← Back
Evidence source 5614Spot Checked

Machine Learning Meets Markowitz - SSRN

SSRN2025-12-20Paper
Executive summary

Researchers have developed an end-to-end machine learning system that combines return prediction and portfolio optimization, directly reflecting investor preferences and real-world trading limits. Tested on data from over 5400 Chinese stocks between 2010 and 2023, the model consistently outperformed traditional methods, delivering higher returns and Sharpe ratios. It adapts to transaction costs and risk appetites, learns to prioritize useful signals, and introduces new methods like optimization layers in neural networks. The study is limited to the Chinese market.

What it examines

This paper introduces an end-to-end machine learning framework for portfolio optimization, directly linking asset return prediction with investor preferences and real-world constraints. The approach aims to improve investment outcomes by integrating prediction and decision-making, addressing the limitations of traditional two-stage methods.

What it concludes

The study finds that the end-to-end framework outperforms traditional methods, especially under real-world constraints like transaction costs and risk preferences. Potential applications include personalized investment strategies, tax-aware portfolio management, and asset-liability optimization for pensions. Future research may extend the framework to tax and liability considerations.

Extracted from this source

Evidence objects

Evidence 555264% extraction confidence
A new end-to-end machine learning framework revolutionizes portfolio management by unifying return prediction and optimization, directly addressing investor preferences and real-world trading constraints like transaction costs and short-selling bans.

key_findings bullet 1 · key_findings · validation V0

Evidence 555364% extraction confidence
Tested on over 5,400 Chinese stocks from 2010 to 2023, the model consistently outperforms traditional methods, delivering higher returns and Sharpe ratios while dynamically adapting to transaction costs and individual risk appetites.

key_findings bullet 2 · key_findings · validation V0

Evidence 555464% extraction confidence
Notably, the approach integrates optimization layers into neural networks, prioritizes practical signals, and reduces costly trading, though its focus on a single market and simplified risk calculations are key limitations.

key_findings bullet 3 · key_findings · validation V0

Evidence 555564% extraction confidence
This paper introduces an original end-to-end machine learning framework for portfolio optimization, uniquely integrating return prediction and portfolio construction while explicitly modeling investor heterogeneity, real-world constraints, and transaction costs. Its novelty lies in surpassing the traditional 'predict then optimize' paradigm, demonstrating empirical superiority and practical relevance, making it compelling and impactful.

key_findings bullet 4 · key_findings · validation V0

Raw abstract and provenance

For fundamental characteristics derived from financial statements, we strictly ensure that the data is incorporated only after its public announcement. The

Source row: 1263 · abstract type: snippet