Neural network trained to predict crises in Russian stock market

Economists from HSE University have developed a neural network model that can predict the onset of a short-term stock market crisis with over 83% accuracy, one day in advance. The model performs well even on complex, imbalanced data and incorporates not only economic indicators but also investor sentiment. The paper by Tamara Teplova, Maksim Fayzulin, and Aleksei Kurkin from the Centre for Financial Research and Data Analytics at the HSE Faculty of Economic Sciences has been published in Socio-Economic Planning Sciences.

How can a stock market storm be predicted? Financial analysts and investors worldwide are eager to find the answer. A study by Tamara Teplova, Maxim Fayzulin, and Aleksei Kurkin from the HSE Centre for Financial Research and Data Analytics presents a novel approach to predicting short-term crises in the domestic stock market. The hybrid deep learning model developed by the researchers combines three architectures—Temporal Convolutional Network (TCN), Long Short-Term Memory (LSTM), and an attention mechanism—marking the first use of such a complex structure for Russian stock data.

The authors analysed data from 2014 to 2024, incorporating market and macroeconomic indicators—primarily the Moscow Stock Exchange IMOEX index—along with measures of investor sentiment. To predict the likelihood of a crisis within the next one to five trading days, the researchers first had to address several methodological challenges. First, market crises are relatively rare—accounting for at most a quarter of all events—which makes the training sample imbalanced and risks the model learning to ignore these infrequent signals. Second, investor behaviour is influenced not only by objective economic factors but also by subjective sentiments, which are difficult to formalise. To address these challenges, the researchers created composite indices of internal and external investor sentiment using the principal component method. These indices complement traditional macroeconomic and market variables, making it possible to capture hidden investor sentiment over longer forecasting horizons.

'We present a hybrid TCN-LSTM-Attention model that combines deep learning with attention mechanisms. The model effectively handles imbalanced data, achieving an accuracy of 78.70% for same-day forecasts and 78.85% for predictions on the following trading day. Monthly retraining and the use of adaptive time windows have increased accuracy to 83.87%. Key factors influencing the forecasts include stock index values (similar to those used in technical analysis), total company capitalisation, and exchange rates,' explains Tamara Teplova, Professor at the HSE Faculty of Economic Sciences.

The developed system can be a valuable tool for investors, financial analysts, and regulators. It not only enables retrospective analysis of crisis periods but also allows reliable prediction of potential threats one to two days in advance. When combined with regular updates using new data, such a system can serve as the foundation for a dynamic risk-monitoring framework tailored to the specifics of the Russian market.

'This work is highly relevant for the national financial sector, providing effective tools for timely detection of market shocks—a critical need in an unstable macroeconomic environment,' Prof. Teplova emphasises.

The study was conducted with support from HSE University's Basic Research Programme within the framework of the Centres of Excellence project.