Climate-Driven Dengue Forecasting in Bangladesh: Division-Specific Feature-Set Design and Lag Structure

TL;DR

This study develops climate-informed dengue forecasting models for Bangladesh, comparing feature sets and predictor configurations across two divisions, and benchmarks various machine learning and statistical methods.

Contribution

It introduces a division-specific feature-set design and lag structure for climate-driven dengue prediction, with comprehensive benchmarking of models and predictor configurations.

Findings

Rainfall peaks at 2-month lag in dengue association.

Humidity correlates positively at 1-month lag.

Best models vary: ANN for Dhaka, SARIMAX for Barishal.

Abstract

Bangladesh exhibits marked year-to-year variability in dengue, partly driven by meteorological fluctuations that shape \textit{Aedes} breeding-site persistence, mosquito development, and transmission. We exploit a contrast between Dhaka (consistently high burden) and Barishal (recently rising burden despite lower population density) and frame feature-set design and predictor structure as the main methodological contributions. Using monthly dengue data from DGHS \cite{DGHS} and meteorological data from World Weather Online \cite{Weather} for January 2022--October 2025, we compare four climate feature sets that vary wetness (rainy days vs.\ rainfall) and sunshine (sun days vs.\ sun hours), while temperature and humidity appear in all sets. We evaluate two predictor configurations: lagged climate covariates only, and lagged climate covariates plus 1-month lagged dengue incidence ($Y_{t-1}$). Climate lags (0--4 months) are applied in correlation and forecasting. Both divisions show similar delayed associations: rainfall metrics peak positively near a 2-month lag, humidity near a 1-month lag, sunshine metrics are most negative around a 2-month lag, and temperature is weakly positive at longer lags. We then benchmark MPR, ANN, XGBoost, and SARIMAX across all sets. Best performance differs: Dhaka favors ANN-1 with SET-1 (RMSE=2176.70, MAE=1282.00, MAPE=31.54\%), whereas Barishal favors SARIMAX(0,1,1)(1,0,0,12) with SET-2 (RMSE=817.56, MAE=717.78, MAPE=39.96\%). Analyses use consistent monthly aggregation and division-specific tuning.