AI-Driven Financial Econometrics: Integrating Machine Learning and Stochastic Calculus for Intelligent Business Decision Support

Models that relate predictive accuracy to economically important risk control are needed to make financial decisions. This study assesses an integrated approach that involves GARCH (1,1) volatility estimation, Random Forest regression, validation-weighted hybrid forecasts, stochastic Monte Carlo simulation and conditional value-at-risk (CVaR) portfolio optimization. The daily adjusted prices of eight US exchange-traded funds (ETFs) were studied over the period 2015–2026, where 652 out-of-sample observations were used for forecasts of each asset, and 653 trading days were used for backtests of portfolio decisions. The hybrid had the lowest mean quasi-likelihood loss (QLIKE) of 0.8743 as compared to 0.8949 for Random Forest and 0.9128 for GARCH, and outperformed all the methods for six out of eight assets. During the portfolio evaluation process, the hybrid showed the smallest maximum draw down (-6.05%) and historical CVaR (0.998), and the annualized volatility dropped by about 36.2% compared to the initially equal-weighted benchmark. But the Sharpe ratio in case of Random Forest was better (1.970 compared with 1.918) and monthly return differences were not statistically significant. Results show that better volatility predictions do not necessarily translate into better investment performance and validate the multidimensional assessment of AI-supported financial decision making under explicit risk restrictions. The analysis identifies sensitivity with respect to constraints on the portfolio, costs of transactions, and assumptions regarding distributions.