In 2023, assets invested in index funds and exchange-traded funds (ETFs) exceeded those invested in actively managed funds for the first time. At the same time, decades of research have shown that actively managed mutual funds tend to underperform the market after fees.
But averages do not tell the whole story. The challenge is identifying future outperformers before that performance materializes.
A new working paper by Dr. Christian Breitung, Manuel Mazidi (Université de Neuchâtel), Prof. Dr. Sebastian Müller, and Prof. Dr. Florian Weigert investigates whether machine learning can help answer this question. Drawing on data from more than 12,500 actively managed equity mutual funds worldwide between 1987 and 2022, the researchers examine whether future risk-adjusted fund performance can be predicted using machine learning models.
The findings provide new insights into when and where machine learning can improve fund selection. We spoke with Prof. Dr. Florian Weigert about the motivation behind the project, the study’s key findings, and what they could mean for the future of investment management.
Prof. Weigert, what motivated you and your co-authors to investigate whether machine learning can improve mutual fund selection?
Every investor would like to know which securities/funds will outperform in the future. However, this is extremely difficult to do so. In our paper, we want to examine whether modern machine learning, applied to a global universe of mutual funds, could uncover more complex patterns in fund characteristics that help predict future performance. We are also interested in understanding whether predictability differs across international markets.
Can machine learning identify mutual funds that are likely to outperform in the future?
Yes, but only to a certain extent. We find that machine learning can successfully distinguish between future winners and losers, particularly in the United States and Europe. However, the models are especially effective at identifying future underperformers, while generating economically significant long-only outperformance remains more challenging.
What was the most surprising finding of the study?
The most surprising finding was that globally trained machine learning models consistently outperform models trained only on local markets. Even when selecting U.S. or European funds, models that learned from international data produced better predictions.
One of your key findings is that globally trained models outperform locally trained ones. Why does a global perspective improve fund selection?
A global model is trained on a much larger and more diverse dataset, allowing it to learn patterns that may not be visible within a single market. Many characteristics associated with successful fund management appear to generalize across countries, so combining international information improves prediction accuracy.
Which fund characteristics provide the strongest signals about future performance?
The strongest predictors are measures of past risk-adjusted performance, especially previous alpha and the statistical significance of that alpha. Fund costs also contain important information, although their effect depends on interactions with other characteristics. Overall, machine learning benefits from combining many variables rather than relying on any single predictor.
What do your findings tell us about the future of active investing in an increasingly passive investment world?
Our results suggest that active investing will continue to have an important role, but not all active managers are equally valuable. Interestingly, markets with a higher share of passive investing exhibit stronger predictability, making it easier to distinguish skilled managers from unskilled ones. In other words, the rise of passive investing may actually increase the value of selecting the right active managers.
What questions remain unanswered, and where does this research go next?
One important question is why predictability differs so strongly across countries. We also want to investigate whether incorporating additional information, such as manager characteristics, portfolio holdings or textual disclosures, can further enhance the predictability of machine learning models.
Looking Ahead
The study contributes to a broader question that is becoming increasingly relevant in finance: under which conditions can genuine investment skills be identified in increasingly competitive financial markets?
Rather than treating fund performance as either universally predictable or universally unpredictable, the findings suggest a more nuanced view. Predictability emerges in specific institutional settings where information quality and market structure allow genuine managerial skill to become visible.
As machine learning becomes an increasingly important tool in investment management, understanding where predictive models work – and where they do not – may become just as important as developing the models themselves.
Read the full paper here:
Breitung, C., Mazidi, M., Müller, S., & Weigert, F. (2026). Machine Learning the Performance of Mutual Funds on a Global Scale. Working Paper, TUM School of Management and Université de Neuchâtel.
