Machine learning mostly speeds up how quant funds find and trade signals rather than handing anyone a lasting edge. In markets that keep changing, the hard problem is signal decay, not model power.
Key takeaways
- Machine learning has changed how systematic funds work, mostly by speeding up signal discovery and cutting execution costs. It has not delivered a durable, magic edge.
- The moat has moved from the model to the data. Core ML methods are commoditised; exclusive, clean, early data is not. Alt-data adoption hit a record 90% of managers in 2025.
- In non-stationary markets the binding constraint is signal decay, not model power. Some estimates put signal half-lives at roughly 18 months now, down from five to seven years pre-AI.
- For an allocator, “we use AI” is a description, not a differentiator. The real questions are about data, retraining speed, capacity, interpretability, and crowding.
Ask most people what machine learning quant funds have really changed about systematic investing and you get one of two answers. Either the machines have quietly taken over and the returns are now magic, or it is all hype and nothing has really changed. Both are wrong, and the gap between them is where a serious investor can lose money by allocating to the wrong story.
Here is the position worth defending. Machine learning has changed how systematic funds do their work, but it has not changed the thing that decides who wins. ML mostly makes signal discovery faster and execution cheaper. It does not hand anyone a durable edge that the rest of the market cannot compete away. In markets that keep shifting under your feet, the hard problem is not model power, it is signal decay. So the managers pulling ahead are the ones with the best data and the tightest risk control, not the ones running the largest neural networks. If you are weighing an allocation to a fund that leads with its AI, that distinction is the whole game.
This connects directly to how hedge funds generate returns more broadly, so it is worth reading alongside that guide.
What a quant fund does with machine learning
A quant fund, short for quantitative fund, is one that makes buy and sell decisions from models and data rather than from a manager’s judgement. Systematic investing is the same idea seen from the other side: rules and code decide the trades, and a human sets the rules rather than picking the stocks.
Machine learning is a subset of that toolkit. Where an older quant model might use a fixed formula a researcher wrote by hand, an ML model learns the relationship from the data itself, adjusting as new data arrives. That is the real shift. It is a change in how signals get found and refined, not a change in what a fund is trying to do, which is still to buy things that will go up and sell things that will go down, at a scale and speed a human cannot match.
It helps to split the work into four stages and ask where ML earns its keep in each: the data, the signal, the execution, and the risk. Those four cover the entire pipeline, and ML behaves very differently across them.
Where ML adds value across the quant pipeline, and where it disappoints
Rather than treat “AI in quant funds” as one thing, look at it stage by stage. The table below is our read of where machine learning genuinely moves the needle and where it tends to let managers down, assembled from how the leading systematic firms describe their own stacks and from the regulatory surveys of what funds deploy.
| Pipeline stage | What ML does well here | Where it disappoints | Net read |
|---|---|---|---|
| Data (sourcing, cleaning, structuring) | Parses messy alternative data at scale: satellite imagery, card transactions, web-scraped pricing, shipping records. Turns unstructured text and images into usable inputs | Cannot fix a biased or discontinued dataset. Vendor changes silently break inputs. Rubbish in, rubbish out, faster | High value. The biggest, most reliable win |
| Signal (finding predictive patterns) | Speeds up the search across thousands of candidate features. Captures non-linear relationships a linear model misses | Overfits readily. Finds patterns that vanish out of sample. The edge it finds is the edge others find too, so it decays | Mixed. Faster discovery, not a durable moat |
| Execution (getting the trade done) | Predicts short-term liquidity and market impact, times child orders, cuts slippage. A quiet, compounding cost saving | Little here. The gains are real but bounded by market microstructure, not by model cleverness | High value, low drama. Underrated |
| Risk (sizing, limits, drawdown control) | Flags regime change and rising correlation earlier than static rules. Stress-tests portfolios against learned scenarios | A model trained on the past cannot see a genuinely new shock. False comfort is the danger | Useful but never a substitute for judgement |
The pattern is worth sitting with. Machine learning is at its strongest at the two ends of the pipeline, the plumbing and the execution, where the work is unglamorous and the gains compound quietly. It is weakest in the middle, in the signal-finding step that the marketing decks are built around. That inversion is the reader’s wrong assumption corrected in one table: the AI is doing its best work in the places nobody puts on a pitch deck.
The edge has moved from models to data and systems
For years a quant fund’s advantage came from having a model nobody else had. That is much harder to hold now. The core ML methods, gradient boosting, random forests, neural networks, are documented, taught, and available as open-source libraries. A talented team at a mid-sized fund can build the same model architectures as the giants. So the model itself has stopped being the moat.
What has not commoditised is the data feeding the model and the systems around it. Two funds running the same algorithm on different data will get different results, and the fund with cleaner, earlier, more exclusive data wins. This is why the money has followed the data.
The alternative data market, meaning the non-traditional inputs funds buy to feed these models, was estimated at roughly $18.8 billion in 2025 and is forecast to grow at around 37% a year through 2033 on some readings (Grand View Research). Adoption among investment managers has become close to universal: Lowenstein Sandler’s 2025 survey put alt-data use at a record 90% of respondents, up from 62% in 2023, with 89% planning to grow their budgets and 96% planning to increase AI budgets in 2026 (Lowenstein Sandler, 2025). Serious spend runs to real money. Morgan Stanley has benchmarked a committed hedge fund at roughly $1m of alt-data spend per $1bn of assets in year one, rising toward $3m by year three (cited in Neudata).
Read those numbers together and the story is not “funds are buying AI”. It is “funds are buying data, and AI is how they turn it into something usable”. Web-scraped data now shows up in 56% of use cases, and 59% of advisers feed it into custom AI systems (Lowenstein Sandler). The model is the cheap part. The data pipeline is the expensive, defensible part.
The hard problem is signal decay, not model power
Now the part that the AI framing tends to hide. Financial markets are non-stationary, which means the statistical relationships in the data keep changing. A pattern that predicted returns last year can weaken or reverse this year, because other funds find the same pattern and trade it away, or because the underlying behaviour of the market shifts. The Bank for International Settlements has made exactly this point in its own research: conventional machine learning approaches, which lean on historical correlations, “often struggle to generalise when the economic environment changes” (BIS Working Paper 1291, 2025).
That is a very different problem from the one the technology is good at. A more powerful model does not help you when the thing you are modelling has moved. This is signal decay, and it is the real constraint on returns.
The scale of it is now being measured. One line of research estimates annual alpha-decay costs at around 5.6% in US markets and 9.9% in European markets, as edges erode (microalphas). More strikingly, work on AI-driven crowding argues that as adoption rises, the average signal’s useful life, its half-life, has compressed from roughly five to seven years in the pre-AI era to something closer to 18 months today (arXiv, 2026). The better everyone’s models get, the faster everyone’s edges die. That is the paradox at the centre of ML in quant funds: the same tools that find a signal faster also kill it faster, because your competitors are running them too.
So the winning behaviour is not building a cleverer model once. It is refreshing signals faster than they decay, and having enough of them that the portfolio does not depend on any single one. Research on this points the same way: more frequent model updates, such as three-month rolling retraining, consistently beat static approaches in non-stationary conditions (arXiv, 2026). The edge is in the treadmill, not the trophy.
What the leaders show, and what they don’t
The track records here are genuinely impressive and worth reading carefully, because they are often used to prove more than they do. Renaissance Technologies’ internal Medallion fund, the most famous machine-driven strategy of all, has reportedly compounded at roughly 39% a year after fees over decades, and its firm-wide funds led quant gains in 2024 (Hedgeweek). In the first half of 2025, the machine-learning fund Voleon Composite rose 12.8%, and Two Sigma’s Spectrum returned 7.6% (Hedgeweek). Two Sigma manages around $60bn; Cubist, the systematic arm of Point72, runs computer-driven strategies across liquid asset classes and describes its own edge as sitting in data access and research, not a single model (Point72).
Two things about that list matter more than the returns. First, Medallion’s returns come with a hard constraint most investors never see: the strategy has limited capacity, which is why Renaissance caps its internal capital at roughly $10bn to $12bn rather than raising the tens of billions it easily could. A brilliant signal that only works on small size is not a business you can scale. Second, the firms themselves describe their advantage in terms of data, research velocity and execution, not model wizardry. They are telling you where the edge really is. It is worth listening.
For a fuller picture of the biggest names in this space, our guide to the top quant hedge funds covers the leading systematic managers, and the largest multi-strategy hedge funds ranks the platforms where much of this quant capital now sits.
What the regulators are watching
The regulatory read matters because it tells you how far and how fast this has actually spread, stripped of marketing. The Bank of England and FCA’s 2024 survey found 75% of UK financial firms already using AI, up from 58% in 2022, with a further 10% planning to adopt within three years (Bank of England, 2024). The uncomfortable finding sits alongside it: 46% of firms reported only a “partial understanding” of the AI they use, largely because they rely on third-party models they cannot fully see inside. The risks the survey flagged as growing fastest were third-party dependency, model complexity, and hidden or embedded models.
That is the interpretability problem, and for an investor it is a due-diligence question, not an abstract one. A manager who cannot explain why a model took a position cannot tell you how it will behave in a regime it has never seen. Regulators are circling this precisely because a market full of similar, opaque, correlated models is a new kind of systemic risk. When everyone’s black box learns from the same data and reaches for the same trade, crowded exits get more violent. The 2007 quant quake, when a wave of similar systematic funds unwound at once, is the historical version of this risk. ML makes the models better and the crowding worse at the same time.
How to think about it as an investor
None of this is advice about what you should buy, and none of it should be read as a promise about returns. It is a lens for judging the AI story when a fund puts one in front of you.
Treat “we use AI and machine learning” as a description of method, not a source of edge. Almost every serious systematic fund uses it now, so on its own it is not a reason to allocate. The questions that separate the managers are downstream of the buzzword. What data do they have that others do not, and is it exclusive or just widely bought? How quickly do they retrain, given signals now decay in months rather than years? What is the strategy’s capacity, and are they raising past the point where the edge survives? Can they explain, in plain terms, what their models do and how they behave when markets break? And how do they control crowding risk, the chance that their positions are everyone else’s positions?
A fund that answers those well is doing the real work. A fund that answers them by pointing at the size of its models is selling you the part that has already been commoditised.
FAQs
Does machine learning give quant funds a permanent edge?
No. The core ML methods are widely available, so the model itself is rarely the advantage. What lasts is exclusive data and disciplined risk control, both of which sit around the model rather than inside it.
What is signal decay, and why does AI make it worse?
Signal decay is the tendency of a profitable pattern to weaken as other funds find and trade it. AI accelerates this because competitors run similar tools on similar data, so edges get discovered and competed away faster. Some research now puts the average signal’s half-life near 18 months.
Is alternative data the same as AI?
No, but they work together. Alternative data is the non-traditional input, such as satellite images or card transactions. Machine learning is what turns that messy data into a usable signal. Funds are spending heavily on both; alt-data adoption reached 90% of surveyed managers in 2025.
Why do the best machine-learning funds stay small?
Many of the strongest signals only work on limited capital. Renaissance caps its internal Medallion capital at around $10bn to $12bn for this reason. Capacity, not model quality, is often the real ceiling on a strategy’s returns.
What should you ask a fund that leads with its AI?
Ask what data they hold that others do not, how often they retrain, what the strategy’s capacity is, whether they can explain how their models behave in a crisis, and how they manage crowding. The answers, not the AI label, tell you whether there is an edge.
Next read
If you are evaluating this part of the market, start with the hedge funds guide for how these strategies fit the wider landscape, then the top quant hedge funds for the specific managers running them.