The Influence of Twitter Data on the Stocks Prices of Russian Banks and their Volatility

The goal of the paper is to measure the statistical correlation of the sentiment of daily tweets, connected with a bank, with the first value of its stock prices the next day. To calculate the Twitter data sentiment, a Natural Language Processing model ULMFiT is pre-trained on a special dataset. The pre-trained model is used as a base for three models, which are trained on different marked datasets and used to predict probabilities of belonging to positive class. The obtained probabilities are expanded using various non-linear transformations and are then used as features. Using several tests and two families of models (linear regressions as linear and gradient boosting model DART as non-linear) we obtain the results of these features' importances. The results demonstrate a pronounced impact of several features on banks first stock prices of the day.