The Fallacies of Forecasting Models for Coronavirus (COVID-19) Pandemic in India during Country-wise Lockdowns

Background: COVID-19 is widely spreading across the globe right now. While some countries have flattened the curve, others are struggling to control the spread of the infection. Precise risk prediction modelling is key to accurate prevention and containment of COVID-19 infection, as well as for the preparation of resources needed to deal with the pandemic in different regions. Methods: Given the vast differences in approaches and scenarios used by these models to predict future infection rates, in this study, we compare the accuracy among different models such as regression models, ARIMA model, multilayer perceptron, vector autoregression, susceptible exposed infected recovered (SEIR), susceptible infected recovered (SIR), recurrent neural networks (RNNs), long short term memory networks (LSTM) and exponential growth model in prediction of the total COVID-19 confirmed cases. We did so by comparing the predicted rates of these models with actual rates of COVID-19 in India during the nationwide lockdowns. Results: Few of these models accurately predicted COVID-19 incidence and mortality rates in six weeks, though some provided close results. While advanced warning can help mitigate and prepare for an impending or ongoing epidemic, using poorly fitting models for prediction could lead to substantial adverse outcomes. Implications: As the COVID-19 pandemic continues, accurate risk prediction is key to effective public health interventions. Caution should be taken when choosing different risk prediction models based on specific scenarios and needs. To improve risk prediction of infectious diseases such as COVID-19 for policy guidance and recommendations on best practices, both internal (e.g., specific virus characteristics in transmission and mutation) and external factors (e.g., large-scale human behaviors such as school opening, parties, and breaks) should be considered and appropriately weighed.