By the Kunkafa team
How Kunkafa Trains and Tests Its AI Market Forecasts
About 10 billion data points, seventy experts, and a test on about 2 billion points they had never seen. How the forecasts were built, and where you check the result.
Ten billion is the kind of number a company puts on its homepage. You have seen it there. What you want to know is what it was for, and whether anything came out the other end that can be checked.
The finding. The scale bought coverage: enough markets and years that a violent week has been seen many times. Seventy experts were trained on it, then tested on about 2 billion points they had never seen.
Where Kunkafa stands. Kunkafa keeps the machinery private and publishes the outcome instead: every morning, on live markets, at its results page.
What to do. Judge the forecasts by the record, not the training figure. Past results do not promise future ones.
What the scale was for
Coverage. About 10 billion data points of price history sit behind the forecasts, drawn from stocks, indices, currencies, commodities and crypto. A market only teaches you what it has already been through. Learn from a stretch of calm and you get something that has never met the week the calm ended.
Price history is unusual because its rules move. A cat in 2010 looks like a cat in 2026. An index in 2008 does not behave like the same index in 2017. Relationships between markets tighten, break and re-form without warning.
There is no clever way round that. There is only breadth. Ten billion points taken from one currency pair over one year would be ten billion rehearsals of a single mood. The range of conditions is the point, not the number.
Seventy experts, one question each
On that history, seventy experts were built. Each judges how far the price could move and which way, on its own. None sees what the others concluded.
When they agree, you see it. When they disagree, the probability drops, and you see that too. Disagreement is not a fault to smooth over. It is the market being unclear, and the honest thing is to say so.
That is also why the other direction is never more than a glance away. The side you asked about comes first, the other sits below it, on any duration from five minutes to seven years.
Tested on about 2 billion points they had never seen
This is the part that shows whether the rest means anything, and the part most easily faked. Something graded on the history it learned from looks extraordinary and tells you nothing. A student who has seen the exam paper looks like a genius.
So about 2 billion data points were held out of training entirely. Every expert was checked on moves it had never encountered. The answer either matched what happened or it did not.
Billions rather than a few hundred matters because rare conditions are rare. A small sample contains almost none of them, and a record built on it would describe calm weather and nothing else.
Kunkafa’s confidence is compared with outcomes, then corrected
A percentage on a screen is only worth reading if someone has counted it. So Kunkafa counts it. Take every past forecast that stated a given value, count how often the move arrived, and compare the two. If forecasts stating 70% arrived far less often than 70% of the time, the number is corrected rather than explained.
It has to hold in both directions at once. If the up side reads 62% and the down side 55%, each is its own value, never made to add up to 100, and both have to survive being counted against what the market did next.
This is why most of what you see looks unremarkable. Whatever is known is already in the price, so most forecasts sit near 50/50, and a 52% is shown as 52%. The few that lean are the ones worth stopping on. The argument is made at length in Probability, Not Prophecy, and its limits in whether AI can predict the stock market at all.
Why the internals stay private
Two reasons, and neither is mystique. A market absorbs whatever it is told, so a full description of what works is a set of instructions for pricing it in. And a description proves nothing anyway. You cannot verify a claim about markets by reading a paragraph about machinery.
What can be checked is the outcome. So the record is published instead, including the unflattering parts. Every word on the screen is defined in the Kunkafa market forecasting methodology and its glossary of forecasting terms.
Where to hold us to account
Every morning Kunkafa counts how past forecasts ended, on live markets, and publishes the result at Kunkafa forecast performance results, with the raw figures at /api/stats.json. The same record sits beside every forecast in the app.
- Forecasts evaluated. How many finished and could be counted at all.
- Confident enough for the filter. How many cleared the confidence you selected.
- Reached the level. How often it travelled the whole distance. The harder question.
Below 100 finished examples the record shows no rate and says why: not enough past examples to estimate performance reliably. A rate under 1% is written as under 1%, because that precision is not something we can honestly claim.
The record does not buy the next move. A rate counted over thousands of past forecasts describes those forecasts. The one in front of you could land on either side of it. It is a description of past outcomes, not a recommendation. The Kunkafa forecasts and alerts FAQ covers that boundary.
The training figure is where the product started. The record is where to judge it. See both sides of a live forecast, with its record beneath, in the Kunkafa market forecast app.
Keep reading
Twelve guides that go with this one, by the question you came with.