Machine Learning Can’t Pick Winning Funds. But It Can Help You Avoid Losers

Key Takeaways

• Original study claimed 2.4% net annual alpha using ML.
• Replication uncovered look‑ahead bias in weight‑updating code.
• Corrected results erased outperformance, dropping returns 1.4%.
• Survivorship bias also inflated original findings.
• ML can still aid fund screening by filtering losers.

Pulse Analysis

Machine learning has long been marketed as a disruptive force in portfolio construction, promising to uncover hidden alpha in actively managed funds. Early research, such as the 2023 Journal of Financial Economics article, fed this narrative by reporting double‑digit outperformance after costs. The allure of algorithmic edge attracted both asset managers and retail investors eager for a data‑driven shortcut to beating the market, reinforcing a broader fintech hype cycle that equates sophisticated models with superior returns.

The 2025 replication study, however, pulled back the curtain on methodological oversights that can easily mislead. By inadvertently using next‑month returns to set current portfolio weights, the original code introduced a look‑ahead bias that is impossible to replicate in live trading. Coupled with survivorship bias—where only funds that survived the sample period were analyzed—the corrected analysis showed no statistically significant alpha, and performance dropped by roughly 1.4 percentage points. This episode highlights the perils of data snooping and the critical importance of out‑of‑sample validation, especially when deploying machine‑learning techniques in finance.

For practitioners, the takeaway is nuanced: while ML may not magically pick winning funds, it remains valuable for risk mitigation and loser avoidance. Robust models can flag underperforming assets, improve diversification, and enhance due‑diligence processes when built on clean, forward‑looking data. The industry must adopt stricter testing protocols, transparent code reviews, and realistic benchmark comparisons to ensure that any claimed edge is genuine. As the field matures, disciplined research will determine whether AI can truly add incremental value beyond traditional quantitative methods.