HYBRID AI MODELS IN FINANCIAL FORECASTING

Abstract

Introduction. Financial forecasting is an applied field in which model errors are rapidly transformed into economic consequences: liquidity losses, incorrect risk pricing, inefficient investment positions, or misleading signals for banking, monetary, portfolio, or regulatory decisions. In the current environment, financial time series forecasting is complicated not only by the growing volume of data but also by regime instability, volatility, information shocks, algorithmic trading, and the increasing complexity of financial behaviour. Under such conditions, single-model approaches often fail to capture the full structure of a financial signal. Classical econometric models remain useful for analysing autoregressive patterns and volatility, but they are limited in dealing with nonlinearities and high-dimensional feature spaces. Machine learning and deep learning models are better suited to nonlinear relationships and large sets of predictors, but they generate additional risks related to opacity, overfitting, instability, and limited auditability.

Purpose. The purpose of the article is to substantiate the theoretical and methodological foundations for applying hybrid artificial intelligence models in financial forecasting and to develop an architecture for their use in financial time series forecasting, taking into account predictive accuracy, regime stability, explainability, and model risk.

Materials and methods. The information base of the study consists of scientific publications on econometric, machine learning, and deep learning forecasting, studies on explainable artificial intelligence, international reports on the use of AI in finance, and open data from the National Bank of Ukraine. Methodologically, the article relies on the systematisation of sources, comparative analysis of modelling approaches, structural and functional modelling, generalisation of empirical research findings, and the principle of walk-forward validation for time series.

Results. The article substantiates a multilayer interpretation of financial time series by distinguishing linear, volatility, nonlinear, noise, and regime layers. The main types of hybrid AI models in financial forecasting are systematised, including sequential, decomposition-based, ensemble, stacking, and explainable hybrid architectures. The article develops a hybrid forecasting architecture that combines data collection, preprocessing, feature engineering, signal decomposition, base models, forecast aggregation, walk-forward validation, explainability, and model risk control. Based on official data from the National Bank of Ukraine, the article demonstrates the regime heterogeneity of the Ukrainian financial context in 2020–2026 using the dynamics of the USD/UAH exchange rate and the NBU key policy rate. It is shown that predictive accuracy is not identical to the financial usefulness of a decision, since practical application of a forecast requires consideration of liquidity, transaction costs, regulatory constraints, risk limits, and human oversight procedures.

Discussion. Further research should focus on full empirical testing of the proposed architecture using daily financial time series from the National Bank of Ukraine. Such testing should compare naive forecast, ARIMA/GARCH, XGBoost, LSTM, and hybrid models using MAE, RMSE, SMAPE, directional accuracy, and error estimates by currency and monetary regimes. A separate research direction should examine how forecasting errors are transformed into financial effects for banks, investors, importers, exporters, and regulators.