New AI Optimizes Sparse Portfolios for Better Sharpe Ratios

Traditional sparse tangent portfolio optimization aims to create interpretable, low-cardinality portfolios that align with the mean-variance frontier. However, this task is inherently challenging due to the NP-hard nature of cardinality-constrained formulations. A common issue with standard "predict-then-optimize" pipelines is a misalignment where highly accurate predictions do not always translate into superior downstream portfolio quality. This disconnect can lead to suboptimal investment decisions. To address these problems, a novel end-to-end decision-focused learning framework has been proposed. This framework redefines Sharpe ratio maximization by incorporating it as a Disciplined Parametrized Programming (DPP)-compliant convex programming layer. Crucially, it replaces the problematic discrete asset selection process with a smooth top-k operator, which ensures an exact cardinality k for the portfolio while maintaining differentiability. This innovative design allows for seamless gradient flow across the entire pipeline, from predictive modeling through asset selection and re-optimization. Consequently, the predictive model can be directly trained to optimize actual portfolio performance rather than just forecasting accuracy. Empirical evaluations across four major equity markets demonstrated that this method achieves competitive, and often superior, out-of-sample Sharpe ratios compared to historical and prediction-focused baselines, especially in larger asset universes. The associated code is publicly available for further exploration and implementation.