Forecasting global monthly mpox trends using an SVR-Bayesian autoregressive-Prophet ensemble model

To address the limitations of single models in balancing nonlinear fitting, trend stability, and error control for mpox epidemic prediction, a multi-model ensemble framework was constructed as a methodological exploration for short-term nowcasting under small-sample conditions. Using global monthly mpox surveillance data from May 2022 to September 2025 (41 observations) from the Our World in Data database, three models—Support Vector Regression (SVR), Bayesian Autoregressive (BAR), and Prophet—were established. A two-level (SVR-BAR) and a three-level (SVR-BAR-Prophet) fusion framework were built via weighted optimization with SLSQP. Rolling one-step-ahead forecasting was adopted, with performance evaluated using R 2 , RMSE, MAE, and MAPE. The BAR model provided prediction intervals for uncertainty quantification. Benchmark models (ARIMA, ETS, LSTM, GRU) and ablation experiments were included for comprehensive comparison. Among single models, BAR achieved the best fitting (R 2 = 0.8876). The three-level fusion model achieved the lowest MAE (27.31) and MAPE (7.34%), while the two-level fusion achieved the highest R 2 (0.9128) and lowest RMSE (43.82). All fusion models outperformed individual models and benchmark baselines. Diebold-Mariano tests confirmed that the fusion model significantly outperformed all single models under MAE loss ( p < 0.05), and ablation experiments validated the contribution of each component. The multi-model fusion framework effectively integrates the advantages of individual models, improving prediction accuracy and stability for mpox epidemics. The framework should be considered an exploratory methodological approach requiring prospective validation with future data, but it provides a reproducible reference for infectious disease forecasting under data-limited conditions.