Sluggish State-Based Neural Networks Provide State-of-the-art Forecasts of Covid-19 Cases

At the time of writing, the Covid-19 pandemic is continuing to spread across the globe with more than 135 million confirmed cases and 2.9 million deaths across nearly 200 countries. The impact on global economies has been significant. For example, the Office for National Statistics reported that the UK’s unemployment level increased to 5% and the headline GDP declined by 9.9%, which is more than twice the fall in 2009 due to the financial crisis. It is therefore paramount for governments and policymakers to understand the spread of the disease, patient mortality rates and the impact of their interventions on these two factors. A number of researchers have subsequently applied various state-of-the-art forecasting models, such as long short-term memory models (LSTMs), to the problem of forecasting future numbers of Covid-19 cases (confirmed, deaths) with varying levels of success. In this paper, we present a model from the simple recurrent network class, The Multi-recurrent network (MRN), for predicting the future trend of Covid-19 confirmed and deaths cases in the United States. The MRN is a simple yet powerful alternative to LSTMs, which utilises a unique sluggish state-based memory mechanism. To test this mechanism, we first applied the MRN to predicting monthly Covid-19 cases between Feb 2020 to July 2020, which includes the first peak of the pandemic. The MRN is then applied to predicting cases on a weekly basis from late Feb 2020 to late Dec 2020 which includes two peaks. Our results show that the MRN is able to provide superior predictions to the LSTM with significantly fewer adjustable parameters. We attribute this performance to its robust sluggish state memory, lower model complexity and open up the case for simpler alternative models to the LSTM.