Towards Intelligent Investment Risk Modeling: Probabilistic Reasoning and Adaptive Decision Systems Under Uncertainty

Investment risk modeling remains challenging due to uncertainty, structural change, and market dynamics. This study develops an integrated empirical framework linking adaptive Bayesian forecasting, predictive-uncertainty evaluation, and risk-sensitive portfolio decision-making in a five-asset system comprising Microsoft, NVIDIA, Tesla, IBM, and Boeing. The TVP-VAR is employed as an established adaptive forecasting component and evaluated against expanding-mean and fixed-parameter VAR benchmarks using point-forecast accuracy, probabilistic calibration, regime-specific performance, and investment outcomes. Results show that the TVP-VAR does not systematically outperform simpler benchmarks in one-step-ahead point forecasting. However, it produces higher mean log scores and 95% predictive-interval coverage closer to nominal levels across all five assets, although these improvements generally require wider intervals and do not yield superior CRPS values. Sensitivity analysis shows that the strongest model-based portfolio performance occurs at λ = 1, where forgetting-factor-driven coefficient variation is eliminated while adaptive covariance updating remains active. Nevertheless, the equal-weight portfolio delivers the strongest overall risk-adjusted performance. Comparison with static covariance indicates that the economic advantage of adaptive covariance estimation is modest and criterion-dependent. Greater parameter adaptivity does not imply universal statistical or economic superiority; its value is more evident in predictive-uncertainty representation and calibration than in consistently improved point forecasts or portfolio outcomes.