Deep Learning Sequence Models for Forecasting COVID-19 Spread and Vaccinations

Abstract

Time series forecasting constitutes an important aspect of traditional machine learning prediction problems. It includes analyzing historic time series data to make future predictions and thus enabling strategic decision-making. It has several practical applications throughout various industries including weather, finance, engineering, economy, healthcare, environment, business, retail, social studies, etc. On the verge of increasing COVID-19 cases globally, there is a need to predict the variables like daily cases, positive tests, vaccinations taken, etc. using state-of-the-art techniques for monitoring the spread of the disease. Understanding underlying patterns in these variables and predicting them will help monitor the spread closely and also give insights into the future. Sequence models have demonstrated excellent abilities in capturing long-term dependencies and hence can be used for this problem. In this paper, we present two recurrent neural network-based approaches to predict the daily confirmed COVID-19 cases, daily total positive tests and total individuals vaccinated using LSTM and GRU. Our proposed approaches achieve a mean absolute percentage error of less than 1.9% on the COVID-19 cases in India time series dataset. The novelty in our research lies in the long-term prediction of daily confirmed cases with a MAPE of less than 1.9% for a relatively long forecast horizon of 165 days.