AI and machine learning in quant research, and how to assess it

AI and machine learning are powerful research tools — but they are tools, not autonomous decision-makers, and never a guarantee of returns. This note explains where these techniques genuinely help in quantitative research, the risks that must be governed, and what an allocator should establish before committing to a manager that relies on them.

Manuel Ernesto De Luque MuntanerFounder & CEO

18 June 20269 min read

Key takeaways

AI and machine learning are research tools for finding and validating patterns in data — not autonomous investors and not a promise of performance.
Their value in quant research lies in signal development, feature engineering and model validation within a disciplined framework.
The central risks are overfitting, poor or biased data, non-stationary markets and opacity — every one of which requires governance.
Responsible use keeps humans in the loop and keeps risk management independent of the models themselves.
Where a systematic manager relies on these techniques, what should decide an allocation is the governance around them — not the sophistication claimed for them.

What these techniques actually do

In quantitative research, machine learning is, at its core, pattern recognition at scale: the ability to detect relationships in large, complex datasets that are hard to specify by hand. Applied well, it helps researchers generate and refine hypotheses, process information that would overwhelm manual analysis, and test the robustness of ideas.

What it does not do is remove the need for judgement, economic reasoning or risk management. A model can find a pattern; it cannot know whether that pattern is meaningful, durable or safe to trade. Treating these techniques as tools within a disciplined process — rather than oracles — is the difference between research and speculation dressed up as science.

Where it genuinely helps

Used within a defined framework, AI and machine learning contribute across several stages of the research process.

Signal development — surfacing candidate relationships in data that can be investigated, understood and tested rather than trusted blindly.
Feature engineering — constructing and selecting informative inputs from raw data more systematically than by hand.
Processing complex data — extracting structure from large, noisy or unconventional datasets at a scale manual analysis cannot reach.
Regime and pattern detection — helping identify shifts in market behaviour that may warrant a change in exposure or approach.
Validation and robustness testing — stress-testing ideas across periods and conditions to distinguish genuine effects from chance.

The risks that must be governed

The same power that makes these techniques useful makes them dangerous when applied without discipline. The failure modes are well known — and each is a governance problem before it is a technical one.

Overfitting — powerful models can fit noise as easily as signal, producing patterns that look compelling in history but do not persist.
Data quality and bias — conclusions are only as good as the data; gaps, errors and biases propagate straight into the model.
Non-stationarity — markets change, so a relationship that held in the past may not hold in the future, however well it was learned.
Opacity — complex models can be hard to interpret, which makes their risks harder to understand and to govern.
Over-reliance — automation can create a false sense of certainty; a model's confidence is not the same as being right.

Using it responsibly

Responsible application is defined less by the sophistication of the models than by the discipline of the framework around them. A few principles matter more than any single technique.

Rigorous out-of-sample testing and deliberate guards against overfitting before anything is relied upon.
Human oversight of models — they inform decisions and are governed by people, not left to run unchecked.
Risk management kept independent of the models, so a model failure does not also disable the controls meant to contain it.
Economic reasoning alongside statistical evidence — a pattern should make sense, not merely appear.
No promises of performance — these are research tools, and no technique guarantees a result.

What we establish before allocating

We do not build these models. The systematic strategies we oversee are implemented using signals and systems produced by an external provider founded by a partner of Block Asset Management — a related party rather than an arm's-length third party. The partner concerned sits on no investment committee and takes no part in the decision to use the provider or to deploy a strategy that relies on it; the arrangement is set out in full on our governance page. Our work is to select that provider, diligence how it operates and supervise it thereafter. Where machine learning forms part of a provider's research process, our questions are about governance rather than technique.

Block Asset Management

How Block Asset Management helps

Our advantage is not the use of AI and machine learning — we do not develop these models. It is the diligence we apply to a provider that does, and the discipline of the framework that governs the allocation thereafter.

Governance before technique

Where a provider's research relies on these techniques, we assess the framework that governs them before we assess the results they produce.

Human oversight, not autonomy

Models inform decisions and remain supervised by experienced people. A strategy is never left to run unchecked, and the authority to reduce or withdraw it stays with the investment committee.

Validation and guards against overfitting

We look for evidence that signals and models were tested for robustness across periods and conditions, with deliberate protection against fitting noise rather than signal — and we ask what failed those tests.

Risk controls apart from the models

We assess whether risk controls sit apart from the models they oversee, so that a model's failure cannot also disable the discipline meant to contain it.

Concentrated on liquid markets

The systematic strategies we oversee operate in liquid markets, where data is richer and execution and risk can be managed with discipline.

Experience and transparency

Eight years focused on digital assets inform how we assess and govern the strategies we oversee, with the transparency and controlled, auditable access institutions expect.

AI and machine learning have real value in quantitative research — but that value is realised only within a disciplined framework that governs their risks and keeps human judgement and risk management firmly in place. The technique is not the edge; the discipline around it is.

If your organisation is evaluating systematic digital asset strategies, our investor relations team can discuss how we select and oversee them.

Important information

This material is provided for information purposes only and is intended for professional and qualified investors. It does not constitute investment advice, an offer or a solicitation to buy or sell any financial instrument, nor a recommendation of any strategy. Digital assets are volatile and involve significant risk, including the possible loss of the entire amount invested. AI and machine learning techniques do not remove risk and do not guarantee results. Past performance is not a reliable indicator of future results. Nothing in this note should be relied upon as a promise or representation as to future performance.

Related research

Continue reading BAM research

This note is part of Block Asset Management's research on institutional digital asset investing. Explore the wider library, or read how we assess managers and structures before any allocation is made.

Browse all research Our investment approach