By Kareem Farid, co-founder of Kunkafa
Kunkafa vs. OpenAI’s ChatGPT for Financial Services
OpenAI brought ChatGPT to Wall Street’s documents. Here is why reading about a market is not forecasting it, and what Kunkafa does instead.
This week OpenAI put ChatGPT inside investment banks. Junior bankers get a model that reads every filing and transcript, cites its sources and drafts the deck. It is a good product for that job, and we read the announcement with interest, because the job it does not claim is the one we built Kunkafa for: saying how far a price could move, which way, by when, and how confident to be about it.
The finding. No, and OpenAI does not say it does. Reading data and putting it into words is one job. Training on the numbers themselves and putting a probability on a precise move is another, and it is the only thing Kunkafa does.
What OpenAI launched, and what it is for
ChatGPT for Financial Services is a tailored ChatGPT for banks: GPT‑6 Astra with built-in financial data from Daloopa, PitchBook, LSEG News and Crunchbase, shaped with Morgan Stanley and Evercore, sold to eligible institutions for investment banking and equity research. It reads filings, transcripts and fundamentals, cites its sources, and produces research, valuation models and client materials in the firm’s own templates. Read the page twice and notice what is not on it: no claim about where a price goes. That is honest, and it is the point of this post.
Reading about a market is not forecasting it
A filing tells you what happened. A transcript tells you what somebody said. A language model can read all of it faster than any analyst and turn it into a fluent paragraph about what might come next. That paragraph is still words about data. It comes from a model trained on words about markets, and it inherits their habit: people write assuredly about the future, so the model writes assuredly about the future. Nothing in it names a level, a direction, a duration or a probability that anyone can check afterwards.
We took the other road. Kunkafa was never handed a document. Kunkafa was trained on the numbers themselves — about 10 billion price data points across every market and every duration, refreshed every day — to do one narrow thing well: put a precise figure on a move. How far the price could go, in which direction, within how long, and Kunkafa’s confidence that it gets there, with the other direction worked out beside it. Because every forecast names its level, its direction and its duration, it resolves on its own, and how often Kunkafa was right is counted every morning. The longer arguments are in why words about prices are not price paths and what a model needs instead.
Language models and Kunkafa, side by side
| Question | Language models | Kunkafa |
|---|---|---|
| Trained on | Text about markets: articles, filings, transcripts | Price history: about 10 billion numbers, no articles |
| Asked | Summarise, model, draft the deck | Could gold go up around 1% within 12 days? |
| Answers with | A fluent sentence about the future | A stated move, its direction, its duration, and Kunkafa’s confidence that it gets there |
| The other direction | Rarely mentioned, never measured | Always shown, a glance below |
| Confidence | The tone of the training text | A calibrated figure with its record beside it |
| When wrong | Nothing resolves, so nothing is counted | Counted every morning and published, misses included |
How Kunkafa is scored
Confidence is cheap until somebody keeps score. Kunkafa’s confidence of 84% on a stated move means about 84% of resolved forecasts like it reached the level, and the count that settles it sits next to the figure, on the results page and in the app, recounted every morning. Where a count is too thin to trust, the screen says so instead of quoting a rate. A market Kunkafa cannot call is said out loud: no more confident one way than the other. Misses are counts, not apologies, and the record shows both.
Where each one fits
Hand Kunkafa a filing and nothing happens. Kunkafa builds no deck, runs no workflow, connects to no data room and carries no compliance console. A bank that needs those needs a document product, and this one was built with two of the finest names on Wall Street.
What we claim is narrower. A figure produced natively from numerical price series rather than written into a fluent sentence. Both directions, always. Every duration from five minutes to seven years, at once. Calibration shown in the open, misses included. And you do not have to be an institution: Kunkafa is a question in plain words, on every plan, for anyone who would rather weigh scenarios than be told a story. Kunkafa never tells you what to do; a confident view is still a view, and the decision stays yours.
Three questions for anything that forecasts a price
- What was it trained on? Words about markets, or the price paths themselves?
- Does the answer name a level, a direction and a duration? If not, nothing resolves, and nothing can be scored.
- Who counted afterwards, and where is the count? A confidence without its record is a tone of voice.
Wall Street now has AI for the deck. For the question the deck cannot answer, create a Kunkafa account and read one live forecast beside its record. Every plan, including Free, comes with a 15-minute call with Kunkafa; request it from Account once you are in.
Keep reading
Twelve guides that go with this one, by the question you came with.