Comparative analysis of transformer models for sentiment classification of UK CBDC discourse on X

Abstract Sentiment analysis is critical in understanding public perceptions of evolving currencies such as central bank digital currencies (CBDCs). This study compares three transformer-based models—DistilBERT, RoBERTa, and XLM-RoBERTa—for sentiment classification of UK CBDC-related tweets. Models were fine-tuned on a domain-specific, English-only dataset and evaluated using accuracy, precision, recall, F1-score, training and validation loss across 3 and 5 training epochs. RoBERTa consistently outperformed DistilBERT and XLM-RoBERTa, achieving the highest accuracy and F1-scores. However, extended training epochs in RoBERTa_5 indicated signs of overfitting, making the model fine-tuned for 3 epochs (RoBERTa_3) a suitable model for classifying sentiments on unseen data. These findings provide practical guidance for selecting optimal transformer models for sentiment analysis in specialised financial domains like CBDCs, emphasising the importance of balancing training duration with generalisation capabilities.