Ten billion data points, seventy experts: how the forecasts are trained and checked

Ten billion is the kind of number people use as a boast, and boasts are how anyone talks themselves into a position. Nasdaq going up forever? That is greed talking. Fear says the opposite — and both can wreck the next decision. Kunkafa does not take a side. It maps the paths, how far each one could run, and the chance of getting there. So the scale is not the claim. What the scale is for, how the result is checked afterwards, and where you can check it without taking our word for any of it: that is the claim.

What the scale is for

Coverage. About ten billion data points of price history sit behind the forecasts, drawn from stocks, indices, currencies, commodities and crypto. The reason is unglamorous: 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. Learn from one asset class and you get something fluent in that asset class and lost everywhere else.

Price history is also unusual among datasets in that 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, and the relationships between markets tighten and break and re-form without warning. There is no clever way around that. There is only breadth: enough different markets, across enough different years, that a violent week has been seen many times over rather than once.

Volume on its own would not do it. Ten billion data points taken from one currency pair over one year would be ten billion rehearsals of a single mood. The point of the number is the range of conditions it covers, not the number.

Seventy experts, one question each

Seventy experts, one question each. Each of them judges how far the price could move and in which direction, on its own, and none of them sees what the others concluded before answering.

Seventy independent views of the same market. When they agree you see it; when they disagree, that tells you something too. Agreement that survives seventy separate looks means more than agreement produced by one look repeated seventy times. Disagreement is not a fault to be smoothed over — it is the market being genuinely unclear, and the honest thing is to say so rather than to manufacture a direction out of a tie.

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. A forecast that showed you only the side it favoured would be hiding the more useful half of the picture, which is how far the other path could run and how likely it is.

Checked on moves it never saw

Every expert is checked on millions of past moves that were held out of its training entirely. This is the part of the work that decides whether the rest of it means anything, and it is the part most easily faked.

Something graded on the same history it learned from will look extraordinary and tell you nothing, in the way a student who has seen the exam paper looks like a genius. Holding data back removes that comfort. The move has to be one the expert has never encountered, in a market it cannot have memorised, and the answer either matched what happened or it did not.

Millions of such moves, rather than a few hundred, is not thoroughness for its own sake. Rare conditions are rare: a sample small enough to fit on one screen will contain almost none of them, and a record built on it would describe calm weather and nothing else.

Stated chances are compared with outcomes, then corrected

A chance printed on a screen is only worth reading if someone has counted it. So we count it: take every past forecast that stated a given chance, count how often the move actually arrived, and put the two numbers side by side. If forecasts stating a 70% chance of reaching a level arrived far less often than 70% of the time, the number is wrong, and the number gets corrected rather than explained.

This is the whole difference between a forecast that sounds careful and one you can use. A stated chance you can take at face value is worth more than a high one you cannot, and it has to hold in both directions at once: if the up side reads 62% and the down side 55%, each its own chance and never made to add up to 100, both of those have to survive being counted against what the market did next.

It also explains why so much of what you see looks unremarkable.

Anything that made most days look decisive would be telling you about itself rather than about the market. The near-50/50 majority is the honest answer arriving on time, and the reason the few forecasts that lean are worth stopping on. That argument is made at greater length in Probability, Not Prophecy, and the limits of the idea in can AI estimate market chances at all.

Why the internals stay private

Two reasons, and neither is mystique. The first is that a market absorbs whatever it is told. Anything that works does so because it is not already priced in, and a full description of how it works is a set of instructions for pricing it in. The second is more awkward for anyone asking: a description proves nothing anyway. You cannot verify a claim about markets by reading a paragraph about the machinery behind it, however technical the paragraph is. Publishing the architecture would buy you the feeling of having checked, not the fact of it.

What can be checked is the outcome. So the record is published instead, in full, including the parts that are unflattering, and it is the thing we would rather be judged on. How the screen works and what every word on it means is written out on the methodology page, with plain definitions in its glossary.

Where to hold us to account

The results page, updated daily, with the raw figures behind it available at /api/stats.json for anyone who would rather do their own arithmetic than read ours.

The same accounting sits beside every forecast in the app, under How similar forecasts performed, over the last day, 7 days, 30 days and all time:

• Forecasts evaluated — how many finished and could be counted at all.
• Strong enough for this filter — how many of them cleared the strength you selected.
• Finished in the forecast direction — how often the price ended the period on the side the forecast leaned.
• Reached the selected movement level — how often it actually travelled the distance, which is the harder question and the one that matters.

What the record does not buy

It 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, and the chance shown beside it is the honest width of that ignorance rather than a formality. Markets keep changing their own rules, which is why the counting never stops and why the results page is dated.

It also does not buy a recommendation. Nothing here tells you what to hold, how much of it, or when to leave; the questions people ask most often about that boundary are answered in the FAQ.

Forecast, not advice. For the rational investor: emotion out, scenarios in.