There is naturally significant interest in identifying mutual funds that can deliver positive risk-adjusted returns. Despite this, actively managed mutual funds often underperform and selecting mutual funds that can deliver superior performance on an ex-ante basis is a notoriously difficult task.
In Barcelona GSE Working Paper 1245, “Can Machine Learning Help to Select Portfolios of Mutual Funds?” Victor DeMiguel, Javier Gil-Bazo, Francisco J. Nogales and André A.P. Santos take on the challenge of identifying outperforming mutual funds. To do so, they propose the use of machine learning (ML) techniques to predict future fund performance from observable fund characteristics.
The authors state that their method has three distinct advantages:
- The method does not rely on one predictor alone, as is the case in previous work, but exploits information contained in a large set of variables.
- ML algorithms are better suited to handle the complex relationships between mutual fund performance and fund characteristics.
- Their approach is dynamic and adapts to changes in the financial landscape. Their method is directly applicable by investors as it combines readily available techniques with public data.
Three models of machine learning
The authors selected three broad classes of ML methods:
- Elastic Net
- Random Forests
- Gradient Boosting
The Elastic Net approach applies the same linear approximation as a linear regression but improves estimation when explanatory variables are correlated. Moreover, to extend the linear approximation and capture other, more sophisticated relationships between the explanatory variables, the authors use two forms of decision trees, Random Forests and Gradient Boosting, as well as Neural Networks. The authors explain that these methods often outperform linear methods in terms of prediction performance in general applications, especially with the type of data used in this project.
The data used in the paper spans the period 1980-2018 and the analysis is carefully designed to resemble as close as possible the actual selection problem faced by investors. This ensures that the results of the paper are both practical and relevant. The authors then employed the following process for each of the three models.
The first 10 years of data are used to train the ML method to predict one-year ahead fund risk-adjusted performance net of transaction costs, fees and other expenses. As predictors, they consider the values of several fund characteristics in the previous year, including age, manager tenure, and volatility. Then the authors ask the algorithms to predict performance in the following year and form a portfolio of funds consisting of those in the top decile of the predicted performance distribution. The funds are kept in the portfolio for 12 months, during which portfolio returns are recorded.
For every remaining year, they roll the sample forward one year, train the algorithms again on the expanded sample, make new predictions for the following year, rebalance the top-decile portfolio and track its return during the next 12 months. This way, the authors construct a time series of monthly returns for the top-decile portfolio. They can then evaluate the overall performance of this top-decile portfolio over the whole sample period.
Simulating market returns
Two of the three algorithms considered, Gradient Boosting (GB) and Random Forests (RF), produced tangible results. Both were able to select a portfolio of funds that delivered positive and statistically significant risk-adjusted returns. In particular, the top-decile portfolio constructed with the GB algorithm earned between 3.5% to 4.2% per year more than the market benchmark, net of all fees, expenses, and transaction costs. When using RF to select funds, results were slightly lower and ranged from 2.4% to 3% per year. Neural Networks performed similarly to RF. The top-decile portfolios selected by both Elastic Net and Standard Linear regression delivered positive excess returns, although substantially lower than those selected by gradient boosting and statistically indistinguishable from zero.
For comparison purposes, the authors also built two baseline portfolios: an asset-weighted and an equally-weighted portfolio of all available mutual funds. These naive methods underperformed the ML methods as well as linear regression and earned negative risk-adjusted returns. This result matches the wealth of previous literature on mutual funds, which outlines the challenges faced by managers in producing positive results when using standard techniques. Overall, while portfolios that exploit predictability in the data, such as the three ML methods utilised here, help investors to avoid underperforming funds, only gradient boosting, random forests and neural networks allow them to benefit from investing in actively managed funds.
The authors go on to analyse the relative importance of different explanatory variables within their data. They find that there is significant variation through time in which variables provide the most information when selecting portfolios. In addition, in none of the models examined does only one predictor dominate when selecting top-performing mutual funds. For example, for the Gradient Boosting model, the second and third most important fund characteristics are almost as important as the first, while the fourth and fifth are half as important.
Importantly, the ranking of the most important characteristics varied across the different models in use. While the Gradient Boosting model relied heavily on realised risk-adjusted returns in the previous year, this was relatively unimportant for Random Forests and almost ignored by the Elastic Net and Linear Regressions. These findings further justify the use of methods capable of handling numerous predictors and adapting to changes over time.
Real world performance
Literature has shown that the predictive ability of fund characteristics with respect to future fund performance has declined through time due to an increase in arbitrage activity and competition among mutual funds. Consistent with such findings, returns produced by the authors’ methods declined through the sample period for all portfolios, including the top performing GB-selected portfolio. This result suggests that the best performing Machine Learning algorithm is able to extract positive risk-adjusted returns from the mutual fund market, but only when any such returns exist in the first place.
The results presented in the paper are of great practical importance for investors, financial advisers, managers of funds, and pension plan administrators. The methods proposed are readily implementable and can be used to improve fund selection. Since the data appear to show a tightening of market returns and as such greater difficulty in selecting the right portfolios, the methods presented in this paper can guide investors towards the right choices.