Eco-epidemiological clustering and cluster-wise forecasting of dengue in Mexico, 2020-2025.

Dengue transmission varies markedly across Mexico, posing challenges for short-term operational planning. We propose a two-layer framework that (i) constructs incidence-free eco-epidemiological clusters for the 32 Mexican states and (ii) produces one-week-ahead forecasts of weekly dengue cases at the cluster level with quantified uncertainty, expressed through P10-P90 prediction bands. Our analysis covers 283 epidemiological weeks per state from 2020 to 2025, with all 32 states retained after meteorological quality control. Clusters are identified using full-covariance Gaussian mixture models applied to seasonal summaries of temperature, relative humidity and vapour-pressure deficit, precipitation, solar radiation, and wind, together with descriptors of weekly seasonality. We select the number of clusters, k, using the Bayesian information criterion or integrated complete likelihood, and we confirm robustness with bootstrap adjusted Rand index, favoring k = 5. For forecasting, we train cluster-specific HistGradientBoosting (Poisson) models using leak-free lagged incidence and lagged or rolling meteorological data, fitting quantile variants (α=0.10, 0.90) to create P10-P90 bands. Over the last 20 weeks, the overall performance measures are a mean absolute error of 17.1 and a root mean square error of 27.7, outperforming a seasonal-naïve baseline in several clusters. The best-performing cluster achieves a mean absolute error of about 5.6 and a root mean square error of about 10.2, while the most variable cluster reaches a mean absolute error of about 38.1 and a root mean square error of about 55.1. Residuals center around zero but show heavy tails during rapid changes. Ablation studies show that climate contributes small, regime-specific improvements at h = 1 (mean absolute error +1.8%, root mean square error -1.4% without climate), whereas removing autoregressive terms more than doubles the error (mean absolute error +109.7%, root mean square error +106.9%). Based on open surveillance and NASA POWER data, this framework provides understandable spatial types and practical, uncertainty-aware forecasts to aid in subnational dengue preparedness.