Fig 1.
Time series of dengue cases from 2004 to 2022, highlighting the initial training period (red), the time-series cross validation (TSCV) phase (blue), and the evaluation period (grey).
The initial training data is used for model calibration, followed by TSCV for model validation, and concluding with an evaluation period for assessing forecast accuracy.
Table 1.
Summary of model performance metrics and features for dengue incidence prediction three months in advance for the baseline model, the ensemble model, and the five individual models included in the ensemble, aggregated across the Mekong Delta Region districts, 2012-2016. Lower CRPS indicates better accuracy, while bias closer to zero suggests closer alignment with observed values, and lower diffuseness reflects more precise predictions. Weight is the weight each of the five individual models contributed to the ensemble.
Fig 2.
Comparison of performance metrics for the baseline (reference model), the ensemble model, and the five individual models included in the ensemble across forecast horizons of 1 to 3 months, aggregated across the Mekong Delta Region districts, 2012-2016.
Panel (a) shows CRPSS, with higher values indicating a better forecast compared to the baseline model. Panel (b) displays bias, with values closer to zero reflecting more accurate predictions. Panel (c) presents diffuseness, which measures the spread of the predicted probabilities, with lower values indicating greater precision. Panel (d) presents Brier score, using Mean+2 SD, with a lower value indicating more accurate outbreak classification.
Fig 3.
Observed vs. forecasted dengue cases (mean and 95% predictive interval) using the ensemble model, by time horizon, aggregated across the Mekong Delta Region districts, 2012-2016.
The solid red line represents observed cases, the solid blue line forecasted cases with a 3-month horizon, the solid green line forecasted cases with a 2-month horizon, and the solid yellow line forecasted cases with a 1-month horizon. The dashed lines correspond to the 95% credible intervals for the forecasts.
Fig 4.
Brier score averaged by month across the three time horizons using the ensemble model with five different threshold definitions, aggregated across the Mekong Delta Region districts, 2012-2016.
The Brier score can range from 0 to 1, with 0 representing the optimal value.
Fig 5.
Performance metrics for the baseline (Reference) model, the ensemble model, and the five individual models included in the ensemble across forecast horizons of 1 to 3 months, Mekong Delta Region, during the 2017-2022 out-of-sample evaluation period.
Panel (a) shows CRPSS, with higher values indicating a better forecast compared to the baseline model. Panel (b) displays bias, with values closer to zero reflecting more accurate predictions. Panel (c) presents diffuseness, which measures the spread of the predicted probabilities, with lower values indicating greater precision. Panel (d) represents the Brier score, using Mean+2 SD, which can range from 0 to 1, with 0 representing the optimal value.
Fig 6.
Time-series plot of predicted dengue cases from the top five individual models compared to observed cases (2017-2022).
This plot illustrates the temporal patterns and predictive performance of different models over the years. Notably, there is a visible misalignment in predictions for the years 2019 and 2022, indicating challenges in model accuracy during these periods.