Machine Learning in Trading — A Reality Check

Supervised learning has three parts. A feature is an input column, one value per row. A label is the outcome you are asking the model to predict, and it must be computable for past data. Training is the fitting process: the algorithm adjusts its internal numbers until its output on the training rows sits as close as it can to the labels you supplied.

That is all it does. It has no concept of a market, no idea what a stock is, and no interest in whether the relationship it found makes any economic sense. It reduces error on the rows you handed it, using whatever in the columns helps — including things that should never have been in there.

The AI module builds this vocabulary properly, including the difference between a supervised model and a language model, and the three defects that ruin studies. This topic assumes those and asks the systematic-trading question instead: given that machine learning is this and not something grander, where does it actually earn its place in a trading process?

There is a narrow set of jobs where it earns its keep, and none of them is prediction of a price. Note the framing throughout: these are places the tool is appropriate, not places it is claimed to work. Whether any of them produces something tradeable is a question only your own validated study can answer, and for most people the answer will be no.

The model class is a small decision. The columns are the whole study, because a model can only find structure that is present in what you gave it.

A feature is not merely a measurement. It is a claim: 'this quantity, computed this way, at this moment, plausibly carries information about the outcome I defined.' Every column should be traceable to a sentence of that form, written down before the column exists. Columns added because they were easy to compute are how a feature set becomes a search over noise.

Two studies with the same universe, the same features and the same model can reach opposite conclusions because they labelled the outcome differently. The label is not a formality applied at the end. It is the question, and the question decides the answer.

A return over the next twenty candles is a hard label: it is dominated by noise, one enormous move can carry an entire study, and the model spends its capacity fitting outliers. A categorical label is usually more honest, because it asks something the data can answer cleanly.

Consider a label defined as: from the signal candle, did price first close above a level a fixed distance above entry, or first close below a level a fixed distance below it, within a fixed number of candles — and if neither, the outcome is recorded as neither. That definition has three parameters hidden inside it. Move the upper distance and the balance of the classes changes. Move the horizon and a different set of situations resolves. Each choice produces a different question, and therefore a different study.

This is the part that separates market work from the domains where machine learning looked miraculous. Signal is the portion of a future move genuinely related to something knowable now. Noise is everything else — a fund rebalancing on a schedule you cannot see, order flow, a headline nobody could have anticipated.

In handwriting recognition, almost all the information needed is in the image. In market data the noise dominates heavily, and weak signal has a direct consequence: you need far more independent examples to separate a real relationship from a coincidence. Market data supplies far fewer independent examples than the row count suggests, because consecutive rows on the same stock share most of their story.

Supervised learning rests on an assumption that is invisible in most fields because it is simply true there: the relationship fitted on past data still holds on future data. Markets violate it routinely. Volatility clusters and subsides. Leadership rotates. A relationship that held across one long stretch inverts. A change in market mechanics quietly rewrites the thing your model spent its training fitting.

And unlike a handwritten digit, the subject responds to being studied. A market is made of participants, some with better data and lower costs than you, and a simple findable pattern that pays reliably gets traded until it stops paying. The relationship is adversarial, and adversarial relationships decay — which is why the previous topic in this module exists, and why a market model is never finished.

The operational consequence is uncomfortable: the shelf life is real, it is unknown in advance, and nothing errors when it expires. The code runs exactly as it always did and the results simply get worse.

Standard k-fold cross-validation shuffles the rows, splits them into folds, and rotates which fold is held out. It is correct, and it is correct because it assumes the rows are independent and interchangeable. Market rows are neither.

Shuffle a time series and you fit on later blocks while scoring on earlier ones — the model is told the future and graded on the past. Worse, a label that resolves over the following candles overlaps in time with rows placed in the test fold, so information leaks across the boundary even when the split looks clean.

A relationship that exists gross and vanishes net is not a finding. The frequency of the model's decisions is what decides whether costs are a rounding error or the entire story, and a machine-learning model producing a score for every name every day can be pushed into very high turnover without anyone noticing.

In most applications, a model that performs is enough — nobody demands an explanation from a spam filter. In systematic trading the explanation is load-bearing, and the reason is the previous topic in this module.

When a strategy enters a bad stretch, you have to decide whether the conditions it depends on are currently present. For a rule you can state in one sentence, that is a checkable question: you named the conditions when you wrote the rule. For a fitted model with hundreds of interacting parameters, there may be no sentence to check. You cannot say what the method needs, so you cannot ask whether it is there, so the only evidence available is the recent result — which is precisely the evidence you already know not to act on.

That is a real cost and it belongs in the decision about whether to use a model at all. A slightly weaker approach whose behaviour you can articulate is often the more usable one, because it survives contact with a drawdown. The opaque one leaves you with nothing to do but watch and hope, and hoping is not a maintenance procedure.

Not because the model was weak. The models are commodities now — the same libraries a fund uses are a free download, and the difference in fitted quality between a well-known model class and a clever one is small next to everything else in this article.

It fails because the problem was framed in a way that could not have worked. The label was never defined from a rulebook, so the study answered a question nobody asked. A feature knew the future by a few hours. The universe was assembled this morning. The validation shuffled rows that carry each other inside them. Costs were subtracted at the end, or not at all. The variant kept was the survivor of a search nobody counted. Each of those is fatal on its own, and the study will not error on any of them — it will produce a number, in the right format, with a chart attached.

The honest summary is that machine learning does not lower the bar for a trading idea. It raises it, because it adds several new ways to be confidently wrong on top of every way that already existed. If you cannot state your hypothesis in one sentence that names a cause, a model will not supply the sentence for you. It will only fit whatever you gave it, and report back with complete confidence.

State the kill conditions before running anything, in the same document as the hypothesis.