SELECTION OF DEEP NEURAL NETWORK ARCHITECTURES FOR FORECASTING FINANCIAL MARKETS UNDER CONDITIONS OF HIGH VOLATILITY

Abstract

The paper addresses the problem of selecting deep neural network architectures for financial market forecasting under conditions of high volatility, nonstationarity, and a high level of noise in financial time series. Traditional statistical and econometric approaches often fail to capture nonlinear dependencies and rapidly changing market regimes, which necessitates the use of deep learning models capable of adaptive representation learning and multiscale temporal analysis.

The study provides a systematic comparative analysis of recurrent neural networks (RNN), LSTM and GRU architectures, convolutional and temporal convolutional networks (CNN, TCN), attention-based models, transformer architectures, and neural language models for integrating textual financial information. Particular attention is paid to the prediction lag problem, which significantly reduces the practical value of forecasts in highly dynamic market environments.

It is shown that recurrent architectures effectively model long-term dependencies but may exhibit delayed responses to rapid short-term fluctuations. CNN and TCN models demonstrate higher robustness to noise and better extraction of local patterns, while transformer-based models enable global context modeling and faster adaptation to new information. Neural language models allow the integration of news flows and sentiment analysis into forecasting systems, improving responsiveness to market events.

Experimental results obtained on volatile financial time series indicate that hybrid architectures combining convolutional, recurrent, and attention mechanisms provide the best balance between forecasting accuracy, robustness, and computational efficiency. Such models simultaneously capture short-term dynamics, long-term trends, and contextual information, reducing prediction lag and improving generalization under changing market regimes.

The proposed systematization of deep neural network architectures and their characteristics can be used to justify the selection of forecasting models and to develop information technologies for financial market prediction aimed at increasing accuracy, stability, and adaptability in high-volatility conditions.