Validating the Value-Add of Machine Learning in Multi-Strategy Portfolio Management

Seiryu Ando
Graduate School of Economics, Kyoto University
Master's Program in Economics

Keywords

Empirical Asset Pricing / Hedge Funds / Multi-Strategy / Machine Learning

Research Question

Does machine learning-based strategy allocation provide additional performance improvements over simple multi-strategy portfolios using Equal Weight or Risk Parity allocation?

Research Overview

This study empirically examines whether multi-strategy portfolio management using machine learning adds value over traditional portfolio construction methods.

Specifically, using hedge fund strategy return data, we construct three types of multi-strategy portfolios:

Equal Weight
Risk Parity
Machine Learning-based Allocation

We then compare their performance using metrics such as Sharpe Ratio and Information Ratio.

Background

In the hedge fund industry, multi-strategy management, which combines multiple investment strategies, is widely adopted.

Traditional portfolio construction methods include:

Equal Weight (uniform allocation)
Risk Parity

These rule-based allocations are commonly used.

Meanwhile, empirical research on hedge fund performance suggests that multi-strategy management may generate significant alpha relative to market indices.

For example, Metzger (2019) analyzed approximately 9,500 hedge funds and found that multi-strategy approaches may outperform the market over the long term.

In this study, we examine whether machine learning-based strategy allocation provides additional performance improvements over simple multi-strategy portfolios using Equal Weight or Risk Parity allocation.

Data

Item	Description
Number of Strategies	10 strategies
Analysis Period	February 2022 - November 2025
Return Data	Daily
Allocation	Monthly

Methodology

We construct and compare the following portfolios:

Portfolio	Description
Equal Weight	Uniform allocation across strategies
Risk Parity	Risk-based allocation
ML Multi Strategy	Machine learning-based strategy allocation

We use the following evaluation metrics:

Sharpe Ratio
Information Ratio
RankIC (Rank Information Coefficient)

Main Results

The analysis shows that:

Machine learning multi-strategy achieved the highest Sharpe Ratio
Information Ratio ≈ 0.94 compared to Risk Parity
Sharpe Ratio improvement was mainly due to return enhancement
RankIC showed MeanIC ≈ 0.086, Hit Rate ≈ 63.6%

These results suggest that machine learning-based strategy allocation may contribute to performance improvement in multi-strategy portfolio management.

Research Contribution

This study contributes in the following ways:

Empirically validated the effectiveness of machine learning-based strategy allocation in hedge fund multi-strategy management
Conducted comparative analysis with simple allocation strategies such as Equal Weight and Risk Parity
Evaluated the information content of strategy allocation models using RankIC and Information Ratio

Future Research

Future research directions include:

Alpha verification using Fama-French 4-factor model
Comparison with TOPIX including dividends
Refinement of machine learning allocation models