APPLICATION OF ARTIFICIAL INTELLIGENCE IN FORECASTING RIGHT-TAIL VAR FOR INTEREST RATE RISK MANAGEMENT

Abstract

The study addresses the problem of forecasting the right-tail Value at Risk (VaR) of short-term interest rates in Ukraine during a period of heightened financial market volatility from January 3, 2020, to February 23, 2022, encompassing pre-crisis stability, the COVID-19 pandemic shock, and the initial phase of geopolitical escalation. Under conditions of financial uncertainty, the application of formalized quantitative risk assessment methods capable of capturing extreme interest rate losses and ensuring forecast stability becomes necessary. The objective of the research is to demonstrate the effectiveness of a hybrid approach to right-tail VaR estimation through the integration of classical econometric models (Historical Simulation, GJR-GARCH), machine learning algorithms (XGBoost), and long short-term memory (LSTM) recurrent neural networks into an ensemble structure. The dataset comprised daily interest rates on loans to corporate clients, published by the National Bank of Ukraine, with preliminary processing including logarithmic transformation, stationarity verification using the augmented Dickey-Fuller test, and heteroskedasticity assessment via the ARCH-LM test. Classical VaR methods showed adequacy during stable periods but exhibited inertia during sharp volatility changes, whereas GJR-GARCH provided more precise tail-risk assessment accounting for distribution asymmetry. Machine learning models XGBoost and LSTM captured nonlinear dependencies and complex temporal patterns, enhancing forecast adaptability to structural market shifts. The ensemble, formed as a linear combination of base model forecasts with weights inversely proportional to NRMSE, achieved the lowest breach rate (0.87%), minimal residual volatility (σ = 0.007), and high consistency of forecasts with actual log-returns. The scientific novelty lies in developing an integrated methodology for forecasting right-tail VaR in the Ukrainian short-term credit market characterized by high volatility, while practical significance is reflected in its applicability for implementation in bank risk management systems and corporate finance departments for effective interest rate risk management in post-crisis and conflict-affected environments. The proposed hybrid approach minimizes forecast errors, stabilizes VaR dynamics, and provides a balance between accuracy, stability, and risk assessment adaptability.