Machine learning meets Markowitz
A new study introduces an end-to-end machine learning framework that directly connects asset return prediction with portfolio optimization, replacing the usual two-step process. Tested on over 5,000 Chinese stocks and 415 features, the model adapts to trading limits, transaction costs, and investor risk preferences. Results show higher returns, better Sharpe ratios, and lower turnover than standard methods. The approach can adjust to individual needs and market frictions, though it does not yet address taxes or asset-liability management.
What it examines
This paper introduces an end-to-end machine learning framework for portfolio optimization, directly linking asset return prediction to investor goals and market constraints. It aims to overcome limitations of the traditional two-stage approach by integrating prediction and decision-making, considering risk preferences, economic restrictions, and market frictions.
What it concludes
The end-to-end approach consistently outperforms traditional methods, delivering higher returns and better risk control, especially under real-world constraints. Potential applications include tax-aware investing and asset-liability management for pensions. Future research may extend the framework to handle taxes and joint asset-liability optimization.
Evidence objects
Using data from over 5,000 Chinese stocks and 415 predictive features, the model adapts to market frictions, delivers higher returns and Sharpe ratios, and lowers turnover, making investment strategies more practical and robust.
key_findings bullet 2 · key_findings · validation V0
A new end-to-end machine learning framework revolutionizes portfolio management by directly linking asset return prediction with optimization, outperforming traditional models, especially under real-world trading restrictions and diverse investor risk preferences.
key_findings bullet 1 · key_findings · validation V0
Notably, the approach customizes portfolios to individual risk profiles, dynamically adjusts to transaction costs, and introduces technical innovations like differentiable optimization layers, though it does not yet address tax or asset-liability management.
key_findings bullet 3 · key_findings · validation V0
This paper introduces a unified, end-to-end machine learning framework for portfolio optimization, directly integrating investor preferences, constraints, and transaction costs into the learning objective. Its originality lies in explicit modeling of investor heterogeneity and real-world factors, demonstrating empirical superiority over traditional methods, making it compelling for quantitative finance advancement.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
- … selection involves two stages: forecast the asset returns and then plug them into an … machine learning tools that unifies the expected return generation process and the final …
Source row: 1262 · abstract type: snippet