Decoding Russian stock market trends through ensemble methods and sentiment analysis of social media

Annals of Operations Research. 2025. Vol. 353. P. 1123–1172.

Teplova T., Fayzulin M.

The purpose of this paper is to investigate the impact of sentiment metrics derived from social media on stock market attributes, including returns and trading volumes, within three liquidity classes in the Russian emerging market. In this paper, we conduct a textual analysis of users' posts from two Russian-language online platforms: Tinkoff Pulse and MFD. We compare different artificial intelligence algorithms to solve the problem of classifying large textual data. This allows us to determine that the Stacking ensemble method with unoptimised hyperparameters of the metaclassifier gives the best results when classifying text data (processed with TF-IDF), achievincy of about 62% for the test set. With this method, we construct a series of sentiment metrics for retail investors to study their impact on stock returns and trading volumes. The results show that investor sentiment is a significant factor explaining stock returns in most quantiles. In the process of analysing retail investor sentiment, it was found that attention and divergence indices affect retail investor behaviour for different quantiles of stock returns and trading volumes. For example, attention indicators have quite different effects on stock returns for different quantiles, suggesting the presence of several heterogeneous groups of retail investors. Young traders tend to rely on the opinions of Tinkoff Pulse users regarding first-tier (blue chip) stocks, which reflects investors' cognitive biases quite accurately. However, investors who express confidence in sentiment on the MFD platform have a different cognitive pattern, which is explained by the characterisation of users on the MFD platform. These users include older individuals with more trading experience who are more reticent to social media optimism and are more likely to perceive optimism as a signal to sell stocks. Based on the findings, this paper enhances the understanding of the role of retail investor liquidity in emerging financial markets. The lack of analysis of private investors' sentiment and explanation of market trends may not cover a significant gap in knowledge about the psychology of market participants in liquidity classes. Our results provide a foundation for future research that should be useful for investors, regulators, and developers of online social platforms.

Keywords: текстовый анализ, textual analysis, сентимент инвесторов, investor sentiment, hyperparameter optimization, Investor attention, Sentiment metrics, Machine learning model, внимание инвесторов, метрики сентимента, модель машинного обучения, оптимизация гиперпараметров

Publication based on the results of: Models and methods for identifying patterns in global and national financial markets (2025)