Explainable Neural Forecasting for Small-Sample Time Series: A SHAP-Based Analysis of Dengue Incidence in Cebu City, Philippines

Accurate and explainable forecasting of dengue incidence is essential for effective early warning and public health response. This study develops an explainable deep learning framework for monthly dengue forecasting in Cebu City using 120 months of historical case data and remotely sensed environmental covariates. Time series diagnostics reveal strong trend persistence, nonlinear lag dependence, and a mean‑reverting AR(1) structure, with lagged cross‑correlation observed for land surface temperature at 4 to 6 months.

AutoARIMA, NeuralProphet, NBEATSx, DeepAR, and DeepNPTS were evaluated under a fixed chronological hold‑out validation scheme to prevent data leakage, using 2015 to 2023 for training, 2024 for validation, and Q1 2025 as the final test set. All neural network hyperparameters were optimized via Optuna, which implements Bayesian search and automated pruning to maximize predictive capacity while avoiding overfitting. DeepAR achieved the highest individual accuracy (MAE = 18.00, sMAPE = 8.44%), while a median ensemble of all models substantially outperformed every single forecast (MAE = 8.67, sMAPE = 4.27%). NBEATSx was selected as the primary framework, balancing competitive performance (MAE = 35.67, sMAPE = 19.20%) with post hoc explainability via SHAP. Although NBEATSx was configured with a generic architecture comprising trend, seasonal, exogenous, and identity stacks, the model imposes no hard‑coded interpretable basis functions.

NBEATSx interpretability is derived solely from SHAP‑based feature attribution. SHAP reveals that the trend component dominates forecasts, while lagged dengue cases contribute significant but oscillatory effects reflecting a nonlinear autoregressive behavior, consistent with the estimated AR(1) half‑life of approximately three months. Exogenous variables provide minimal marginal predictive gains at the monthly level. These findings indicate that monthly dengue dynamics are primarily driven by intrinsic temporal structure in this small‑sample setting. The proposed framework provides a transparent, uncertainty‑aware tool for dengue early warning, deliberately trading modest accuracy reductions for full explainability to support data‑driven public health decision‑making.