Evaluating ARIMA, LSTM, and simple benchmarks for short-term COVID-19 case forecasting

Abstract

This study compares Autoregressive Integrated Moving Average (ARIMA) and Long Short-Term Memory (LSTM) models for short-term forecasting of new confirmed COVID-19 cases in the United States. Using national-level data derived from the Johns Hopkins University CSSE COVID-19 repository, we evaluate both models under multiple experimental configurations that vary data frequency, training-test split, historical lookback length, in-put features, and model training strategy. Forecast performance is further compared with persistence and constant-growth benchmarks using rolling one-week-ahead evaluation. The results show that conclusions about model performance are highly sensitive to the experimental configuration. LSTM generally performed better than ARIMA in weekly forecasting, whereas ARIMA performed better under some daily-data settings. Additional epidemiological inputs and expanding-window retraining improved performance in several configurations. However, under the central weekly evaluation, both ARIMA and LSTM were outperformed by the simple persistence and constant-growth benchmarks. These findings demonstrate that increased model complexity does not necessarily lead to better short-term forecasting performance and highlight the importance of experimental design and strong baseline comparisons when evaluating time-series forecasting models.