-based specifications improve with non-traditional information but remain less accurate than the best penalized regressions. Forecast gains stem from disciplined variable selection. MSAENET selects 10.4 predictors on average, with the Google Trends topic Drought as the only non-traditional predictor consistently retained, reflecting Mexico’s industrial vulnerability to climatic conditions. A Ridge model estimated on the seven most consistently selected predictors achieves a relative RMSFE of 0.266—statistically indistinguishable from the best full-sample specifications—using only publicly available series, offering a robust and operationally feasible nowcasting implementation