| Abstract [eng] |
Financial markets are often perceived as unpredictable due to their chaotic nature and the overwhelming volume of information influencing them. However, with the advancement of data processing capabilities, leveraging this information has become more feasible. In this study, we address two often-overlooked aspects of market prediction using news sentiment analysis. First, we treat sentiment not as a classification problem but as a quantifiable regressor, enabling more nuanced modeling of market impact. Second, we explore the contextual interpretation of sentiment, for example - a headline such as “$120 million in long positions liquidated in the last hour” may appear negative but could imply a market reversal, representing a potential positive opportunity. To investigate this, we firstly try to define our custom headline-to-price impact scores from Bitcoin price to our headlines dataset and select the best one for further proceeding. Later, we employ regression-based machine learning models, as well as try to adapt fine-tuned financial language models from classification to regression tasks and fine-tuning to our weights. After choosing viable sentiment analysis approach, we integrate the derived sentiment scores with market data and Bitcoin technical indicators to integrate it to Transformer-based time series forecasting model. Using model, we evaluate the performance in predicting Bitcoin during different market conditions through backtesting with various trading strategies. Our findings indicate that sentiment analysis is less effective with minimal inputs (price and volume only), but performs better when integrated into more advanced datasets that combine price data with Bitcoin technical indicators and broader market health signals. |
