Entropy-based text feature engineering approach for forecasting financial liquidity changes

Changes in individual and institutional financial behavior leading to shifts in liquidity flows often depend on events reflected in news. However, the task of establishing relationship between financial behavior and news remains challenging and understudied. We propose a news-based feature generation approach that allows accounting for news events in liquidity flow time-series predicting tasks, thereby improving the forecasting quality. These features are constructed as different types of entropies and calculated at different levels of text abstraction based on word counts, TF-IDF values, probabilistic topics, and contextual embeddings. We show that this feature engineering procedure is effective for predicting changes in two types of liquidity flows: stock market trading volume and the volume of ATM cash withdrawals. As the first type, we use our original collection of 651, 208 business news articles from a Russian news agency dating to 2013-2021 to predict abnormal jumps in the trade volume of 32 leading Russian companies. With our approach, 97% of them experience an increase in the quality of predicting the differences in daily trading volumes from their median values. For the ATM withdrawals task, we test the impact of economic news from three leading Russian media sources (N = 55, 712) on withdrawals from 100 ATMs located in Moscow. For 95% of them we improve the quality of prediction of year-to-year weekly withdrawal volume change. Additionally, we find that some news sources have a higher predictive power than others. The approach is potentially generalizable for other domains of financial behavior across the globe.