Do AI-Powered Mutual Funds Perform Better? (2022)
Rui Chen, Jinjuan Ren
Finance Research Letters, Volume 47, Part A, URL/SSRN
Almost one year ago, I concluded my post on Big Data & Machine Learning in Asset Management with the following statement. “Picking an outperforming stock is not as easy as identifying a cat and probably will never be. That said, there is also no reason why a smart model cannot be better at it than a human portfolio manager.”
This week’s AGNOSTIC Paper is about exactly this statement. The authors compare the performance of AI-powered- and human mutual funds in the US.1I know, “human funds” sounds somewhat strange but I didn’t came up with a better label. And of course, AI-powered funds are ultimately also managed by humans. Luckily, the results are in line with my post one year ago.
Everything that follows is only my summary of the original paper. So unless indicated otherwise, all tables and charts belong to the authors of the paper and I am just quoting them. The authors deserve full credit for creating this material, so please always cite the original source.
Setup and Idea
For more background on AI and machine learning in the investment industry, I recommend reading my first post on the issue. But in general, the setup of the paper is quite simple. The authors obtain data for US mutual funds and compare the performance of those that are solely powered by AI-models to the overall US market and their “traditional” human peers.
The authors also present two hypotheses why AI-powered funds might be superior. First, AI-models and de facto unlimited computing power allow to analyze more data in less time. Second, portfolio managers are by definition just humans and research shows that even professionals sometimes suffer from behavioral biases.2Statman (2019) has written a good overview on this topic for the CFA Institute. On the other side, research also shows that investing is a very difficult application for AI and that we shouldn’t expect the same breakthroughs as in other areas.3Israel et al. (2020) give a good overview on the reasons for this. So there are valid arguments on both sides and therefore, let’s look at the data.
Data and Methodology
The authors obtain data on US mutual funds from CRSP, a high-quality data provider for financial research.4CRSP stands for Center for Research in Security Prices and is an affiliate of the University of Chicago. More information here. For the period between 2017 and 2019, they also collect the prospectuses of 2,133 (equity) funds and manually label them into different categories. The start of the sample is determined by the inception of the first AI-powered US mutual fund in October 2017.5Information according to the authors. I haven’t checked if there were any AI-funds before that date available.
According to the definition of the paper, AI-powered funds use (only) machine learning methodologies to select stocks and/or construct portfolios. Importantly (and correctly), the authors exclude thematic funds that invest in companies related to AI but are managed by “traditional” portfolio managers. On the other side, human funds mostly rely on human judgement to make investment decisions.
Overall, the authors identify 15 AI-powered funds between 2017 and 2019. This is obviously not a large sample but there is not much we can do about it. The first AI-powered US mutual fund launched in October 2017, so this is the entire history. Note however, that this is not representative for the application of AI in the investment world in general. More sophisticated players like hedge funds adopted such techniques much earlier than the regulated mutual funds and there are definitely more than 15 funds that use AI.6For example, the famous Medallion Fund of Renaissance Technologies presumably uses machine learning since the 1990s.
To compare the performance of AI-powered- and human funds, the authors construct a peer group for each AI-powered fund. For this purpose, they select human funds with comparable investment styles, track records, and descriptions. Finally, they also obtain holding data of funds and test for the disposition- and rank effect, two well-known behavioral biases.
Important Results and Takeaways
AI-powered mutual funds did not outperform the US market
In the first part, the authors calculate various performance statistics for the AI-powered mutual funds. They also estimate the skill of fund managers (i.e. AI-models) via Picking and Timing. Picking captures the ability to select stocks that perform better than the benchmark. Timing is the ability to correctly adjust weights of stocks before the corresponding market movements. The following table summarizes the results.
Panel A shows the monthly performance of AI-powered funds versus the market-cap-weighted US stock market. The results already suggest that there is no persistent outperformance. The monthly returns are more or less comparable and none of the t-statistics suggests statistically significant outperformance. Panel B also supports those results. Market-adjusted returns and five-factor alphas are negative (although not significant) for both the equal- and value-weighted average of AI-powered funds.
So at least during the short period from 2017 to 2019, AI-powered funds did not outperform the US market. In line with those results, Panel C indicates that AI-models lack both picking- and timing skills. The picking-component for the equal-weighted average is somewhat positive but the overall pattern is once again no sales pitch for actively managed mutual funds (even if they use AI). But as we will see in the next part, it can get even worse.
But AI-powered funds outperformed their human peers
The following table shows the same performance statistics but now for the comparison of AI-powered funds and their human peer group. The results are quite different. Although AI-powered funds underperformed the market, they still performed significantly better than their human peers. For example, the average outperformance of AI-powered funds versus their human peers was 0.48% per month.
Panel B shows a similar pattern. Market-adjusted returns and alphas of AI-powered funds remain statistically indistinguishable from zero. But their human peers are with significantly negative values even worse. Regarding the timing- and picking skill, the picture is not so clear. AI-powered funds seem to be better in picking while their human peers are somewhat better in timing. But neither of this differences is statistically significant, so it’s hard to make reliable statements.
Once again a summary of the results so far. Both AI-powered- and their “traditional” human peers did not generate significant outperformance during this short sample period. This is of course unsatisfying, but the AI-powered funds were at least significantly better than their human peers.7Or more accurately, “AI-powered funds were less worse”. So although passive investing remains probably the best deal for most people, AI-powered funds could still improve the investment industry.
And AI-powered funds avoided the disposition- and rank effect
One of the most evident explanations for the better performance of AI-powered funds is that statistical models don’t suffer from behavioral biases.8AI is of course developed by humans, therefore biases will also find their way into the models. So the danger remains… The authors use fund-holdings to examine how AI-models and humans cope with two well-known behavioral biases. The disposition effect9To the best of my knowledge, first mentioned by Shefrin and Statman (1985). (the tendency to sell winners too early and hold losers too long) and the rank effect10To the best of my knowledge, first mentioned by Hartzmark (2015). (the tendency to focus and trade mostly the best- and worst-performing securities in a portfolio). The following table summarizes the results.
The proxies for the two behavioral effects (see caption of the table) indicate that AI-powered funds indeed suffer less from those biases. The negative differences in the third column suggest that AI-powered funds were more willing to sell losers (disposition effect) and traded more stocks from the “middle” of the portfolio (rank effect). In my opinion, those results are quite plausible and highlight one key advantage of systematic investment processes.
Conclusions and Further Ideas
Although I like the paper, the small sample of course remains a problem. Two years of data for only 15 AI-powered funds is simply not that much. But unfortunately, this is what it is. AI in investment management is a relatively new phenomenon and the oldest fund in the sample was just launched in October 2017. So at least for the regulated mutual fund world, history and breadth is limited.
Another potential issue with the sample is somewhat beneath the surface. 12 of the 15 AI-powered funds just use AI to construct more sophisticated industry classifications.11This information comes from the appendix of the unpublished version of the paper. I also tried to look up some of these funds and found that many of them were closed in the meantime. The authors classify some of them as active, however, I doubt that the comparison to their human peers is fair. Portfolio managers of industry funds usually try to select winners within a defined industry instead of coming up with their own industry classifications. So this might be a different objective and I don’t know if the authors control for that.
Despite those problems, the main results of the paper remain interesting and also promising. While AI-powered funds are not the holy grail some investors may have hoped for, they still added value compared to their “traditional” human peers. I think that especially this second part is important. Even with sophisticated AI-models, it is very hard to consistently beat a passive benchmark. But apparently, it is even harder without them. So I still stand by my statement from last years post. Yes, applying AI or machine learning in investment management is much harder than in other areas and the pioneering funds struggled to outperform the market. But humans without the help of AI did even worse. So again, “there is no reason why a smart model cannot be better at picking stocks than a human portfolio manager.”
- AgPa #76: ESG Myth Debunking
- AgPa #75: Optimal Investment Committees
- AgPa #74: Peer-Reviewed Research is Not Helpful to Predict Returns – Really?
- AgPa #73: Country and Industry Momentum
This content is for educational and informational purposes only and no substitute for professional or financial advice. The use of any information on this website is solely on your own risk and I do not take responsibility or liability for any damages that may occur. The views expressed on this website are solely my own and do not necessarily reflect the views of any organisation I am associated with. Income- or benefit-generating links are marked with a star (*). All content that is not my intellectual property is marked as such. If you own the intellectual property displayed on this website and do not agree with my use of it, please send me an e-mail and I will remedy the situation immediately. Please also read the Disclaimer.
Endnotes
| 1 | I know, “human funds” sounds somewhat strange but I didn’t came up with a better label. And of course, AI-powered funds are ultimately also managed by humans. |
|---|---|
| 2 | Statman (2019) has written a good overview on this topic for the CFA Institute. |
| 3 | Israel et al. (2020) give a good overview on the reasons for this. |
| 4 | CRSP stands for Center for Research in Security Prices and is an affiliate of the University of Chicago. More information here. |
| 5 | Information according to the authors. I haven’t checked if there were any AI-funds before that date available. |
| 6 | For example, the famous Medallion Fund of Renaissance Technologies presumably uses machine learning since the 1990s. |
| 7 | Or more accurately, “AI-powered funds were less worse”. |
| 8 | AI is of course developed by humans, therefore biases will also find their way into the models. So the danger remains… |
| 9 | To the best of my knowledge, first mentioned by Shefrin and Statman (1985). |
| 10 | To the best of my knowledge, first mentioned by Hartzmark (2015). |
| 11 | This information comes from the appendix of the unpublished version of the paper. I also tried to look up some of these funds and found that many of them were closed in the meantime. |