AI expert

How accurate is our AI expert?

AI-generated. Every signal counted on this page was produced by software — YOLO’s in-house AI expert — with no human picking the trades. This page is a record of how well its stated confidence has matched reality. It is not financial advice, it is not a forecast, and it is not a promise of profit. Trading puts your capital at risk and you can lose money.

When our AI expert publishes a trade signal it also says how confident it is, in plain words rather than a falsely precise score — a confidence band such as Strong, Moderate or Weak, and sometimes no claim at all. The only honest way to judge that is to check it: for every band, how often did the thing it predicted actually happen? That is calibration, and it is what this page publishes. A well-calibrated AI that says “58%” should be right about 58% of the time — no more, and no less.

Wins and losses below are shown side by side, in the same size and the same colour, because a track record that flatters its wins is not a track record. Nothing here is entered by hand: the numbers are computed by the platform from signals it recorded and scored itself.

Public

Public — these signals are published to followers who subscribe to the AI expert. Every one still needs your own approval before anything is placed.

Snapshot as of 09 Sept 2026, 03:00 — a trailing-window record, recomputed on a schedule. It does not move minute by minute, so nothing on this page pulses or ticks.

What we said, and what happened

Measured over the shadow window, from 970 scored signals out of 1,121 recorded.

Strong: we said 69.1%, it happened 40.4% (n=52, the shadow window)
Moderate: we said 56.0%, it happened 65.8% (n=79, the shadow window)
Weak: we said 28.8%, it happened 32.2% (n=460, the shadow window)
Unproven: we published no likelihood for this band, so there is nothing to check against. Its signals are still counted: n=379, of which 125wins and 254losses.

What we have left out, and why

Every signal in this record is AI-generated. Capital is at risk and past results are not a promise of future returns.
Losses are published with the same prominence as wins — the wins and losses behind every band are shown side by side.
A confidence band with fewer than 20 scored signals is withheld rather than shown as a percentage — a handful of outcomes is noise, not a hit rate.
151 of 1121 recorded signal(s) have not yet reached their d7 checkpoint and are excluded from every rate below. An unscored signal is never counted as a win.
The “Unproven” band publishes no probability at all, so it has no predicted rate — only what actually happened to the signals we explicitly labelled unproven.
Conformal coverage shortfall: the 80% prediction bands covered only 63% of outcomes over 60 scored signals. Treat the stated confidence as optimistic by roughly that margin. The auto-demotion floor is currently set to 50% rather than the registered 70%, so this shortfall is not currently halting publication.

Happened
Said

Each pair is one confidence band: what the AI expert said would happen (amber) against what actually happened (dark). Equal-height bars mean the stated confidence was honest. Bands without enough matured signals are not drawn — they are listed as withheld in the table below. Neither bar is a gain or a loss, so neither is coloured green or red.

Every band, in full

Wins and losses are shown in the same size and the same colour, deliberately. A band with too few finished signals is marked withheld rather than given a flattering number.
Confidence band	We said	It happened	Signals	Outcomes	How the odds were set
Strong	69.1%	40.4%	52	21wins31losses	Fitted to past outcomes (Platt)
Moderate	56.0%	65.8%	79	52wins27losses	Straight from the model — not yet corrected against outcomes
Weak	28.8%	32.2%	460	148wins312losses	Straight from the model — not yet corrected against outcomes
Unproven	No claim made	33.0%	379	125wins254losses	No correction applied

Did our ranges hold?

Alongside a confidence band, the AI expert publishes a range it expects the price to land inside. If it says “80% of the time”, then about 80% of outcomes should actually land inside that range. Aim: 80.0%; actual: 63.3% (n=60).

Where these numbers come from

Measurement window: the shadow window. Every figure on this page is drawn from that window and no other.
Signals recorded: 1,121 · scored (finished, so countable): 970.
Model version: aix2-f3 · strategy library aix2-s3. Versions are published so a change in the numbers can be traced to a change in the model.
Computed 09 Sept 2026, 03:00, server-side, from signals the platform recorded.

A trailing-window record — it changes when signals finish, not second by second.

What this page does not claim

It is not a profit figure. A well-calibrated forecast is not a money-making one. Nothing on this page states, or implies, what you would have earned.
It is not a forecast. Past calibration does not tell you how the next signal will land. Markets change, and the model can drift — which is why the coverage check above can automatically demote the AI expert back to shadow.
Nothing is hidden while it looks bad. Losses sit next to wins, at the same weight; a band we cannot yet stand behind is marked withheld rather than quietly dropped; and if the record gets worse, this page gets worse with it.
No promise of profit, ever. Every AI signal still passes through YOLO’s risk controls and still needs your own two-step approval before anything reaches your broker.

More on how we are paid and why our numbers cannot be self-reported: Trust & fees. The platform in full: About YOLO. The full risk disclosure: Legal & risk.