Enhancing Portfolio Optimization with Transformer-GAN Integration: A Novel Approach in the Black-Litterman Framework
Enmin Zhu, Jerome Yen
TL;DR
This work addresses robust portfolio optimization by improving the generation of investor views within the Black-Litterman framework. It introduces BL-TGAN, a Transformer-GAN architecture where a Transformer-based generator predicts stock prices and a Transformer-based discriminator refines these forecasts, producing refined views $Q$ linked by $P$ and weighted by $\Omega$ with $\tau=0.025$. Integrating these views into BL yields posterior returns $\mu_{BL}$ and an adjusted covariance $\hat{\Sigma}$, enabling optimized allocations that outperform traditional benchmarks, as evidenced by superior MAE, MSE, and NMSE metrics in the AMZN case study. The approach promises enhanced predictive accuracy and robust asset allocation across market regimes, offering practical gains in risk-adjusted performance and adaptable decision-making for portfolio managers.
Abstract
This study presents an innovative approach to portfolio optimization by integrating Transformer models with Generative Adversarial Networks (GANs) within the Black-Litterman (BL) framework. Capitalizing on Transformers' ability to discern long-range dependencies and GANs' proficiency in generating accurate predictive models, our method enhances the generation of refined predictive views for BL portfolio allocations. This fusion of our model with BL's structured method for merging objective views with market equilibrium offers a potent tool for modern portfolio management, outperforming traditional forecasting methods. Our integrated approach not only demonstrates the potential to improve investment decision-making but also contributes a new approach to capture the complexities of financial markets for robust portfolio optimization.