Tabular data: the AI problem nobody talks about

AI has solved forecasting, and anyone still doing arithmetic by hand is behind — that is the hype talking. AI is all invention and cannot count, so none of it is worth your attention — that is the backlash, and it is just as wrong. Both leave you holding the wrong tool for the decisions you actually make, which are almost never about pictures or sentences. They are about rows and columns: a price, a date, a quantity, a rate. That problem is unsolved, it is harder than the ones getting the attention, and it is worth understanding why.

What tabular data is

Structured information in rows and columns, where each row is one observation and each column is one measured thing. It is the least glamorous data in the world and it runs almost everything:

• Finance: prices, transactions, credit records.
• Healthcare: patient histories, lab results, trial data.
• Business: sales, inventory, customer records.
• Science: measurements, sensor readings, survey responses.

# Example only: illustrative rows, not real market data
timestamp,open,high,low,close,volume
2026-01-28 09:30,185.50,186.20,185.30,185.90,1250000
2026-01-28 09:31,185.90,186.45,185.80,186.30,980000
2026-01-28 09:32,186.30,186.50,186.10,186.25,1100000

# One row per observation, one column per measurement.
# The meaning lives in the structure, not in the words.

Anyone who has used a spreadsheet or a database has worked with this. It is the substrate of nearly every consequential decision made in an organisation, and it is where the AI conversation is quietest.

Why nobody talks about it

It does not demo well. A picture conjured from a sentence is instantly impressive to anyone in the room. A better-calibrated chance attached to next month’s demand is not, even when it is worth far more, because appreciating it requires knowing what the old number was and waiting to see which one was right.

There is a second reason, and it is the expensive one. Progress on images and language was so dramatic that people assumed the same approaches would carry over to everything else. They do not. The properties that made those problems tractable are missing here, and the properties that make this problem hard have no equivalent there.

Why it is hard

There is no structure you can assume

Pixels next to each other are related, always. Words next to each other form phrases, always. Those guarantees are what let a system be trained on enormous piles of unlabelled images or text and still learn something general.

No such guarantee exists in a table. Whether column three has anything to do with column seven depends entirely on which table you are looking at, and the answer does not transfer to the next one. There is no universal warehouse of rows and columns to learn from, because the structure is different every time.

Every column is a different kind of thing

In an image, every pixel is the same kind of thing: a colour. In a table, one column is a date, the next a price, the next a category with four possible values, the next a count that is usually zero. Each behaves differently, each needs different handling, and methods that thrive on uniform data are on unfamiliar ground.

The predictable part is tiny

A cat is unmistakably a cat in most photographs. A price tomorrow is mostly noise with a faint, unstable pattern buried in it. In markets this is not an accident of measurement; it is the mechanism. Anything obvious has already been acted on by someone, which removes it from the price and from your reach.

The rules change while you work

Cats looked the same in 2010 as they do now. Markets do not. Relationships that held for years dissolve in a fortnight and re-form somewhere else. Whatever was learned has a shelf life, which means learning once and shipping it is not a strategy, and any claim of a fixed accuracy figure should be read with that in mind.

Other people are trying to do the same thing

Photographs do not rearrange themselves to defeat a classifier. Markets are full of participants doing exactly that, not out of malice but because acting on a pattern is what erases it. You are not reading a fixed system; you are reading one that reacts to being read.

Why a language model does not solve it

Because it was built for a different substance. This is not a complaint about capability; the tool is extraordinary at what it was made for. It is a mismatch, and it shows up in three places.

Numbers arrive as fragments of text

A price is broken into pieces before the system ever sees it, so 185.50 becomes a handful of characters rather than a quantity sitting between 185.49 and 185.51. The ordering that makes a number a number is not there to be used. It can be reconstructed, sometimes, with effort, but it is not native, and every reconstruction is a place to be wrong.

A number in a sentence is not a chance

A language model can write “there is a 70% chance of an increase” with total fluency. That string is a plausible continuation of the preceding words. Nothing in it was counted, and nothing obliges it to match how often such increases actually occur. It reads exactly like a measured chance and is not one, which is a worse failure than being obviously wrong.

The wrong instincts

It has been shaped to be good at grammar, meaning and reasoning over prose — genuinely hard things, none of which is the thing a column of numbers requires. The longer version of this argument is in why ChatGPT cannot predict the stock market.

Seventy experts trained on numbers, not one language model trained on words

That contrast is the whole of Kunkafa’s position. Seventy experts, one question each. Each judges how far the price could move and in which direction, on its own, and none of them has ever read a sentence about markets.

Seventy independent views of the same market. When they agree you see it; when they disagree, that tells you something too. They learned from about 10 billion data points of price history across stocks, indices, currencies, commodities and crypto, each was checked on millions of past moves it never saw, and the chances they state are compared with what actually happened and corrected when the two drift apart. That process is set out in how the forecasts are trained and checked.

Markets are the proving ground

Not because finance is the point, but because it is the harshest available test that also grades itself. Every one of the difficulties above is present at full strength, and unlike a patient outcome years away or a demand figure nobody reconciles, the answer arrives on its own within a known duration and cannot be argued with. The price either reached the level in the time it had, or it did not.

That makes markets an unusually honest laboratory, and it makes evasion impossible. It also sets the expectation properly.

The same shape of problem sits in patient risk, demand planning and fraud review: rows and columns, weak patterns, moving rules, decisions that matter. What is learned under market conditions travels; what is learned under gentle conditions does not.

What to ask of anything that forecasts numbers

Three questions, in order. What was it trained on, and was it checked on anything it had not already seen? Does it show you both directions at once, or only the one it likes? And where is the record published, in full, including the periods it would rather not discuss?

Ours is on the results page, updated daily, with raw figures at /api/stats.json; how every number on the screen is produced is on the methodology page, and the questions people ask most are answered in the FAQ. Where fewer than 100 finished examples exist, no rate is shown at all, because a number counted on a handful of cases is a decoration.

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