Improving vector-borne disease early warning system through environmental risk fore-casting: ARIMA vs. Phrophet models for daily temperature prediction
Kh. Ali1,2, I. Ma’rufi3, L.F. Nuraidah4, I.М.D.М. Adyana5, S. Paudel6
1Jenderal Soedirman University, 708 Professor Kh.R. Benjamen St., Purvokerto, 53122, Indonesia
2Oxford University Clinical Research Unit Indonesia, 10 Mega Kunigan Barat III St., build. 1–6, South Jakarta, 12950, Indonesia
3The University of Jember, 37 Kalimantan St., Jember, 68121, Indonesia
4Airlangga University, 4–6 Airlangga St., Surabaya, 60115, Indonesia
5Universitas Hindu Indonesia, Tembau, Penatikh, Salangait St., Denpasar, Bali, 80238, Indonesia
6London School of Hygiene and Tropical Medicine, Keppel St., London, WC1E 7HT, Great Britain
Effective vector-borne disease management increasingly depends on accurate environmental forecasting. In Indonesia, just like in other countries, air temperature plays an important role in vector biology and acts as a main risk factor for pathogen transmission. However, in Indonesia, limited data availability and quality pose significant challenges. This study compares the predictive performance of Facebook Prophet and ARIMA (autoregressive integrated moving average) models for daily temperature forecasting in Surabaya, Indonesia, to assess their potential for supporting climate-sensitive disease surveillance.
Daily temperature data from 2019 to 2024 were modeled using both approaches under consistent conditions. Training datasets were divided into short-term (2022–2023) and long-term (2019–2023) periods, with 2024 data used for out-of-sample validation.
Facebook Prophet effectively captured seasonal patterns but consistently underperformed compared to ARIMA. Quan-titatively, ARIMA models achieved 16.7–17.0 % lower RMSE and 18.8–20.9 % lower MAE across all scenarios. The ARIMA (0, 1, 2) short-term and ARIMA (4, 0, 3) long-term models demonstrated strong residual diagnostics, with error magnitudes below 0.53 °C – within the thermal sensitivity thresholds relevant to dengue transmission. Performance remained stable regardless of training data length.
Statistical tests confirmed significant differences in model performance, favoring ARIMA. The model’s robustness, even with shorter historical data, supports its use in early warning systems for vector-borne disease control, especially in resource-limited settings. Future work should integrate these models with disease surveillance infrastructure and address the impact of climate oscillations on forecasting accuracy, supporting targeted vector control timing, outbreak risk mapping, and early warning system integration for vector-borne disease management.
fcrisk.ru
