Enhanced forecasting model for pandemic: hyperparameter optimisation using Q-learning integration

Hamidreza Rasouli Panah, Abtin Ijadi Maghsoodi, Samaneh Madanian, Jian Yu

Abstract: The effective management of pandemics relies heavily on timely and accurate forecasts to support policymaking and public health decision-making. However, traditional forecasting models ARIMA, SARIMA and ETS often lack the adaptability needed to handle the rapidly evolving and unpredictable dynamics typical of pandemics. This paper introduces an enhancement to the Prophet Forecasting Model by integrating Q-Learning, a Reinforcement Learning algorithm, which treats hyperparameter tuning as a sequential decision-making process. This adaptive framework continuously adjusts parameters such as changepoint flexibility and seasonality components based on forecast accuracy, significantly improving the model’s responsiveness to nonlinear trends, seasonal patterns, and sudden shifts commonly observed in pandemic data. Applied to New Zealand COVID-19 data, the Q-Learning-enhanced Prophet model outperformed ARIMA, SARIMA, ETS, and standard Prophet. At the 7-day horizon, Absolute Error fell from 564.53 to 433.80 and Symmetric Percentage Error from 6.68 to 4.71%, with similar gains across 14- and 30-day forecasts. By providing policymakers and healthcare officials with forecasts that reduced error rates by up to 23% at the 7-day horizon and also longer horizons, this approach can enhance evidence-based policy decisions, resource allocation, and overall pandemic preparedness, with potential applications extending broadly across healthcare operations and dynamic forecasting challenges.

External IDs:doi:10.1007/s10742-025-00367-4