Evaluating COVID-19 Pandemic Prediction with PatchMixer and RoBERTa: Integrating News Data with Time Series Forecasting

This thesis explores how integrating multimodal data can affect pandemic forecasting, with a focus on COVID-19, using the PatchMixer model. In structured and cyclical fields like epidemiology, combining textual information—such as news reports and public health guidelines—with numerical time series data is believed to enhance predictive accuracy. We propose a new framework using PatchMixer to predict pandemic trends based on traditional numerical data. This data is then compared with results that were generated by concatenating time series data with textual word embeddings generated using the RoBERTa model. The text is derived from World Health Organization reports and is further refined using prompt engineering with ChatGPT-4 and passed through PatchMixer. A comparison of evaluation metrics—including MAE, MSE, SMAPE, and RMSE—shows a slight improvement when textual data is included, suggesting that it can help improve forecasting performance. However, the study also highlights the limitations of multimodal integration for pandemic prediction. In cases with irregular patterns, traditional time series methods may perform just as well.