Nowcasting with Novel High-Frequency Data: A Cross-Method Comparison for Colombia’s ISE

This research addresses two key gaps in the nowcasting literature: the lack of systematic comparisons across heterogeneous methodologies and the limited use of novel, high‑frequency datasets. Focusing on the monthly economic activity indicator ISE for Colombia, published by DANE, we evaluate a broad suite of models to close the first gap. These include the traditional Dynamic Factor Model (DFM), regularized regressions (Elastic Net and LASSO), tree‑based methods (XGBoost and Random Forest), a simple neural‑network specification, and a point‑forecast combination approach. The results show that regularized regressions consistently outperform all other models in terms of predictive accuracy. To address the second gap, we construct three Google Search indexes and develop a set of economic‑activity indicators derived from Redeban’s transactional data—the largest payment processor in Colombia. Although these novel variables enrich the information set, the findings reveal that the lagged structure of the ISE and traditional hard economic indicators—such as coffee production, oil production, and cement production—remain the most influential predictors of short‑term economic activity.