A1288
Title: Sentiment-driven forecasting of fractional trading activity
Authors: Christina Erlwein-Sayer - University of Applied Sciences HTW Berlin (Germany) [presenting]
Sakir Kepezkaya - HTW University of Applied Sciences Berlin (Germany)
Natalie Packham - Berlin School of Economics and Law (Germany)
Alla Petukhina - HTW Berlin (Germany)
Abstract: Fractional trading has expanded market access by allowing investors to purchase partial shares of high-priced securities, thereby broadening participation among retail investors. However, fractional trading introduces significant risk assessment challenges for providers. Fractional trading platforms cannot fully hedge their positions, making accurate short-term forecasting of retail order flow essential for effective market risk control. A statistical learning framework is developed that leverages time series data on investor holdings across more than 8,000 securities, combined with daily news sentiment indicators. The approach follows a two-stage structure. First, securities are clustered based on similarities in trading behavior, and representative assets are selected to capture group-level dynamics. Second, long short-term memory (LSTM) neural networks, enhanced with sentiment features, are employed to model and predict trading activity. The results demonstrate that LSTM-based models provide meaningful predictive accuracy for fractional trading flows, offering a forecasting tool for proactive risk management. Incorporating news sentiment further improves model performance in most cases, highlighting the importance of exogenous information in anticipating retail behavior. Additionally, feature importance is analyzed through SHAP values, revealing the key drivers of trading activity and supporting more transparent and informed decisions.
