Governing the Trade-Off Between Time Period Length and Observations’ Number with Machine Learning: A Number of Previous Days Needed for Prediction of Future COVID-19 Positives’ Count Using Czech Data

The ongoing impact of COVID-19 waves on daily life since early 2020 highlights the need for accurate prediction of new daily positive cases, even with the significant progress in vaccination efforts. Concerns persist about potential new variants that could be highly mutated and resistant to post-vaccination immunity. In this study, we focus on predicting the daily count of new positive cases using various factors, including the number of deaths, hospitalizations, vaccinations, and reproduction numbers from recent days, utilizing publicly available COVID-19 data from the Czech Republic. One crucial aspect we examine is the varying time period length, representing the number of previous days used for predicting the next-day positive cases. Longer time periods are typically thought to improve prediction performance, but they come at the cost of having fewer complete series of observations with the last days available for prediction. This balance between time period length and the number of observations is a critical consideration in our analysis. To navigate this trade-off, we employ machine learning methods, including multivariate regression, least absolute shrinkage and selection operator, ridge regression, support vector machines, and random forests. Within each algorithm, we search for the optimal time period length that minimizes the root mean square error of the prediction, helping us determine the most accurate number of previous days to use for predicting the next-day COVID-19 positives.