By the Kunkafa team
Can ChatGPT Predict the Stock Market?
You paste a ticker into a chatbot and get a fluent paragraph. Kunkafa was trained on price paths, not on words about them. Here is the difference.
You paste a ticker into ChatGPT and ask where it goes next. The answer comes back in the voice of an analyst: measured, balanced, sure of itself. It reads like knowledge.
The finding. A language model learns words about prices, not price paths. Research finds some predictive information in headlines and over-extrapolation from recent returns. A fluent paragraph is not a tested forecast.
Where Kunkafa stands. Kunkafa is trained on about 10 billion data points of price history. It puts a probability on a stated move, both directions, and publishes its record every morning.
What to do. Use a chatbot to read and explain. Ask a forecast for a level, a period, a probability and a record.
What you asked, and what the chatbot learned
The question is fair. You want to know how far the stock could move, which way, and by when.
A language model learned from text: filings, transcripts, articles, arguments about prices. It learned how people write about markets. It did not learn the path a price took from Tuesday to Friday, ten million times over.
So the answer is fluent about reasons and weak on paths. And it sounds equally sure whether the base rate is 80% or a coin flip.
Words versus price paths
A forecast is a claim about numbers. How far, which way, how long it has, how likely. Numbers reach a language model as fragments of text, and a percentage written into a sentence is not a calibrated probability.
The same confident style accompanies a sourced analysis, an assumption or an error. Tone tells you nothing about reliability. Evidence does.
The tell is the missing record. Ask the chatbot how its last hundred calls ended. It has no count. There is nothing to grade.
How Kunkafa is trained on price history
Kunkafa is an AI financial markets app for stocks, indices, currencies, commodities and crypto. Seventy experts, not one, each judging how far a price could move and which way. They were trained on about 10 billion data points of price history and tested on about 2 billion they had never seen.
A forecast has four parts: how far, as a percentage; which way; how long it has; and Kunkafa’s confidence that it gets there. Durations run from five minutes to seven years. The side you asked about comes first. The other side sits below it. When the seventy disagree, the probability drops, and you see that too.
The homepage example, with illustrative figures: could gold go up 1% within 12 days? Kunkafa’s confidence: 84%. Below it, down 0.6% in the same period, 70%. Either, both or neither could happen.
Then the step a chatbot cannot take. Every morning Kunkafa counts how past forecasts ended, on live markets. The count is published at Kunkafa forecast performance results, with the raw figures at /api/stats.json. A 52% is shown as 52%. Past results do not promise future ones.
Model count alone proves nothing. The record is what to judge, and the scoring rules are on how Kunkafa calculates market probabilities.
What research shows about ChatGPT stock forecasts
Lopez-Lira and Tang report predictive information when a model reads news headlines, in their study of LLM stock forecasts. That is evidence against the claim that language models can never help. It is a specific result, not a promise that a chatbot answer predicts the next move.
Chen and colleagues found excessive extrapolation from recent returns in the models they tested, alongside some better risk calibration than their human comparison. Findings like these belong to particular experiments and prompts, not to every model for all time.
Fine-tuning, structured inputs and calculation tools change what a system can do. They do not remove the need to test it on later, unseen data against a baseline.
Where a chatbot earns its place
Language models are good at reading. Summarise a filing, explain a term, organise a thesis, draft the questions you should be asking. Check the summary against the source.
That is a different job from forecasting a price path. Machine learning for tabular financial data explains why the two jobs need different training. How we trained an AI on 10 billion data points describes ours.
Four questions for anything that claims to see ahead
- Is a probability attached, labelled with what it measures?
- Is there a level and a period, so the claim can be scored rather than reinterpreted later?
- Are both directions shown, or only the one you were hoping for?
- Is there a public record, and does it admit when the sample is too thin to judge?
Kunkafa’s answers are on the record page, and the rest is in the Kunkafa forecasts and alerts FAQ. Next time someone says AI can call the market, ask what it was trained on and who counted afterwards. Then read a live forecast in the Kunkafa market forecast app.
Keep reading
Twelve guides that go with this one, by the question you came with.
