Sentiment-Driven Forecasting of Fractional Trading Activity Using LSTM Networks

Fractional trading has emerged as an opportunity for investors to invest in fractions of shares, enabling access to high-priced stocks even with small budgets, and thus reaching a bigger investment community. Aside from the inventory risk faced by all market makers, providers of fractional trading are unable to perfectly hedge their positions and therefore need to forecast retail investors’ order flow over short time horizons to manage their market risk. We develop machine learning models to predict trading activity in fractional trading using time series data on investor holdings of a large set of securities across different sectors combined with daily news sentiment indicators. Our approach proceeds in two stages. First, securities are grouped into clusters with similar trading patterns, and representative firms are selected from each cluster. Second, trading activity is modeled using long short-term memory (LSTM) neural networks augmented with sentiment features. The results show that fractional trading activity can be modeled with meaningful predictive accuracy and that incorporating news sentiment improves LSTM performance in most cases. In two stylised examples, we demonstrate approaches to hedging and market risk management using the LSTM predictions. Feature importance analysis using methods from explainable artificial intelligence (XAI) identifies the primary drivers of trading activity.