Where the Delegation Decision Still Sits With You

The short answer

We are not going to tell you that human judgement beats a model. We looked for evidence of that and did not find it, and the thing we did find does not amount to it.

What the evidence supports is narrower and more useful: measured model performance is middling rather than decisive, it degrades on data the model has not seen, and a regulator considers overreliance on these systems a risk worth naming. Together those say the decision about how much to delegate has not been settled for you. It remains yours, and it is worth making deliberately.

The trap this article was nearly written into

An earlier plan for this page carried the title "Where Human Judgement Still Beats a Model". We cut it, and the reason is a good illustration of how easy the error is.

The research we have read shows an absence of demonstrated model superiority. That is not the same as evidence of human superiority. The first is a statement about what has been measured; the second is a claim about a comparison nobody in our sources ran. Treating one as the other is exactly the inversion our research process exists to catch, and it would have been easy to write without noticing.

So this page makes the smaller claim, which has the advantage of being true.

What the evidence actually says

Mokhtari, Yen and Liu applied machine learning to price data and to public tweets. Their conclusion:

"The results show a median performance, implying that with the current technology of AI, it is too soon to claim AI can beat the stock markets."

One study, from 2021, attributed as such. Separately, a 2023 benchmark evaluated fifteen state-of-the-art deep-learning models under a single common framework and found that all of them "exhibit a significant performance drop when exposed to new data".

What the published research shows goes through both in detail, including why this page carries no accuracy percentage.

Neither result says models are useless. Both say the confident version of the claim is not yet supported — and the second says something sharper about what happens when conditions change.

The regulator's angle

ESMA's 2024 guidance to firms lists the risks it associates with AI in investment services:

"Algorithmic biases and data quality issues; Opaque decision-making by a firm's staff members; Overreliance on AI by both firms and clients for decision-making; and Privacy and security concerns linked to the collection, storage, and processing of the large amount of data needed by AI systems."

Overreliance on AI by both firms and clients is a regulator's phrase, not ours, and it is listed alongside bias and opacity as a risk in itself. Note what it is not: a statement that people decide better than models. It is a statement about the hazard of leaning too hard on one.

The decisions a forecast cannot make

Here is the part that holds regardless of how good models become, because it is not a question of accuracy.

How much to risk. A forecast about an instrument says nothing about how much of your money should be exposed to it. That depends on your obligations, your horizon and what a loss would actually do to you — none of which the model has.

Whether to be in the market at all. "Which way will this move" presupposes you should have a position. That is a prior question, and it is yours.

When the question itself has changed. Models are fitted to conditions that held in their training data. The judgement that circumstances are no longer the ones a tool was built for is the hardest to automate, precisely because the tool keeps producing confident output throughout.

What you will do when it is wrong. Deciding in advance where you are wrong, and what happens then, is a commitment rather than a prediction — the same thing that separates a signal from a tip in how signals are generated.

So how much should you delegate?

We are not going to answer that for you either, and any page that does is overreaching.

What is reasonable to say: the measured evidence does not currently justify handing over the decisions above, the degradation-on-new-data result is a specific reason for caution, and a regulator has named overreliance as a risk. That is a case for keeping the delegation deliberate — knowing what the tool decides, what you decide, and being able to say why.

If you are weighing a particular product, the eight questions turn this into something concrete, and how model-based tools work explains what you are actually delegating to. Nothing here is personal advice; see the risk disclosure.