Yellow Fever Vaccine Demand Forecasting With ARIMA, SARIMA, Linear Regression, and XGBoost

The demand for vaccines is significantly increasing in various countries due to heightened population mobility and the prevalence of epidemics. This study employed machine learning methods to predict optimal vaccine stock levels, aiming to prevent both shortages and oversupply, and to compare the effectiveness of these predictions. The data utilized in the prediction models were sourced from the General Directorate of Border and Coastal Health. This study analyzed a 21-year retrospective dataset collected between 2003 and 2023, which contains monthly vaccination coverage data. Four different methods commonly used in the literature were applied to estimate annual vaccine demand. Among these, the most widely utilized method was the Autoregressive Integrated Moving Average (ARIMA). Additionally, Seasonal Autoregressive Integrated Moving Average (SARIMA), Linear Regression, and XGBoost models are employed. Certain events, such as the COVID-19 pandemic, disrupt patterns within the dataset. In pruning tests, variations in data frequency within the raw dataset are analyzed. The models are evaluated using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). The entire dataset is then transformed to achieve stationarity. The models are re-evaluated after removing seasonality and white noise. Cross-validation is applied to the models that yield the most accurate predictions. The forecast results obtained from the optimized model serve as input for the Value at Risk (VaR) model. Actual, projected, and average vaccination numbers are presented with 95% and 99% confidence intervals (critical stock range) based on SARIMA, Linear Regresion and XGBoost estimates. Due to the vaccine forecast range balance, XGBoost's outputs are input into the Value at Risk (VaR) model and the cost risk related to the safe vaccine stock that may arise in the coming days is evaluated. Throughout the study, the conditions under which models can continue to learn effectively, as well as the rationale for selecting these models, can be monitored.