Abstract

This study investigates the phenomenon of social stock pumping in the Russian equity market and explores effective machine learning models for its detection. Social stock pumping is defined as a market anomaly in which coordinated publications on social media trigger abnormal increases in stock prices and trading volumes without fundamental justification. The paper proposes a methodology for identifying such events based on a combination of behavioral and market indicators. A dataset of 615 social pumping episodes across 104 Russian companies over the period 2019–2025 was constructed. To assess the impact of social media, two proprietary indices were developed: the Russian Social Media Intensive Index (RSMII) and the Russian Social Media Sentiment Index (RSMSI). Five classification models were trained to detect manipulation events: logistic regression, KNN, random forest, SVM and CatBoost. The CatBoost model showed the best performance (AUC-ROC = 0.97, F1 score = 0.91). A comparison with normal trading days confirmed the presence of statistically significant anomalies in prices, trading volumes, and social indicators on pump days. The results demonstrate that machine learning models (particularly CatBoost and KNN) substantially outperform logistic regression in terms of accuracy and recall when detecting social pumping cases. The proposed methodology can be applied by regulators and market participants to monitor informational influence and manage associated risks.

Statistics