Topological data analysis for cluster-based dengue forecasting in Mexico

Topological Data Analysis (TDA) offers shape-aware descriptions for complex epidemic patterns. We have created a clustered forecasting pipeline for weekly dengue incidence at the state level in Mexico. This pipeline integrates TDA, using persistence diagrams and persistence images (PI), into feature engineering, Gaussian-mixture clustering, and cluster-level count Generalized Linear Models (GLMs). We compare Poisson and Negative-Binomial (NB2) models using a rolling-origin evaluation with naïve and seasonal-naïve baselines. With three clusters, NB2 consistently performs better than Poisson in skill, showing improved Mean Absolute Error (MAE) across clusters. The results for NB2 are {0.264,0.475,0.604}, compared to {0.131,0.382,0.495} for Poisson. An analysis across the Baseline, +TDA, and +TDA+PI demonstrates that adding TDA/PI does not significantly alter one-week forecast skill. However, it enhances interpretability and stratification. The means of PI and differences between cluster and global PI highlight unique, cluster-specific dynamic patterns. The workflow is fully reproducible, including seeds, versions, and configuration, and it generates operational, cluster-based forecasts suitable for regional dengue surveillance.