Modelling intermittent time series and forecasting COVID-19 spread in the USA

Forecasting intermittent time series represents a challenging task whose importance increases together with the number of series sporadically observed. However, given the difficulties in modelling the presence of zeros, few methods are available. This article introduces a novel state-space approach defined as Intermittent Local Level (ILL). Our approach allows integrating the intermittent nature of time series and forecasting efficiently. Indeed, the proposed state-space model assumes a Bernoulli dynamics that allows switching between zeros and positive values. Moreover, we derive the unobserved dynamics of the time series and provide a simple method for estimating and forecasting. In addition, our approach allows deriving prediction intervals for intermittent observations. Finally, we compare our method's performance with those of standard intermittent models as well as other benchmarks, using the daily number of new cases of COVID-19 observed in nearly 3000 American counties. Predicting the number of newly infected people is important, not only for hospitals but also for policy makers in general. Empirical results show that the suggested approach clearly outperforms the Croston model and its variants when forecasting the number of new Coronavirus cases over a two-week period. In addition, it compares well with non-intermittent benchmarks both in point forecast and prediction intervals.