0%
READ
What actually happens inside a machine-learned trading model — and why nobody can read it back. A written rule is a hypothesis somebody had; a fitted model is a residue of what worked.
The short version
A trading rule used to be a sentence somebody wrote down. Buy when this happens. Today the rule is a set of numbers a computer arrived at by trial and error — and there is no sentence anywhere in it. The firm still decides what to trade and how much to risk. It just cannot say what its own model will do tomorrow.
The model behaves as though it holds views nobody gave it. It can treat the same signal as important on a calm day and ignore it on a violent one, without anything inside it saying so. That flexibility is why these models work. It is also why they cannot be explained.
They are trusted because they performed well on past years held back as a test, not because anyone understood how they think. Models do stop working — a pattern stops holding, or enough people find it and trade it away. With a written rule you can look up the assumption that failed. Here there is nothing to look up.
There is a particular frustration familiar to anyone who has sat on a desk during an unexplained move. The market goes somewhere it had no business going, on no news, in a size that suggests intent. And you cannot work out who.
You run the usual checks. Rebalance calendar — nothing. Month-end — no. Option expiry — clean. Nothing in the flow you can attribute. So you wait. If the selling dries up in a day or two, it was mechanical: somebody with a position to shed, now finished. If it keeps coming, it was a view, and the move has further to run. Those call for opposite responses, and you find out which one you were looking at only after it no longer matters.
That gap between the move and the explanation appears to be widening, and several things are pushing on it at once. This piece is about one of them, and it is narrow: a change in how a trading rule comes into existence in the first place.
Let's start with something completely unmysterious.
You build a large spreadsheet. Each row is one instrument on one day, holding every measurable market parameter about that instrument on that morning — the level, the change over the past week, the shape of the curve, how volatile it has been, what it costs to hold. Broadly the things you would look at yourself, converted into numbers and lined up in a row.
Next to each row you write what actually happened afterwards: the return over the following week. Real number, real week, no cleverness involved.
Twenty years of that is a million rows. Any analyst could assemble it from a terminal in a fortnight, and there is no machine learning anywhere in it. It is a record of what markets looked like and what they returned afterwards.
Machine learning begins where the spreadsheet ends, and what it searches for is weights.
A weight is the importance the model assigns to one input. Give carry a weight of 0.3 and volatility −0.4, and every unit of carry adds 0.3 to the answer while every unit of volatility subtracts 0.4. The weights are the model; everything else is arithmetic. So the question is how much should carry count, and does that change when volatility is high.
Nobody knows. And — this is the part worth sitting with — nobody could look it up either. Carry exists and can be measured; its correct weight does not exist until somebody builds a model that needs one, and then it belongs to that model rather than to the market.
So the search has nothing to start from. The weights begin as arbitrary numbers, which makes the model's first output worthless. That output is a score: its guess at the return that row will earn over the following week. Producing one takes a single pass through the structure — the inputs blended into a first set of intermediate quantities, those into a second, and a final step turning the last set into the score.
The fitting is the loop around that pass. Score a row, compare it against the return recorded beside it, nudge every weight fractionally in the direction that would have made it less wrong. Then the next row, and the next, over the whole table. Several million times.
And there is only one set of weights to find. Any single row has weights that would score it exactly, but those weights would misprice the row beside it; the same numbers have to serve all million at once, so the fitting is a compromise rather than a solution.
Nobody supervises that compromise. Nobody decides that carry should outrank momentum. The weights settle wherever the combined error is smallest — across all million rows together, the gap between the scores and the returns that followed as small as that set of weights can make it. Not right on any particular row. Least wrong overall.
And then they stop. This is the detail most descriptions skip, and it matters: once fitting ends, the weights are frozen. The same numbers on a calm Tuesday and on the worst morning of the decade.
Two things have now been described and they are easily run together. The pass through the structure happens every time the model is used, for every row. The fitting — the search that decided what the weights in that structure should be — happened once, and is over. So when the model runs live, one row goes in, arithmetic happens through the layers, a score comes out — a guessed return for that instrument, on that morning. No more iteration, no searching, no learning. Microseconds.
Which gives us the distinction the rest of this piece rests on: a written rule is a hypothesis somebody had; a fitted model is a residue of what worked.
Fitting alone does not make a thing unreadable. If the answer comes out as three parts carry, two parts momentum, less four parts volatility, you can say it aloud in a breath. Somebody found those numbers by trial rather than theory, but the result is still a sentence, and a model of that shape stays legible however it was arrived at.
The trouble is that such a sentence holds one opinion and holds it forever. Whatever weight it lands on for carry applies on calm days and crisis days alike.
But what every practitioner actually believes is conditional. Carry works when volatility is low and gets you carried out when it isn't. No arrangement of fixed weights can express that, because the weight on carry would need to depend on volatility — and a fixed weight, by definition, depends on nothing.
Quants have worked around this for decades, elegantly: build a new input equal to carry multiplied by volatility, and let the model weight that. It works. Its limitation is that you must know in advance which interactions matter, and with fifty inputs there are over a thousand pairs and tens of thousands of triples. So you specify the ones you suspect — and the model can only find relationships you had already thought of.
Layering is the general solution. Instead of reading the raw inputs and producing an answer, the model builds a set of intermediate quantities, each a blend of the inputs. Then a second set, each a blend of the first. Then a third. Only at the end does a final step produce the score.
Blending alone would achieve nothing. What makes layering work is a rule so simple it sounds like a triviality: if an intermediate quantity comes out negative, replace it with zero. Whatever reads from it that day receives nothing. This is standard machinery rather than anything proprietary, but it is worth watching it work, because the whole argument rests on it.
Take a two-layered model, small enough to check by hand: three inputs feed two intermediate quantities, A and B, which feed one score. The weights and readings below are invented for illustration; the arithmetic that follows from them is not, and can be checked with a calculator.
A = carry − volatility
B = (0.2 × carry) + momentum − (0.3 × volatility)
Score = (0.6 × A) + (0.4 × B)
Six weights, fixed for the life of the model. Note that carry enters A and B with different weights: every intermediate quantity has its own set.
Calm morning: carry 3, momentum 2, volatility 1. A comes out at 2, positive, so it passes through. B comes out at 2.3, passes too. The score is 2.12.
Violent morning: same carry, same momentum, volatility at 5. Now A is −2, so it is bent to zero and passes nothing on. B is 1.1. The score is 0.44.
Now ask what carry is worth on each day. Raise it from 3 to 4 and see how far the score moves.
| Morning | Score at carry 3 | Score at carry 4 | Carry is worth |
|---|---|---|---|
| Calm | 2.12 | 2.80 | 0.68 |
| Violent | 0.44 | 0.52 | 0.08 |
Carry matters nearly nine times more on the calm day. The model is behaving as though it holds the view carry works when volatility is low — and nobody wrote that. There are six numbers and a rule about negatives.
Now delete the bend. Let A stay at −2 on the violent morning instead of being flattened, and run the same test again.
| Morning | Score at carry 3 | Score at carry 4 | Carry is worth |
|---|---|---|---|
| Calm | 2.12 | 2.80 | 0.68 |
| Violent, bend removed | −0.76 | −0.08 | 0.68 |
Look at what has changed and what hasn't. The scores still differ on the two mornings — of course they do, volatility is one of the inputs. That was never the question. The question was whether the importance of carry shifts with conditions, and the last column says it does not: 0.68 on a calm morning, 0.68 on a violent one. The model still responds to volatility; it no longer lets volatility change how much attention it pays to anything else. That is what conditionality means, and without the bend it is gone.
You can see why by substituting A and B into the score:
0.6(c − v) + 0.4(0.2c + m − 0.3v) = 0.68c + 0.4m − 0.72v
The two layers have collapsed into a single line with three fixed weights, and there sits the 0.68, the same on every morning through the model's life.
"Collapsed" is meant literally, and it matters what is identical to what. Not the scores — those move as the inputs move. The recipe is identical, the way multiplying by 2 and then by 3 is identical to multiplying by 6: different answers every time, from a rule that never changed. A stack of weighted sums behaves the same way, however many layers you pile on: the answers vary with the inputs, the relative importance of each input never does. Without the bend, all that structure buys nothing at all. Other model families get their conditionality differently — decision trees split the data rather than bending it — but they arrive at the same place, and everything that follows applies to them too.
Two consequences follow, and they are the whole story.
The pattern of open paths differs for every row on every day. Ask which part of the model is the carry rule and there is no answer. The rule is the pattern, and the pattern is different tomorrow.
So how does anyone know the model learned something sensible rather than a coincidence of the sample?
They don't, in the sense you would want. What they have is one check, and it is statistical rather than explanatory.
Before fitting begins, you cut the history in two. Fit on the first stretch; leave the rest untouched. When fitting finishes, run the model on the years it has never seen and measure whether its scores still show a positive, non-zero correlation with the returns that followed.
The test sifts patterns from accidents. The table holds only so many rows, and a model with enough weights can account for quirks in those particular rows rather than finding anything general — a stretch of 2011 where high carry happened to precede a rally, say. Noise, not structure, and it almost never repeats on the untouched years, so the correlation collapses. What survives tends to be conditional structure of the kind the bend allows, because that reflects how markets behave rather than which rows were in the sample.
But notice what the check delivers. It tells you that whatever the model learned generalises. It does not tell you what the model learned. You can go further and probe the thing from outside — hold volatility high, sweep carry across a range, watch the score respond — but that is an experiment of exactly the kind you would run on a market, and it only answers the questions you thought to ask. Either way you have verified the output without inspecting the reasoning.
None of this is for want of effort. A substantial field has grown up around making these models explainable, and its methods — attribution techniques that assign each input a share of a given prediction — are standard enough that the CFA Institute has published guidance on them.¹ But look at what they deliver: an account of what mattered on a particular row, after the fact. You learn what the model did. You still cannot say what it will do.
Which leaves the industry in an unusual position. These models are trusted because they worked on data they had never seen. Nobody confirmed that they learned something sensible; they confirmed that whatever was learned kept working.
That is a defensible way to run money, and the design around the model exists to make a weak, unreadable signal usable:
All of which follows from how thin the edge is. Signal quality is measured as the correlation between a model's scores and the returns that follow — the information coefficient — and for anything predicting the direction of returns, a working signal lives somewhere between 0.01 and 0.06.² A coin landing right barely more than half the time. Anything approaching 0.10 is treated less as a triumph than as a warning that something has been fitted too closely to the past. And the models described in this piece tend to sit at the thinner end rather than the thicker, at coefficients small enough that transaction costs can consume them entirely. Worth nothing on a single trade. Spread across thousands and refreshed continually, it becomes a business.
None of that scaffolding is opaque. It is written down, and it exists precisely because everyone involved knows how little the model inside it actually knows.
But it remains a different position from we wrote down a rule we believed in and tested whether it held. When a written rule stops working, you can go back to the sentence, find the assumption that no longer holds, and know what broke. When a fitted model stops working, there is no sentence to return to — only a set of weights that were never a claim about anything, and a record of the days on which they used to do better.
It would be convenient to present this as an overlooked problem. It isn't, and the argument is still being had. Supervisors have been circling it for years:
| Body and date | What it covers |
|---|---|
| IOSCO, 2021³ | Report on AI and machine learning at market intermediaries and asset managers. Lists transparency and explainability among its core concerns. |
| FSB, Nov 2024⁴ | Financial stability implications of AI. Names limited explainability directly as something that can raise model risk at firms without strong governance. |
| IOSCO, Mar 2025⁵ | Consultation on AI in capital markets. Returns to the same concerns. |
| FSB, Oct 2025⁶ | Work on the data gaps that make AI adoption hard to monitor at all. |
| IOSCO, May 2026⁷ | Supervisory toolkit, produced by an AI working group whose members include the Monetary Authority of Singapore. |
The sharpest version came from the CFA Institute, responding to the IOSCO consultation in April 2025⁸: the report framed the risks too narrowly around large language models, when model complexity is fundamental to all deep learning models. Their own survey work ranked the black box effect as the second largest organisational and governance risk, behind only data privacy.⁹ Not a property of chatbots, then, but of fitted parameters arranged in layers.
Two things cut against a tidy conclusion.
The first is that the headline adoption figures do not measure what they appear to. Surveys reporting that most managers now use AI are generally counting generative tools somewhere in the working day, which is a different thing from a machine-learned trading signal. And the CFA Institute's letter notes what the percentages miss: the data, teams and capital required mean only a few firms can run these methods at the frontier. Concentration, rather than uniform adoption.
The second is that the traffic runs in both directions, and this should give the argument some pause.
In February 2026, Lauren Cohen, Yiwen Lu and Quoc Nguyen published work through the National Bureau of Economic Research¹⁰ that pointed a model of exactly this kind at active fund managers rather than at markets. Same table, different question: what sat beside each row was the direction a named manager subsequently took in a given stock — buy, sell or hold.
Fitted on rolling five-year windows from 1990 to 2023, then asked to call quarters it had not been shown, it got the direction right 71% of the time. It was not watching the trades and inferring the rest; it worked from public information available before the fact. And the managers whose trades were hardest to call were the ones who outperformed.
So the same technology that removes readability from a written rulebook is being used to add it back somewhere else. A systematic rule that used to be inspectable becomes a fitted model nobody can read; a discretionary manager whose thinking was never inspectable becomes largely predictable from the outside. Whether the market as a whole is getting harder or easier to anticipate is an open question, and not one this piece can settle.
The mechanism described above is not in dispute; it is arithmetic. What it adds up to across a whole market is a question supervisors are still framing, researchers are still testing, and this section will keep returning to.
Which returns us to the move nobody could attribute.
The flow that did it was not mysterious. It was mechanical, unsentimental, indifferent to the news, and would have traded the same way whether or not anybody was watching — no different, in that respect, from the rebalance you can compute at two in the afternoon.
The only difference is that its rule was never written in a form anyone can read, including the desk that ran it. Ask them what they will do if the market gaps twenty basis points overnight and they can tell you what happened on the handful of historical days that vaguely resembled it. A range drawn from the past, rather than a statement about the future.
That is the same answer you would have given. They just have better data.
The worked example in "Watching the Bend Work" uses invented weights and readings, chosen to make the mechanism visible. The arithmetic is reproducible from the figures given.
Next in this section: what happens when a great many firms fit similar models to the same few decades of market history. Last Friday of October.