Detecting Social Stock Pumping Using Machine Learning: Empirical Evidence from the Russian Market

In recent years, the Russian stock market has experienced a growing influence of social media on the behavior of retail investors, which creates preconditions for short-term anomalies in stock prices and trading volumes. While the role of social media publications in driving stock dynamics has intensified since 2019, no comprehensive academic or practical studies have yet explored this phenomenon in depth, which highlights the relevance of the present research. The study aims to identify and quantitatively assess cases of social stock pumping—situations where spikes in discussions and sentiment on social media precede abnormal increases in prices and trading volumes of individual securities. The research tests two hypotheses: first, that social media activity can serve as a causal factor in market anomalies; and second, that machine learning models outperform traditional econometric approaches in detecting such cases. The analysis is based on data for 104 Russian companies covering the period from 2019 to 2025, incorporating market features along with aggregated indices of discussion intensity and investor sentiment. The methodology includes constructing a model of normal stock behavior based on historical data and calculating deviations on days identified as potential social pumping events. For classification, four machine learning models (CatBoost, SVM, Random Forest, KNN) and logistic regression were applied. The results confirm the existence of persistent anomalies in price and volume behavior on social pumping days, along with a statistically significant rise in social indicators. Machine learning models achieved ROC-AUC scores exceeding 0.9, thus outperforming many earlier published approaches. The findings contribute to the theoretical development of behavioral finance and offer practical tools for regulators, analysts, and traders to detect market anomalies and incorporate social influence into trading strategies.