Temporal Dynamics of COVID-19 Outbreak and Future Projections: A Data-Driven Approach.

Long-term predictions for an ongoing epidemic are typically performed using epidemiological models that predict the timing of the peak in infections followed by its decay using non-linear fits from the available data. The curves predicted by these methods typically follow a Gaussian distribution with a decay rate of infections similar to the climbing rate before the peak. However, as seen from the recent COVID-19 data from the US and European countries, the decay in the number of infections is much slower than their increase before the peak. Therefore, the estimates of the final epidemic size from these models are often underpredicted. In this work, we propose two data-driven models to improve the forecasts of the epidemic during its decay. These two models use Gaussian and piecewise-linear fits of the infection rate respectively during the deceleration phase, if available, to project the future course of the pandemic. For countries, which are not yet in the decline phase, these models use the peak predicted by epidemiological models but correct the infection rate to incorporate a realistic slow decline based on the trends from the recent data. Finally, a comparative study of predictions using both epidemiological and data-driven models is presented for a few most affected countries.