Graph- and machine-learning-based framework for short-selling risk assessment

We present a novel framework that integrates graph analytics with machine learning to assess the factors influencing the short-selling of publicly traded company shares. This approach centers on constructing a knowledge graph representing selected companies in the banking sector, along with their corporate and individual owners as nodes connected by weighted ownership relations. By extracting network-based features such as the PageRank centrality alongside traditional financial indicators like firm size, ownership concentration, and insider trading activities, this framework identifies and ranks the factors that drive short-selling behavior.

This study employs a regression analysis using models such as random forests, support vectors, and neural networks to quantify the relationship between these features and the short-selling position averages and standard deviations. Features like the largest shareholder’s stake and the Herfindahl–Hirschman Index (HHI) capture the concentration of ownership, while normalized insider trading data provide insights into market sentiment and stock volatility. A comparative analysis using the Shapley Additive Explanation (SHAP) values reveals that although liquidity-related measures are key predictors of the average short-selling positions, the ownership concentration and insider trading are also influential, especially in explaining fluctuations in short-selling activity.

Overall, these results underscore the transformative potential of combining a graph-based network analysis with machine learning techniques to enhance financial risk modeling and governance transparency. This integrated framework not only improves the detection of governance vulnerabilities but also offers valuable insights for regulators and investors. Future research could extend this approach to other sectors with complex ownership networks, further refining the predictive accuracy by incorporating real-time data and additional alternative data sources.