A mixture GARCH-based recurrent neural network for financial volatility forecasting

This paper is devoted to the development of mathematical models of stock price volatility in financial markets, with a focus on the GARCH family models. The paper proposes to consider these models from a new perspective: as recurrent rather than autoregressive. The main idea is that GARCH econometric models can be interpreted as recurrent neural networks, especially after introducing an activation function into the equation of variance dynamics. The relevance of the study stems from the constant need to improve the accuracy of volatility forecasting in modern financial markets, especially in the context of the Russian financial system, where accurate forecasts play a key role in financial decision making. The aim of the study is to evaluate the possibility of representing GARCH models in the form of recurrent neural networks and to assess their applicability for volatility forecasting in Russian financial markets. The main objectives are to develop and test recurrent neural networks based on GARCH, combining the advantages of econometric models and machine learning models. The article proposes a modification of the standard GARCH model called GARCH-RNN, which is a recurrent neural network with multidimensional hidden state and the ReLU activation function. The methods used include econometric analysis of stock price volatility and comparison of forecast accuracy using Moscow Stock Exchange data with GARCH and GARCH-RNN models. The results of experiments on said data showed that the GARCH-RNN model provides volatility forecasting accuracy comparable to that of traditional GARCH models. Results of the study confirmed the potential of the new approach for volatility forecasting on financial markets in Russia, opening prospects for improving forecasts and making informed decisions in the market.