AI-Augmented finance for improving risk assessment and portfolio management

Traditional portfolio management methods often struggle to capture nonlinear market dynamics and tail risks, necessitating AI-augmented approaches for improved risk assessment. The paper aims to propose a new hybrid model combining feature extraction through Transformers, the use of generative adversarial network (GAN) to generate scenarios, and Black-Litterman portfolio optimization model to improve risk estimation and portfolio optimization. The framework was tested against CAPM, Markowitz mean-variance optimization, LSTM-based allocation, and conventional GAN- Black-Litterman models using the Multi-Indicator Market and Investment Corpus (MiMIC) dataset (20002023), which captures a wide range of macro-financial indicators and market cycles. The results show that the suggested model consistently outperforms the benchmarks, lowering value-at-risk, conditional VaR, and maximum drawdowns while obtaining the greatest Sortino (0.78) and Sharpe (0.85) ratios. Rolling-window back testing also reveals the solidity of the framework under stable and crisis periods, with smoother portfolio weight paths and greater stability indices, reducing rebalancing expenses. The significance of data-driven embeddings and artificial stress scenarios in capturing nonlinear relationships and reducing tail risks is highlighted in discussion. The results demonstrate that Transformer-GAN Black Litterman is a dynamic, data-enhanced tool that institutional investors can use to gain significant, risk-adjusted returns.