AI Predicting Fund Manager Trades: NBER Reveals 71% Accuracy

The Research Behind AI Predicting Fund Manager Trades

A groundbreaking NBER working paper is challenging fundamental assumptions about active fund management. The study titled Mimicking Finance, authored by Lauren Cohen, Yiwen Lu, and Quoc H. Nguyen, demonstrates that AI predicting fund manager trades achieves a remarkable 71% accuracy rate — using nothing more than historical behavioral patterns. This finding has profound implications for every investor evaluating whether active management fees are worth paying.

Published as NBER Working Paper 34849, the research introduces a novel framework that decomposes each fund manager’s decision-making into two distinct components: the predictable portion that machines can replicate and the truly novel portion that represents genuine human insight. The results reveal that what many managers consider innovative thinking is often algorithmic repetition of past behavior.

The study falls within the NBER’s Asset Pricing program and intersects with Behavioral Finance and Personnel Economics working groups. It builds on decades of research questioning whether active managers consistently generate alpha, but takes a fundamentally different approach by using frontier AI to measure the novelty of each decision rather than simply tracking returns. For readers interested in how AI is transforming financial analysis, our Interactive Library offers deep dives into the latest research at the intersection of technology and finance.

How the Mimicking Finance Framework Works

The Mimicking Finance framework represents a significant methodological advance in understanding financial decision-making. Rather than evaluating managers solely on returns, it uses frontier AI and machine learning to classify every trade a manager makes along a spectrum of predictability.

The process works in three stages. First, the algorithm ingests a fund manager’s complete trading history — every buy decision, every sell order, every position adjustment. Second, it builds a behavioral model that captures the manager’s habitual patterns, preferred sectors, typical timing, and characteristic responses to market conditions. Third, it uses this model to predict what the manager would do in future periods, comparing those predictions against actual trades.

What makes this framework powerful is its generality. The authors designed it to apply to any key economic agent’s behavior, not just mutual fund managers. Corporate executives, central bankers, and policy makers all operate with behavioral patterns that could theoretically be decomposed into predictable and novel components. The NBER’s Asset Pricing program has long studied how behavioral patterns affect market efficiency, and this paper extends that tradition with cutting-edge computational tools.

The framework’s output is a predictability score for each manager and, crucially, for each individual position within a portfolio. This granularity allows researchers and investors to identify not just who is predictable, but which specific decisions represent genuine insight versus behavioral habit.

71% Prediction Accuracy: Breaking Down the Numbers

The headline finding is striking: across the entire universe of mutual fund managers studied, AI can predict 71% of trade directions without the manager making a single trade. The algorithm knows, with better than two-in-three accuracy, whether a manager will buy or sell a given position in any quarter.

But the average conceals important variation. For some managers, prediction accuracy approaches nearly 100% of their quarterly trades. These are managers whose behavior is so formulaic that an algorithm, trained only on their past decisions, can replicate virtually everything they do. At the other extreme, some managers exhibit genuinely novel behavior that consistently surprises the model.

To understand what 71% means in practical terms, consider a typical mutual fund that makes 200 position changes per quarter. The AI model would correctly predict approximately 142 of those changes. Only 58 trades — less than a third — represent decisions that the algorithm could not anticipate. This ratio raises an uncomfortable question: if machines can replicate most of what a manager does, what exactly are investors paying active management fees for?

The prediction task is specifically focused on trade direction — whether the manager increases or decreases exposure to a given security. This is a meaningful measure because direction captures the core investment thesis. Predicting exact position sizes would be a harder problem, but the directional accuracy alone demonstrates that the fundamental buy-or-sell decision is largely formulaic for most managers. Explore how AI-driven research is reshaping investment strategies in our collection of finance and technology interactive experiences.

Why Some Fund Managers Are More Predictable

The research identifies three primary factors that determine how predictable a fund manager’s behavior will be. Understanding these drivers is essential for investors seeking to evaluate whether their chosen managers are genuinely active decision-makers or merely following habitual patterns.

Length of Trading History

Managers with longer track records are significantly more predictable. This makes intuitive sense: the more data the algorithm has to learn from, the better it can model behavioral habits. A manager who has been trading for 15 years provides the AI with a rich dataset of responses to various market conditions — bull markets, bear markets, sector rotations, volatility spikes. Over time, their behavioral fingerprint becomes clearer and more replicable.

Competitive Environment

Managers operating in less competitive fund categories exhibit more predictable behavior. When competitive pressure is low, there is less incentive to innovate or deviate from comfortable patterns. In highly competitive categories, by contrast, managers must constantly adapt and evolve their strategies to differentiate themselves, leading to less predictable — and often more profitable — behavior.

Personal Ownership Stake

Perhaps the most intriguing finding: managers with larger personal ownership stakes in their funds are significantly less predictable. When a manager has substantial skin in the game, they appear to make more genuinely novel decisions. This aligns with research from the SEC on mutual fund governance, which has long emphasized the importance of manager alignment with investor interests. Personal financial exposure appears to break habitual patterns and motivate truly independent thinking.

AI Predicting Fund Manager Trades and Performance

The most consequential finding in the Mimicking Finance paper is the relationship between predictability and performance. The researchers discover a clear monotonic pattern: the more predictable a manager’s behavior, the worse their returns.

This relationship operates at three distinct levels. At the manager level, less predictable fund managers strongly outperform their peers. The most predictable managers — those whose trades the AI can replicate with near-perfect accuracy — significantly underperform. At the portfolio level, within any individual manager’s holdings, the positions that were hardest for the AI to predict outperform the positions that were easiest to predict. And at the market level, across the entire universe of fund managers each quarter, stocks whose position changes are least predictable by the AI model significantly outperform stocks with the most predictable position changes.

This three-level consistency is remarkable. It suggests that the predictability-performance link is not a statistical artifact but a fundamental feature of financial markets. Novel decisions — by definition, those that defy pattern recognition — appear to capture information or insight that is not yet embedded in market prices. Predictable decisions, conversely, are essentially redundant with the information already reflected in historical behavior.

For investors, the implication is clear: the value of active management lies exclusively in the unpredictable component. The 71% of trades that AI can replicate contribute nothing that a machine could not provide at a fraction of the cost. Only the remaining 29% of truly novel decisions represent the potential justification for active management fees.

Active vs Passive Investing in the Age of AI

The active versus passive debate has raged for decades. Index fund pioneer John Bogle built Vanguard on the premise that most active managers fail to beat their benchmarks after fees. The Mimicking Finance research adds a powerful new dimension to this argument by explaining why most active managers underperform: their behavior is not actually active.

If 71% of a typical fund manager’s trades can be predicted algorithmically, then the majority of what is marketed as active management is functionally passive — it follows predictable patterns rather than generating novel insights. Investors are paying active management fees, typically 0.5% to 1.5% annually, for behavior that could theoretically be replicated at passive-fund cost levels.

However, the research also provides a defense of genuine active management. The unpredictable portion of manager behavior — the approximately 29% of trades that AI cannot anticipate — is associated with strong outperformance. Truly active managers do exist, and they do add value. The problem is distinguishing them from the majority who are, in the words of the paper’s framework, merely mimicking their own past behavior.

This nuance is critical. The research does not suggest that all active management is worthless. Instead, it provides a quantitative tool for separating genuine active management from formulaic decision-making. Investors who can identify the least predictable managers — those operating with genuine novelty — may find that active fees are well justified. The challenge is that most investors lack the tools to make this distinction, which is why innovations in AI-powered financial analysis are becoming increasingly valuable.

Incentive Alignment and Manager Ownership

One of the most actionable insights from the NBER research concerns the role of incentive alignment in driving genuine active management. The finding that managers with larger personal ownership stakes make less predictable — and more profitable — decisions carries significant implications for fund governance and investor due diligence.

When a fund manager has substantial personal wealth invested alongside their clients, several behavioral dynamics shift. First, the personal financial consequences of habitual, underperforming behavior become more painful. A manager who invests $5 million of their own money in the fund feels losses differently than one with no personal exposure. Second, ownership creates a longer time horizon. Managers with significant stakes are less likely to engage in short-term pattern-following behavior driven by career risk, and more likely to make contrarian decisions based on genuine conviction.

The research echoes findings from the CFA Institute’s research on manager alignment, which has documented the relationship between co-investment and fund performance. But the Mimicking Finance framework goes further by providing a quantitative mechanism: ownership reduces predictability, and reduced predictability drives outperformance.

For institutional investors and fund selectors, this creates a practical screening criterion. Rather than relying solely on past returns — which are noisy and often mean-reverting — investors could evaluate manager predictability as a forward-looking indicator of genuine active skill. Managers who are highly predictable, regardless of their recent track record, are essentially offering commoditized behavior at premium prices.

Implications for Investors and Asset Management

The Mimicking Finance research has practical implications across the investment industry. For individual investors, institutional allocators, fund companies, and regulators, the findings demand a reassessment of how active management is evaluated, priced, and governed.

For Individual Investors

The 71% predictability figure should prompt retail investors to ask harder questions about the active funds in their portfolios. If a fund manager’s behavior is largely algorithmic, the investor is effectively paying active fees for passive-like exposure. Tools that measure behavioral novelty — which may emerge as commercial products based on frameworks like Mimicking Finance — could help investors identify truly active managers worth their fees.

For Institutional Allocators

Pension funds, endowments, and sovereign wealth funds allocate trillions to active management. The research suggests that a significant portion of these allocations may be directed toward managers whose behavior adds little beyond what AI could replicate. Integrating predictability analysis into manager selection processes could materially improve allocation outcomes. According to the OECD’s pension fund research, even small improvements in manager selection compound into billions in value over multi-decade horizons.

For Fund Companies

Asset managers face a strategic reckoning. If AI can replicate 71% of their portfolio managers’ decisions, the competitive moat around active management narrows significantly. Forward-thinking firms may respond by developing hybrid approaches: using AI to handle the predictable component of portfolio construction while directing human capital toward the genuinely novel decisions that drive outperformance.

For Regulators

The research raises questions about fee disclosure. If a quantifiable framework can determine what percentage of a manager’s behavior is formulaic versus novel, regulators might consider requiring funds to disclose their predictability scores alongside traditional risk and return metrics. This would give investors a clearer picture of what they are paying for.

The Future of AI in Fund Management

The Mimicking Finance framework opens several avenues for future development. As AI capabilities continue advancing, the 71% prediction rate may increase further, potentially reaching levels that fundamentally alter the economics of active management.

One likely development is real-time predictability monitoring. Rather than analyzing managers retrospectively, AI systems could provide continuous assessments of whether a manager’s current behavior is novel or habitual. This would allow investors to adjust allocations dynamically, increasing exposure when a manager enters a period of genuine novelty and reducing it when behavior becomes formulaic.

Another frontier is applying the framework beyond mutual funds. The authors explicitly designed the methodology to generalize to any financial agent. Hedge fund strategies, corporate capital allocation decisions, central bank policy actions, and even venture capital investment patterns could all be decomposed into predictable and novel components. Each application would provide valuable insights into where human judgment adds genuine value versus where it merely follows patterns.

The interplay between AI prediction and manager behavior may also create interesting feedback loops. If managers know that their behavior is being modeled and scored for predictability, they may actively seek to be less predictable — which, based on the research findings, would likely improve their performance. This awareness effect could paradoxically make markets more efficient by encouraging genuine novelty in professional investment decisions.

For organizations looking to make complex financial research more accessible, interactive experiences offer a compelling format. Research papers like Mimicking Finance contain insights that matter to every investor, but their academic format limits their reach. Transforming them into engaging, interactive content bridges the gap between academic insight and practical investment decision-making.