A comparative analysis of time series models for forecasting recovered and deceased COVID-19 cases in India

Time series forecasting methods play a critical role in estimating the spread of an epidemic. Providing insights into the severity of the infection and trend of the outbreak not only helps the general masses to acknowledge the severity of the pandemic but also prompts state officials to take appropriate decisions in due time. Recognising that the use of only one model is never wise and may lead to unreliable forecasts, an attempt has been made to forecast by employing different statistical techniques and their combinations. Specifically, the autoregressive integrated moving average model, innovations state space models for exponential smoothing, the neural network autoregression model, and the trigonometric exponential smoothing state space model with Box–Cox transformation, ARMA errors, and trend and seasonal components were used to forecast the number of deceased and recovered cases of COVID-19. The optimal hybrid combinations of individual models were also tested. The overall goodness of fit of all models was tested with the MAE, MAPE, and RMSE. Compared to single models, hybrid statistical models captured a higher number of properties in the data structure. The forecasts obtained in the case of NNAR from March to July, 2025, for deceased cases are 116.27, 120.30, 113.29, 60.34, 64.71.