Forecasting Weekly Influenza Detections in South Africa Using SARIMA and LSTM Models
Abstract
This study compares Seasonal Autoregressive Integrated Moving Average (SARIMA) and Long Short-Term Memory (LSTM) models for forecasting weekly laboratory influenza detections in South Africa using World Health Organization (WHO) FluNet surveillance data. Weekly laboratory detection counts from January 1997 to March 2026 were preprocessed, regularized, and treated for missing values, yielding 1,522 weekly observations. Data were divided chronologically into training (20 January 1997–11 May 2020; 1,217 weeks) and a fixed-holdout test period (18 May 2020–16 March 2026; 305 weeks). SARIMA incorporated a 52-week seasonal period, while LSTM used a 52-week lookback window with Min–Max scaling fitted exclusively to training data. A Seasonal Naïve model with a 52-week seasonal lag was included as a benchmark. Forecast performance was assessed using RMSE and MAE, while Diebold–Mariano (DM) test assessed differences in forecast accuracy. LSTM achieved the lowest errors, with RMSE of 9.49 and MAE of 6.31, compared with 23.58 and 15.56 for SARIMA and 29.93 and 16.02 for Seasonal Naïve, respectively. Relative to SARIMA, LSTM reduced RMSE and MAE by 59.8% and 59.4%. DM tests indicated significantly lower forecast loss for LSTM than SARIMA using MSE (DM = 6.6661, p < 0.001) and MAE (DM = 9.7569, p < 0.001). The study provides an empirical comparison of statistical and deep-learning approaches using a nearly three-decade WHO FluNet surveillance series. LSTM provided significantly more accurate forecasts than SARIMA under the fixed-holdout design. Future research should examine robustness using rolling-origin validation and multiple forecasting horizons.
Related
- Weekly sales forecasting using time series decomposition with irregular seasonality: A case study of Sony TV sales in the United States
- Forecasting seasonal influenza epidemics with physics-informed neural networks
- Generative diffusion models for spatiotemporal influenza forecasting
- Informed forecasting: Leveraging auxiliary knowledge to boost large language model performance on time series forecasting
- postshock: An R Package for Donor-Adjusted Forecasting After Structural Shocks
BibTeX
@article{ajuka2026forecasting,
title = {{Forecasting Weekly Influenza Detections in South Africa Using SARIMA and LSTM Models}},
author = {Ebelechukwu Ajuka and Ebenezer Esenogho and Cameron Modisane},
journal = {Journal of Information Systems and Informatics},
year = {2026},
month = oct,
doi = {10.63158/journalisi.v8i5.1860},
url = {https://doi.org/10.63158/journalisi.v8i5.1860},
note = {CC BY},
}
RIS
TY - JOUR
TI - Forecasting Weekly Influenza Detections in South Africa Using SARIMA and LSTM Models
AU - Ebelechukwu Ajuka
AU - Ebenezer Esenogho
AU - Cameron Modisane
PY - 2026
DA - 2026/10/24
JO - Journal of Information Systems and Informatics
DO - 10.63158/journalisi.v8i5.1860
UR - https://doi.org/10.63158/journalisi.v8i5.1860
AB - This study compares Seasonal Autoregressive Integrated Moving Average (SARIMA) and Long Short-Term Memory (LSTM) models for forecasting weekly laboratory influenza detections in South Africa using World Health Organization (WHO) FluNet surveillance data. Weekly laboratory detection counts from January 1997 to March 2026 were preprocessed, regularized, and treated for missing values, yielding 1,522 weekly observations. Data were divided chronologically into training (20 January 1997–11 May 2020; 1,217 weeks) and a fixed-holdout test period (18 May 2020–16 March 2026; 305 weeks). SARIMA incorporated a 52-week seasonal period, while LSTM used a 52-week lookback window with Min–Max scaling fitted exclusively to training data. A Seasonal Naïve model with a 52-week seasonal lag was included as a benchmark. Forecast performance was assessed using RMSE and MAE, while Diebold–Mariano (DM) test assessed differences in forecast accuracy. LSTM achieved the lowest errors, with RMSE of 9.49 and MAE of 6.31, compared with 23.58 and 15.56 for SARIMA and 29.93 and 16.02 for Seasonal Naïve, respectively. Relative to SARIMA, LSTM reduced RMSE and MAE by 59.8% and 59.4%. DM tests indicated significantly lower forecast loss for LSTM than SARIMA using MSE (DM = 6.6661, p < 0.001) and MAE (DM = 9.7569, p < 0.001). The study provides an empirical comparison of statistical and deep-learning approaches using a nearly three-decade WHO FluNet surveillance series. LSTM provided significantly more accurate forecasts than SARIMA under the fixed-holdout design. Future research should examine robustness using rolling-origin validation and multiple forecasting horizons.
SN - 2656-5935
SN - 2656-4882
ER -