Can AI estimate market probabilities? Look at the record, not the claims

A screenshot of a 94% win rate is greed talking; it is selling the feeling of certainty, which is the one thing markets never hand out. "Nobody can forecast anything" is fear talking, and it costs just as much, because it leaves you with no way to compare one path against another. Neither belief survives contact with a record. So ask for the record.

The useful version of the question is not whether a machine can see the future. It is whether a number it prints today can be checked against what happens afterwards, by you, on a page you did not have to trust in advance.

The question only means something in its checkable form

"Will the market go up?" cannot be graded, because it never said by how much, by when, or how sure. Rewrite it as "of finished forecasts that looked like this one, how often did the price reach that level, in that direction, inside that duration?" and it becomes a counting problem with an answer that either holds up or does not.

That rewrite is the whole argument. Kunkafa Predictions answers one question per market — how far the price could move, in which direction, and how likely each outcome is, across every duration at once — and every number it prints is a rate taken from forecasts that have already finished. There is nothing to believe. There is something to check.

What a record has to specify before it means anything

Five things, and if any of them float, the headline number can be made to say almost anything: which forecasts were counted, which were dropped, what counts as finishing in the forecast direction, what counts as reaching the level, and over what window.

Ours are fixed and shown beside each forecast, in a panel called "How similar forecasts performed". It counts forecasts evaluated, how many were strong enough for the filter you set, how many finished in the forecast direction, and how many reached the movement level you selected — over the last day, the last 7 days, the last 30 days, or all time.

How similar forecasts performed        [ 30 days ]

  Forecasts evaluated                        4,812
  Strong enough for this filter              1,046
  Finished in the forecast direction           58%
  Reached the selected movement level          41%

Illustrative shape, not live figures.
Under 100 finished examples, the panel prints
no rate at all - it says the sample is too small.

Two rules do most of the work. Under 100 finished examples the panel refuses to quote a rate and says so: "Not enough past examples to estimate performance reliably." And a rate below 1% prints as "under 1%", because a decimal there would claim a precision the sample cannot carry.

Outcomes are named rather than graded on a curve: reached the movement level, passed the movement level, period ended ahead, period ended behind, still running. A forecast that is still open is counted as still open, not quietly parked until it looks better.

Where to check ours

The results page is public and updated daily: kunkafa.com/performance, with the raw figures behind it at /api/stats.json for anyone who would rather load the numbers than read a chart. The counting rules are written out in the methodology, and every term used on screen is defined in its glossary.

You can also check it from the inside without an account. Open the Kunkafa demo, pick a market, and move the strength filter: the ladder of levels narrows, and the counts in the record panel move with it. If a filter that keeps almost nothing still reported a comfortable rate, you would have caught us.

There is a third way to check, slower and more convincing than either: grade us yourself. A saved outlook is a frozen copy of a forecast exactly as it stood when you saved it, so nothing about it can be revised afterwards. Keep a handful across different markets and durations, come back once they have run their course, and compare what you kept against what the price did. That is the same arithmetic the results page performs, on a sample you chose rather than one we picked.

Why a 90%-accuracy claim is a warning sign

Because accuracy with no stated question is a free parameter, and there are at least four ways to manufacture a big one. Narrow the filter until only a handful of examples remain, and any rate becomes reachable. Count only the forecasts that closed well and leave the open ones out. Pick the window that flatters. Or grade an easy question — "did the price move at all, in either direction?" — and report the high rate without mentioning that it was never in doubt.

None of that requires dishonesty. It is what happens by default when a number is chosen for a landing page instead of being produced by a rule. The defence is boring on purpose: publish the sample beside the rate, keep the dull forecasts in the denominator, refuse to quote anything under 100 finished examples, and let a reader change the filter and watch the counts change.

The same instinct applies to the products themselves; the problem with AI trading apps covers what tends to be hidden and why.

Why most forecasts sit near 50/50

Because the market has already read the news. Markets are efficient — whatever is known is already in the price, so most forecasts sit near 50/50. Our models watch every update and bring you the few that do not.

That is not a hedge, it is the finding. A tool that produces a strong lean on every market at every duration is not seeing more than the market; it is reporting noise with a firm voice. Ours prints the balanced cases as balanced, and leaves them in the counts, which makes the record duller and the rates lower than they would otherwise look. That trade is deliberate — why exact price forecasting is impossible goes further into why the honest ceiling is lower than the advertised one.

Seventy experts, one question each

Seventy experts, one question each. Each of them judges how far and which way on its own, and what reaches the screen is what came back. Seventy independent views of the same market. When they agree you see it; when they disagree, that tells you something too.

They were built on about 10 billion data points of price history spanning stocks, indices, currencies, commodities and crypto, and each was checked on millions of past moves it never saw. Stated chances are then compared with what actually happened and corrected, which is the step that turns a number into a rate you can hold us to.

What it still cannot do

Four limits, stated plainly, because a tool that hides them is back to selling certainty.

• It cannot tell you what happens next. A 70% chance of reaching a level leaves a 30% chance it is not reached, and that side arrives often.
• It cannot turn a rate into a return. Sizing, costs, timing and the trades you skip decide that, not the ladder.
• It describes finished forecasts, not the next one. The record is evidence about a population, and your market today is one draw from it.
• It cannot rescue a market it has too little history on. Where the sample is thin, the panel says so instead of guessing.

Common questions about coverage, durations and what each plan opens are answered in the FAQ. Looking costs nothing in any plan, including the free one, so nothing about checking the record depends on paying for it.

So: can AI put a number on where a price could go? It can produce a rate, publish the rule that made it, and let you count. Everything after that is your judgement, which is exactly where it belongs.