Inhalt

Using Machine Learning for Fund Selection

Swiss pension funds had approximately CHF 1’000 billion in assets under management at the end of 2022, of which more than 20% is managed in the form of equity mutual funds. Worldwide, there are more than 10’000 open-end equity funds available for investment and the number has been increasing over the past years. As a fund selector, how can you pick funds that deliver substantial performance for your clients in the future?

 

By Prof. Dr. Florian Weigert
Chair of Financial Risk Management, University of Neuchâtel    

 

Actively vs. Passively Managed Funds

There is a strong argument that the average passively managed equity fund beats the average actively managed equity fund. As already indicated by Sharpe (1991), before costs, the benchmark-adjusted performance among active traders must be equal to the performance of passive traders. However, as (i) there exist trading costs on financial markets and (ii) fees of active managers are usually higher than those of passive managers, net performance of passively managed funds exceeds the performance of actively managed funds. This arithmetic of active management holds for the average fund; but – would it not be great to pick those few actively managed funds that consistently beat the index (if they exist)?

 

Predictors of Fund Performance

The academic literature has investigated fund and fund manager variables that are positively related to future fund performance. Among those, it has been documented that (a) smaller funds beat larger funds, (b) funds with a distinctive investment style beat quasi-indexers, and that (c) education, experience, and the composition of the fund management team matters for outperformance. And this is not all – there exist more than 50 variables which could be used for fund performance prediction. The question is – how should we combine this information properly?

 

Machine Learning

In a problem like this – i.e., seeking a prediction based on the information of a large set of different prediction variables (so-called features) – machine learning can help you find an optimal solution. Machine learning algorithms are designed to learn from historical data patterns without being assisted by humans; they derive a most favorable prediction given a “big data” input.

Frequently used machine learning algorithms for such prediction tasks are “Regression trees”, “Random forests”, or “Feed-forward neural networks”. These models are able to recognize the importance of nonlinearities and interaction effects of features. As an example, a machine learning algorithm can learn that, during recessions, small funds with an experienced management team perform particularly well. Hence, it will select such a combination during hostile market environments. An observation that a human fund selector is very likely to miss out…

 

Can Machine Learning Select Future Winner and Loser Funds?

How successful are machine learning models to pick funds for the future? Recent research from DeMiguel, Gil-Bazo, Nogales, and Santos (2023) as well as Fausch, Frigg, and Weigert (2023) shows that they seem to perform very well!

Applying a “Random Forest” machine learning algorithm to select funds leads to substantial performance gains over index funds. A portfolio of funds, which was selected by the model to have the best performance over the next 5 years, outperforms the average passive fund by approximately 3% per annum. Moreover, the machine learning can also identify underperforming funds. A portfolio of funds, which was selected by the model to have the worst performance over the next 5 years, underperforms the average passive fund by approximately -4% per annum. Hence, one can apply machine learning also to exclude funds in the selection process.

 

Conclusion

Quantitatively predicting fund performance is a very difficult task. In order to be successful, one has to combine “big data” from different predictor variables in an optimal way. Machine learning models are designed for this task and show promising results to detect future out- and underperformers in the market of actively managed equity funds. To summarize, machine learning can be a potential game-changer when allocating money to investment funds!

 

 

Biography

Florian Weigert is Full Professor of Financial Risk Management at the University of Neuchâtel. He is also a Visiting Professor at HEC Lausanne and Research Fellow at the Centre for Financial Research Cologne. Florian was an Assistant Professor at the University of St. Gallen and visiting scholar at the University of New York and Georgetown University. He obtained his Ph.D. in Finance from the University of Mannheim. Florian’s research interests are within empirical asset pricing, mutual funds, hedge funds, and machine learning. His research was published in the top academic journals, such as the Review of Financial Studies, the Journal of Financial Economics, and the Review of Finance.