The method and architecture of an adaptive decision support system based on hybrid forecasting

The article considers the problem of decision support in financial markets under conditions of non-stationarity and high volatility, typical for the Russian market during the period of increased macroeconomic uncertainty in 2022–2024. Traditional approaches based on technical analysis with fixed parameters lose effectiveness during abrupt changes in market regimes. A method for dynamically adapting the calculation periods of technical indicators based on a power-law dependence on current market volatility is proposed, ensuring a nonlinear response of the system to changing market conditions. The architecture of the hybrid LSTM-TI-Adaptive model is developed, integrating 33 input features (28 adaptive technical indicators and 5 basic OHLCV features) with a two-layer LSTM recurrent neural network. The proposed method automatically configures the parameters of all indicator classes—trend indicators, oscillators, volume indicators, and volatility indicators—based on a single adaptation mechanism, eliminating manual intervention. The proposed method's effectiveness has been confirmed by experimental testing on 28 financial instruments from the Russian, American, and cryptocurrency markets for the period 2015–2024: the adaptive architecture provides a 4.2 percentage point increase in average annual returns and a 3.1 percentage point decrease in maximum drawdown compared to a system with fixed parameters, confirming the practical significance of the developed adaptation method.