Abstract

This study evaluated statistical and machine learning time-series models for forecasting monthly nationwide tuberculosis notifications in the Philippines from January 2016 to December 2023. The dataset contained 96 valid monthly observations after removing one trailing summary row. Total notifications were used as the primary forecasting target, while new, relapse, and other retreatment notifications were used for data verification. A chronological split assigned January 2016 to December 2022 to model training and January 2023 to December 2023 to testing. Seasonal naive, exponential smoothing, seasonal ARIMA, STL-ETS, random forest, XGBoost, and ridge lag regression models were compared using RMSE, MAE, MAPE, and sMAPE. The reported Total column matched the annual workbook summary, although four monthly component checks showed small inconsistencies between Total and the sum of the three registration groups. SARIMA (0,1,2)(0,1,1)[12] with drift achieved the best test performance, with RMSE = 8,123.41, MAE = 7,063.50, MAPE = 13.95%, and sMAPE = 15.32%. Residual diagnostics did not indicate significant autocorrelation in the 2023 test errors. A 24-month projection for 2024-2025 was generated from the final SARIMA model fitted to the full series, with prediction intervals used to express uncertainty. Results support seasonal ARIMA as the most useful model in this dataset, while machine learning lag models require longer series or external predictors for stronger performance.

Authors

Samuel John E. Parreno
UM Digos College, Philippines

Keywords

Tuberculosis Forecasting, Time Series, SARIMA, Machine Learning, Philippines