{"id":8079,"date":"2026-03-05T12:33:20","date_gmt":"2026-03-05T11:33:20","guid":{"rendered":"https:\/\/bicon.li\/?p=8079"},"modified":"2026-03-06T20:45:56","modified_gmt":"2026-03-06T19:45:56","slug":"ai-predictability-and-the-real-value-of-active-management","status":"publish","type":"post","link":"https:\/\/bicon.li\/en\/2026\/03\/05\/ai-predictability-and-the-real-value-of-active-management\/","title":{"rendered":"AI, Predictability and the Real Value of Active Management"},"content":{"rendered":"

Recent advances in artificial intelligence and machine learning are beginning to reshape how we think about decision-making in financial markets. A new study by Lauren Cohen, Yiwen Lu and Quoc H. Nguyen \u2013 \u201cMimicking Finance\u201d<\/em> (NBER Working Paper 34849, 2026)<\/a> \u2013 provides an intriguing perspective on the extent to which the behaviour of professional investors can be anticipated using modern data-driven methods.<\/p>\n\n\n\n

The authors analyse decades of mutual fund holdings data and train machine learning models to predict the directional trading decisions of portfolio managers \u2013 whether they buy, sell or hold individual securities in the following quarter.<\/p>\n\n\n\n

Specifically, the study employs a Long Short-Term Memory (LSTM) neural network, a class of recurrent neural networks designed to capture temporal patterns in sequential data. Using this model trained on historical trading behaviour and a wide set of market, fund and firm characteristics, the authors show that approximately 71% of mutual fund managers trading decisions can be predicted.<\/p>\n\n\n\n

At first glance, this finding may appear unsettling for the active management industry. However, the more interesting implication lies not in the predictability itself, but in the economic meaning of the remaining, unpredictable portion of decisions.<\/p>\n\n\n\n

Predictability as a Measure of Routine Behaviour<\/h3>\n\n\n\n

The paper proposes a useful conceptual distinction between routine and non-routine components of portfolio management.<\/p>\n\n\n\n

Routine behaviour includes decisions that follow recognisable patterns:
portfolio rebalancing, style-consistent adjustments, responses to flows, or factor-driven allocation shifts. These actions tend to repeat over time and are therefore detectable in historical data.<\/p>\n\n\n\n

Machine learning models excel precisely at identifying such regularities.<\/p>\n\n\n\n

Non-routine behaviour, by contrast, reflects judgement, interpretation and strategic decision-making. These decisions are less systematic and therefore more difficult to predict.<\/p>\n\n\n\n

From a technological perspective, this distinction mirrors a broader pattern seen across many industries:
AI systems are particularly effective at replicating structured and repeatable tasks, while human expertise remains most valuable where interpretation, ambiguity and strategic judgement play a central role.<\/p>\n\n\n\n

Predictability and Investment Performance<\/h3>\n\n\n\n

The study reveals a particularly important empirical relationship. <\/p>\n\n\n\n

\n

Funds whose trading behaviour is more predictable tend to underperform, while those whose actions are harder to predict outperform their peers.<\/p>\n<\/blockquote>\n\n\n\n

The performance spread between the least predictable and most predictable managers is economically meaningful and statistically robust across different risk-adjusted measures.<\/p>\n\n\n\n

Even more interestingly, the outperformance appears to be concentrated in the subset of portfolio decisions that the model fails to predict correctly. In other words, the alpha seems to originate precisely from those trades that do not follow easily identifiable patterns.<\/p>\n\n\n\n

This suggests that much of what machine learning can replicate represents the systematic component of investment processes, whereas the value-creating element may lie in the smaller set of discretionary decisions.<\/p>\n\n\n\n

Implications for the Economics of Asset Management<\/h3>\n\n\n\n

The findings have less to do with replacing human portfolio managers and more with re-evaluating the economic structure of the industry.<\/p>\n\n\n\n

If a large share of portfolio adjustments can be anticipated algorithmically, the justification for active management fees may increasingly depend on the smaller subset of decisions that deviate from predictable patterns.<\/p>\n\n\n\n

Or as Cohen summarises:<\/p>\n\n\n\n

\n

The truly skilled part \u2013 the unpredictable, non-routine component \u2013 is real, but small.<\/p>\n<\/blockquote>\n\n\n\n

From an economic perspective, this shifts the discussion away from the question of whether machines will replace managers, and towards a more nuanced issue: what exactly investors are paying for.<\/p>\n\n\n\n

In this sense, AI may act less as a substitute for investment professionals and more as a tool that helps separate routine portfolio mechanics from genuine skill.<\/p>\n\n\n\n

A Changing Role for Human Expertise<\/h3>\n\n\n\n

Naturally, machine learning models may become even more powerful as data availability and modelling techniques continue to improve. Yet the current implications appear to be more economic than existential.<\/p>\n\n\n\n

Portfolio management will likely continue to combine systematic processes with discretionary judgement. What may change, however, is the relative weight of these components.<\/p>\n\n\n\n

As predictive technologies improve, the industry may gradually move towards a clearer separation between replicable portfolio mechanics and true investment insight.<\/p>\n\n\n\n

For investors, regulators and asset managers alike, this distinction could become increasingly important in defining how value is created \u2013 and how it should be priced \u2013 in the evolving landscape of digital finance.<\/p>\n\n\n\n

Passive, Systematic \u2013 or AI-Augmented Active?<\/h2>\n\n\n\n

One natural question follows from these findings: what happens to the predictable part of active management?<\/p>\n\n\n\n

If a large share of portfolio decisions can be replicated algorithmically, the economic rationale for traditional discretionary management becomes less clear for that part of the investment process. Investors may increasingly ask whether these predictable portfolio adjustments should simply migrate to low-cost passive vehicles such as ETFs or other rules-based strategies.<\/p>\n\n\n\n

However, the outcome may not be purely binary.<\/p>\n\n\n\n

Rather than a full shift from active to passive, a hybrid structure may emerge in which systematic portfolio mechanics are handled algorithmically, while human portfolio managers focus on the smaller subset of non-routine investment decisions.<\/p>\n\n\n\n

In this sense, the future of asset management may increasingly combine:<\/p>\n\n\n\n