Can Machine Learning Predict Fund Performance? Evidence from 33,030 Funds

The Challenge of Fund Selection

With over 33,000 investment funds available across 25 sectors, institutional allocators face a fundamental information problem: how to distinguish genuine skill from historical noise. Traditional approaches—star ratings, peer-group rankings, Sharpe ratios—are backward-looking metrics that conflate past performance with future capacity.

The Fynup Ratio project, initiated in early 2019 using fund data through 2018, takes a radically different approach: it uses only historical annual returns as inputs to a machine learning classification framework. No fund names, no manager information, no strategy descriptions—nothing but the return time series.

Methodology: Science-Based Fund Quality Prediction

The dataset comprises 33,030 investment funds spanning 25 sectors, with track records ranging from 1 year (4,672 funds) to 29 years (84 funds). The deliberately large variation in history length is itself a feature—shorter histories carry higher prediction uncertainty, which the model must internalize.

The sector-agnostic design is critical. By withholding sector labels, the framework forces the model to learn patterns that generalize across fund types rather than memorizing sector-specific return characteristics.

Results and Validation

The out-of-sample validation demonstrates that the Fynup Ratio captures persistent patterns in fund return distributions that carry predictive power for future performance. Funds classified as high-quality by the model exhibit statistically significant outperformance over subsequent periods, while low-quality classifications reliably predict underperformance.

The key insight is that return time-series structure—not just return level—contains information about manager skill, strategy capacity, and risk management quality. A fund that achieves 10% annualized through smooth monthly returns reveals different characteristics than one that achieves the same annualized return through volatile swings.

When the model was tested across sectors, the framework maintained classification accuracy despite the fundamental differences between equity, fixed income, commodity, and alternative strategy return distributions. This cross-sector robustness is the strongest evidence that the model is capturing genuine quality signals rather than sector-specific artifacts.

Implications for Institutional Allocators

The Fynup Ratio framework has direct implications for how institutional investors approach fund due diligence:

• Quantitative pre-screening at scale. With 33,000+ funds, human review of every candidate is impossible. The ML framework provides a systematic first filter that narrows the universe to funds exhibiting quality-consistent return patterns.
• Bias reduction. By excluding fund names, brand recognition, and strategy narratives from the quantitative assessment, the framework reduces the well-documented tendency for allocators to over-weight recent performance and familiar brands.
• History-adjusted confidence. The model inherently weights predictions by track record length—a fund with 20 years of history receives a more confident classification than one with 3 years, reflecting the statistical reality of inference from limited samples.
• Complement, not replacement. The quantitative classification is a screening tool. It identifies which funds deserve deeper qualitative investigation, not which funds to buy.