What Machine Learning Can and Can't Do in Discretionary Trading

By Simon Cotterill, Founder & Lead Portfolio Manager, Vadantia

Machine learning has become one of the most over-promised ideas in markets. It is described, often by people selling something, as a way to predict prices, remove human error, or turn uncertainty into a solved problem. None of that is true, and pretending otherwise tends to end badly.

It is also genuinely useful — when you are honest about where its usefulness begins and ends. At Vadantia, machine learning sits inside a discretionary process, not in place of one. The most important design decision we made was not which model to use, but where to draw the line between what the machine does and what a person does. This post is about that line: what the technology can carry, and what it cannot.

What it can do

Test ideas against history with discipline. A human looks at a chart and sees a pattern. The honest question is always: how often has this configuration actually preceded the outcome I think it implies, and under what conditions did it fail? Answering that consistently, across years of data, without quietly forgetting the times it didn't work, is something people are bad at and machines are good at. Used this way, machine learning is less a crystal ball than a memory that refuses to flatter you.

Surface structure that isn't obvious. Markets contain relationships that aren't linear and aren't visible to the eye — interactions between volatility, liquidity, positioning and time that don't show up in a single indicator. Statistical and machine-learning methods are good at detecting that something is present in the data without you having to specify its exact shape in advance. That is a real edge, provided you treat what comes out as evidence to be weighed, not an instruction to be followed.

Express uncertainty honestly. A well-built model doesn't just produce an answer; it produces a sense of how confident that answer deserves to be, and flags when current conditions look unlike anything it was trained on. Calibrated uncertainty is, in our view, more valuable than any single output. Knowing when the system doesn't know is what keeps you out of the worst situations.

Stay consistent. A model applies the same standard at 7am on a quiet Tuesday and in the middle of a violent session. It doesn't get bored, anchored to a losing position, or seduced by a story. That consistency is a genuine contribution to a process — not because it is smarter than a person, but because it is steadier than a person on their worst day.

What it can't do

Predict the future. This is the claim worth retiring permanently. A model estimates conditional structure from the past; it does not foresee what hasn't happened. Markets are shaped by human behaviour, policy, and events that have no precedent in any training set. Anyone offering forecasting as a product is, at best, describing probabilities in confident language.

Understand a regime it has never seen. Models learn from history, and history is not evenly informative. A system trained largely through one kind of market can be quietly miscalibrated the moment conditions change in a way it hasn't experienced. Recognising that the world has shifted — before the numbers catch up — remains a human judgement. The machine tells you what usually happened; it cannot tell you that this time is genuinely different.

Supply conviction, or own the decision. A probability is not a decision. Sizing a position, choosing to stand aside, deciding that the evidence is real but the risk-reward isn't worth it — these require judgement and accountability that no model carries. The model can inform the call. It cannot be the one held responsible for it.

Manage its own risk. Left alone, an optimisation process will happily pursue an objective straight off a cliff. The discipline that decides how much to risk, when to stop, and what "wrong" looks like is imposed from outside the model, by design and by a person. We treat that boundary as non-negotiable.

The honest division of labour

Put simply: the machine generates and validates evidence; the human weighs it, decides, and is answerable for the outcome. We log how each decision was reached, so the process can be reviewed rather than just remembered. The technology earns more influence only as it demonstrably helps — never by default, and never to the point of taking the final call.

This is a less exciting story than "the AI does it for you." It is also the only version we'd be willing to put our name to. Discretion informed by good evidence, with clear lines of responsibility, is more robust than either pure intuition or a black box — precisely because it keeps a thinking person in the loop, with the humility to know what the tools can and cannot tell them.

The firms that get into trouble are usually the ones that forgot where the line was. The work, for us, is in keeping it exactly where it belongs.

For professional and institutional investors only. This article is for information purposes only and does not constitute investment, legal or tax advice, a personal recommendation, an offer or solicitation, or a financial promotion. It does not describe or recommend any specific transaction. Vadantia Quant FX involves leveraged products; capital is at risk and you may lose more than the amount invested. Past performance is not a reliable indicator of future results. Vadantia FZCO acts as strategy provider and investment adviser to a regulated asset manager and does not guarantee any return.