COVID-19: some simple filtering algorithms, November 2021

ABSTRACT

Introduction/ Background: COVID-19 was declared a global pandemic on March 11, 2020 by the World Health Organization. Susceptible-Exposed- Infected-Recovered-Dead (SEIRD) has been used to predict its outbreak. However, it should be handled with precaution when it comes to predicting the end of the waves. Methods: In this study, we use some basic filters to show how bad the prediction of COVID-19 data can be. This study uses the publicly available data from the Center for Systems Science and Engineering (CSSE) through their github repository. Reinforcement learning, machine learning, exponential fitting, exponential smoothing and ARIMA are used on the same COVID data set and same time window. Their root mean square errors as well as their l2 errors are investigated as performance criteria. Results: Using the time horizon of 605 days, the RMSE are 0.6619 for reinforcement learning, 5.7549 for exponential smoothing, 274.3350 for machine learning, 274.3350 for single exponential and 137.5769 for ARIMA for short-term. On a longer-term basis, machine learning, exponential smoothing and single exponential were evaluated using RMSE and the results are 173.2891 for machine learning, 909.5221 for exponential smoothing and 289.2051 for single exponential. l2 errors were plotted on a graph as well. Impact The filters used in this study do not allow us to estimate unreported cases, unreported deaths, hospitalized cases etc. “S+E+I+R+D=N” does not hold in the filter. The use of improved filtering techniques is to be investigated. Conclusion: The methods above can be reasonably good enough for short-term tracking and filtering by designing the parameters properly. For long-term forecasts, however, the trend is different. The basic machine learning method appears to be progressively performant as the training data size increases. The l1-norm needs to be investigated.