Application of Machine Learning and Deep Learning in Forecasting Covid Daily Death Counts

Time series modeling and forecasting has fundamental importance to a wide range of applications. While classical time series models dominated the forecasting field for years, their applications have been limited to single time series data at low frequency such as monthly, quarterly or annually. The goal of this paper is to build multi-series time series models to forecast future daily death counts for each county in the state of Pennsylvania. The data used in this paper include JHU daily death counts and confirmed cases and CDC vaccination rates from 1/22/2020 to 1/7/2022 at the county level for Pennsylvania. Both machine learning (Extreme Gradient Boosted Tree "XGBoost") and deep learning (Keras Slim Residual Neural Network Regressor, "Keras") algorithms were explored and time series modeling related steps such as feature engineering, data partition and project setup are discussed in detail. In addition, four metrics were calculated to evaluate the algorithms’ performance. The comparison with a baseline time series model indicated that machine learning and deep learning algorithms did improve forecasting accuracy significantly and Keras has slightly better performance than XGBoost. Finally, the Keras model was utilized to forecast daily death counts for 60 days after 1/7/2022, i.e., 1/8/2022 to 3/8/2022. Based on the model forecasts, daily death counts should gradually ease off by mid-February which has been validated by the subsequent observations. (Abstract)