Forecasting Russian stock returns based on investor sentiment analysis in social networks

The study explores the sentiment of Russian private investors in social networks and its impact on the dynamics of the stock return of 78 companies on the Russian stock market (MOEX) in the period from 2018 to 2022. To take into account sentiment when forecasting returns, the authors RSMI index (Russian social media index) is used, which is based on a unique sample of messages from the most popular social networks among Russian investors — “Telegram” and “Tinkoff Pulse”. The RSMI index includes quantitative (the number of publications in relation to each company) and qualitative (investor reactions) characteristics, allowing to determine the real impact of a particular publication on investors. Using the RSMI index, several models for predicting stock prices of Russian companies were used: lasso regression, random forest, gradient boosting, extreme gradient boosting, ensemble learning and long short-term memory. It is demonstrated that for a wide sample of stocks, indicators of technical and fundamental analysis play a large role in building forecasts of changes in stock returns based on hourly data. Although the addition of the sentiment index improves the results of predicting returns for a wide sample of stocks, it does not significantly improve the predictive ability of the models and shows mixed results. The best results of adding the sentiment index to forecast models are shown for the top 15 most discussed Russian companies. For individual models, we achieved an average error reduction of 4.9%, and at the level of specific companies, the MAE error rate was reduced by more than 10% and MSE by 20%. It has been proven that the returns of low-liquidity stocks of the second and third tiers of the Russian stock market are not significantly influenced by the sentiment of private investors on hourly data, and the addition of the sentiment index does not improve the results of forecast models.