Training-Testing Data Ratio Selection for Accurate Time Series Forecasting: A COVID-19 Case Study

Time series forecasting is an essential task in data science and machine learning with applications ranging from finance to healthcare. One pivotal aspect of building accurate models is the selection of an appropriate training and testing data splitting ratio. In this study, an investigation into the effects of varying dataset-splitting ratios on the performance of time series forecasting models is performed. Leveraging a diverse set of time series datasets, we conducted experiments with ratios ranging from 50:50 to 95:05, evaluating multiple forecasting algorithms, including AES, ARIMA, WMA, and SES. Interestingly, splitting ratios between 75:25 and 90:10 consistently yielded the best results across all algorithms and datasets. This range demonstrated remarkable stability, robustness to class imbalance, and superior generalization to unseen data, surpassing extreme ratios like 95:05. Notably, the 85:15 ratio emerged as the top performer, achieving an impressive average accuracy of 95.43% across all algorithms and datasets. These findings highlight the importance of considering factors beyond average accuracy when selecting dataset splitting ratios for time series forecasting. By focusing on a range between 75:25 and 90:10, data scientists and analysts can achieve optimal performance while ensuring model stability and generalization capability. This knowledge provides valuable guidance for informed decision-making in real-world forecasting applications.