Limited Partners versus Unlimited Machines: AI and PE Fund Performance

Abstract

We assemble a proprietary dataset of 395 private equity (PE) fund prospectuses to analyze the relationship between fund characteristics and both fundraising success and future performance. Using econometric methods and machine learning techniques to analyze quantitative and qualitative information in these documents, we find that while PE fund performance is unrelated to quantitative metrics like prior performance and document readability, fundraising success correlates strongly with most fund characteristics but fails to predict future performance. Remarkably, machine learning tools analyzing qualitative information from strategy sections can predict future fund performance with a spread of about 25% between top and bottom terciles of predicted success. These findings suggest that in opaque and non-standardized markets, investors fail to incorporate qualitative information in their asset manager selection process while successfully processing salient quantitative information.

Introduction

Private equity markets have experienced explosive growth over the past two decades, with assets under management multiplying more than tenfold since 2004 to reach $8 trillion in 2022. Despite PE’s growing importance in institutional portfolios and the considerable resources investors dedicate to PE investment decisions, fundraising success and performance remain understudied due to information scarcity. The pervasiveness of asymmetric information in PE markets, combined with non-standardized presentation of information that relies heavily on textual form, may render the investors’ task particularly difficult compared to public markets.

Tamayo et al. address this gap by examining whether investors effectively process both quantitative and qualitative information when allocating capital to PE funds, and whether the information they incorporate actually predicts future performance.

Methodology

The authors collect a unique dataset of close to 400 PE fundraising prospectuses (Private Placement Memorandums or PPMs) and employ a multi-pronged analytical approach:

Data sources and measures:

• 395 PE fund prospectuses from 2003–2016.
• Fundraising success measured through (1) time to raise a fund from launch to final closing, and (2) oversubscription ratio (realized fund size / targeted fund size).
• Performance measured using PME (Public Market Equivalent) at 6 years post-fundraising.
• Strategy sections averaging 2,774 words analyzed using Natural Language Processing.

Analytical techniques:

• Traditional econometric analysis for quantitative variables (past performance, vintage year, fund size, fund sequence, document readability).
• Term Frequency–Inverse Document Frequency (TF-IDF) for textual analysis.
• Machine learning algorithms: Lasso Regression (linear, interpretable) and Gradient Boosting (non-linear, higher performance).
• Strict out-of-sample testing using funds raised between 2014–2016 as test set.

Performance evaluation:

• Area Under the Curve (AUC) for classification accuracy.
• Portfolio analysis comparing top versus bottom terciles of predicted performance.
• Risk-adjustment analysis to ensure predictions aren’t capturing systematic risk factors.

Key Findings

• Quantitative information paradox: While investors successfully incorporate quantitative information (firm reputation, size, past performance) into fundraising decisions, none of these variables predict future performance. Funds with longer prospectuses take longer to raise.
• Fundraising success ≠ future performance: Despite being highly sought after, oversubscribed funds or those that raise capital quickly show no superior future performance, consistent with the Berk and Green (2004) model and diseconomies of scale.
• Machine learning predictive power: Algorithms analyzing qualitative text from strategy sections achieve 60–65% accuracy (AUC) in predicting fund outperformance, with a 25% performance spread between top and bottom predicted terciles.
• Stable linguistic patterns: TF-IDF vectors from strategy sections have remained remarkably stable since 2003, and content is not significantly influenced by law firms drafting the PPMs, suggesting genuine GP-driven information.
• Risk-independent predictions: Machine learning predictions are uncorrelated with standard risk factors (market beta, size, value factors), indicating the algorithms identify manager skill rather than systematic risk exposure.

Implications for Practice

This study reveals a fundamental inefficiency in private equity capital allocation:

• Institutional investors successfully process readily comparable quantitative metrics but fail to incorporate valuable qualitative information embedded in strategy descriptions.
• The 25% performance spread achievable through machine learning analysis represents a significant opportunity for improving LP selection processes.
• Natural language processing tools could democratize access to sophisticated fund analysis, particularly benefiting smaller LPs without extensive due diligence resources.
• The stability of predictive linguistic patterns suggests these techniques could be reliably deployed across vintage years.

Conclusion

The paper demonstrates that machine learning can successfully extract performance-predictive signals from the qualitative content of PE fundraising documents, achieving what traditional LP due diligence processes apparently miss. While investors efficiently incorporate quantitative information into their demand, this information proves uninformative about future performance. In contrast, the qualitative information in strategy sections, which investors appear to overlook, contains substantial predictive power.

For the PE industry, this research highlights a paradigm shift opportunity: moving from reliance on easily quantifiable but ultimately uninformative metrics toward systematic analysis of qualitative information using artificial intelligence. This evolution could lead to more efficient capital allocation in private markets and better alignment between fundraising success and actual fund performance.