In recent years, machine learning has been touted as a game changer in investment management. Author of “Machine Learning and Fund Characteristics Help Choose Mutual Funds with Positive Alpha” published in the December 2023 issue Journal of Financial Economicsclaimed that machine learning techniques can identify long-only mutual fund portfolios that earn a significant out-of-sample annual alpha of 2.4%, net of all costs. For believers in active management, this was the economic equivalent of searching for the Holy Grail.
What new research reveals about machine learning
Two years later, new researchers conducted a replication analysis of the 2023 study for their paper, “Does Machine Learning Really Help to Select Mutual Funds With Positive Alpha?” These researchers discovered that the original result was caused by a coding error that incorrectly gave the algorithm access to future information. This is a classic case of look-ahead bias.
The error was technical, but consequential. When constructing portfolio returns, the original code used next month’s returns rather than current month’s returns to update the portfolio weights. This means that you are essentially looking into the future when making investment decisions, which is not possible in real-world investing. After fixing this error, the impressive outperformance completely disappeared. Annual returns for the best-performing algorithms decreased by 1.37 to 1.42 percentage points, with no statistically significant algorithms remaining. The authors also identified survivorship bias in the original study.
What actually works (and what doesn’t) with machine learning
The researchers didn’t stop at just identifying the error. They conducted a complete independent replication using new data from 1980 to 2024 and uncovered several important insights.
First, machine learning cannot identify mutual fund portfolios that outperform the market by being long alone. Whether using advanced techniques such as random forests or gradient boosting, or simpler linear models, no approach produced statistically significant positive returns when investing only in top-predicting funds.
This finding holds true across the following areas:
- Various time periods (up to 2024).
- Various forecast periods (12-36 months).
- Multiple risk adjustment models.
However, machine learning has proven effective in identifying funds that are likely to underperform.
The algorithm consistently identified bottom decile portfolios with significantly negative returns of approximately -2% to -3% per year. Both sophisticated machine learning techniques and simple linear regression were successful in flagging the worst performing companies.
When the researchers constructed a long-short portfolio (buying expected winners and shorting expected losers), they found annual returns of 3.00% to 3.05% and were statistically significant. But virtually all of this performance was from the short side, avoiding losers.
Linear models remain unique
Remarkably, simple linear models (ordinary least squares and elastic net regression) performed as well as, and in some cases better than, advanced nonlinear machine learning techniques over a 12-month period.
The benefits of complex machine learning emerged only at longer prediction periods (36 months), where nonlinear methods maintained predictive power while linear models lost statistical significance.
Their findings led the authors to conclude that: “Our findings suggest that ML primarily adds value by avoiding funds that consistently underperform, rather than identifying funds that perform well only over long periods of time.”
Key points for investors
1. Beware of the illusion of picking winners.
The dream of using artificial intelligence to systematically identify top-performing mutual funds remains elusive. Even with sophisticated algorithms and comprehensive data, it has proven impossible to generate consistent long-only outperformance over benchmark returns.
2. Real value: eliminate underperformers
The practical application of machine learning in fund selection is not to pick winners, but to avoid losers. Algorithms can effectively flag funds that exhibit characteristics associated with poor future performance.
- Negative historical alpha t-statistic
- Insufficient value-added indicators
- Exposure to adverse risk factors
3. Simple may be better
For investors with a 12-month time horizon, advanced machine learning offers little advantage over simpler statistical approaches. A basic regression model using fund characteristics such as historical alpha, expense ratio, and factor loadings can be nearly as effective and much easier to interpret.
4. Timeline considerations
For institutional investors with longer horizons (more than three years), more sophisticated machine learning methods performed better, but only identified stocks with future underperformance, rather than outperformers. The researchers found that the nonlinear method maintained its superiority over the linear model only over a 36-month forecast period.
big picture
This study serves as an important reminder that in finance, extraordinary claims require extraordinary evidence. The original discovery was remarkable. This is because it contradicts decades of empirical research showing that active fund management rarely outperforms the market excluding fees, and while many attempts (such as active shares) have failed, there has been no evidence of a strategy that has succeeded in identifying a small number of winners in advance.
The revised analysis is consistent with the following lines of evidence. On average, actively managed mutual funds underperform, and while some tools can help identify funds that may underperform, the search for funds that systematically outperform remains unsuccessful.
Practical advice for investors
Instead of chasing the prospects of the best companies chosen by machine learning, consider these evidence-based strategies.
- Avoid red flags with screening tools: Poor past performance, high expenses, exposure to adverse factors, and the fund’s young age are warning signs worth looking out for.
- focus on cost: It’s very difficult to predict the winners, which makes minimizing expense ratios even more important to your bottom line.
- Consider systemic (negative) alternatives: The difficulty of both identifying winning active managers and implementing effective active selection strategies strengthens the legitimacy of low-cost index funds and other systematic strategies such as Avantis, AQR, Bridgeway, and Dimensional.
- Be skeptical of backtesting: This episode shows how easily illusory performance can be created through coding errors and methodological choices. Always look for independent replication and out-of-sample validation.
Even in the age of AI, machine learning is not magic.
While these tools have some value in flagging funds to avoid, they fail to solve the fundamental challenge of active management: consistent identification. future Out performer.
For most investors, this lesson remains clear. I am deeply skeptical of anyone who claims to have discovered a systematic way to beat the market by focusing on controlling costs, maintaining diversification, and using active management strategies.
