Information Bottlenecks in Forecasting COVID-19

Abstract

Reliable short term and long term forecasting of the number of COVID-19 incidences is a task of clear importance. Numerous attempts for such forecasting have been attempted historically since the onset of the pandemic. While many successful short-term forecasting models have been put forward, predictions for mid-range time intervals (few weeks) and long-range ones (few months to half a year) appeared to be largely inaccurate.In this paper we investigate systematically the question as to what extend such predictions are even possible given the information available at the times when the predictions are made. We demonstrate that predictions on the daily basis is practically impossible beyond the horizon of 20+ days, and predictions on the weekly basis is similarly impossible beyond the horizon of roughly half a year. We arrive at this conclusion by computing information bottlenecks arising in the dynamics of the COVID-19 pandemic. Such bottlenecks stem from the “memoryless” property of the stochastic dynamical systems describing COVID-19 evolution, specifically from the so-called mixing rate of the system. The mixing rate is then used to gage the rate at which the information used at a time when predictions are made no longer impacts the actual outcomes of the pandemic.