A contrastive deep learning framework for measuring central bank monetary policy scores

Machine learning and deep learning algorithms have recently been applied to analyze central bank communication texts, thereby providing valuable insights for financial market forecasting. However, as most deep learning methods require thousands or even more training examples, data scarcity often stands in the way when dealing with monetary policy report texts, especially for central banks in developing countries, which communicate their policies less frequently. To address this, we propose a contrastive deep learning framework designed to operate efficiently with small datasets. Despite being trained on fewer than 200 training samples, excellent performance was demonstrated by applying this modeling framework in two scenarios: Measuring China's central bank monetary report's hawkish-dovish score and predicting its next quarter's tightening-easing moves.