International Conference on Technological Intelligence and Business Strategies

Authors - Gaurav Gupta, Kumar Shashvat, Gunjan
Abstract - In India, dengue fever poses a significant threat to public health which continues to worsen. Forecasting methods are crucial to developing effective disease surveillance systems. This study provides an empirical comparison between classical time series forecasting methods, and various machine learning techniques, applied to dengue forecasting for the period of 2013 - 2019 in Chandigarh, India. Seven methods are explored - ARIMA, SARIMA, Exponential Smoothing (ETS), AutoReg, Linear Regression with lagged variables, Decision Tree Regression, and Random Forest Regression. The models are evaluated on multiple criteria which include Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE) and Mean Absolute Scaled Error (MASE), and for the statistical models, the Akaike Information Criterion and the Bayesian Information Criterion (AIC/BIC) are used. Random Forest Regression produced the lowest predicted error (MAE 26.95, MASE 0.19), while SARIMA, with seasonal modeling, demonstrated the best and most useful epidemiological interpretability (MAE 45.36, MASE 0.39) of the models. The outcome of the study shows the balance between predictive power of a public health forecasting model, and the interpretability of the model. In this case, SARIMA had the best balance of both and thus, is recommended as the best model for dengue surveillance systems.