Early warning system for Russian stock market crises: TCN-LSTM-Attention model using imbalanced data and attention mechanism

Abstract

This research is devoted to the development and evaluation of the effectiveness of machine learning and deep learning models for forecasting crisis phenomena in the Russian stock market. The work covers the period from the beginning of 2014 to June 2024, using the IMOEX index as the main indicator of the market condition. Special attention is paid to the problem of the imbalanced data structure and accounting for investor sentiment.

The study presents a hybrid TCN-LSTM-Attention model, which showed the best performance in predicting crisis events. The model achieved an accuracy of 78.70 % for forecasts on the day of observation and 78.85 % for forecasts on the next trading day. Analysis using the Integrated Gradients method identified key factors affecting forecasting, including stock index values, total capitalization of companies and exchange rates.

The study found that the quality of forecasts declines as the forecast horizon increases, but the importance of considering investor sentiment metrics becomes more important. Validation of the model using different time windows and monthly retraining showed a significant improvement in results, achieving an accuracy of up to 83.87 %.

The developed models demonstrate the potential for building early warning systems for stock market crises, which can be useful for individual investors, financial institutions and market regulators alike. Future research could be aimed at incorporating additional factors and developing decision-making strategies based on the obtained forecasts.