Artificial intelligence is improving quickly. Foundation models from OpenAI, Anthropic, Google and independent labs reason further, hold more context, and synthesise messier material than they did a year ago. That raises a question worth answering carefully rather than loudly:
As models get better at reasoning, do they get better at markets?
Nobody has a good public answer, partly because almost everyone asking has something to sell. Claims arrive as screenshots and end at the profitable trade. ForecastNest Insights is our attempt to answer it the other way round — in the open, on a fixed task, with the losses left in.
What we publish, and what we do not
We do not publish price targets, trading signals, or a view on where anything is going. That is deliberate, and it is the line the whole project sits behind: the subject of this research is the models, not the assets.
What we publish is how well models do on a task that never changes. Every model gets the same question, in the same shape, over the same window, and is graded afterwards by a method we have written down in public.
What you will find here
- Can it even find the price? We no longer tell the model what an asset is trading at — establishing that is the first part of the task, and it is graded on its own. It turns out to be the sharpest question we ask. On one run where the baseline was mistyped as $1.00 for an ETF trading near $760, three of five leading models forecast a range around a dollar. Two checked. That is not a subtle failure, and it is invisible unless you look.
- How far apart the models are. On the same asset, at the same moment, the price each model reports for itself has differed by anything from 0.5% to nearly 2%. Before any forecast is judged, that spread tells you which answers were grounded in anything.
- Whether the ranges hold up. Did the close land inside the stated range, and does a stated 70% confidence hold up 70% of the time? A range wide enough to cover every outcome is a hedge, and it is scored as one.
- Model leaderboards over time. Not one lucky call — the same models, the same task, repeatedly, so that improvement can be told apart from noise.
- Build in public. The prompts, the scoring weights, and the parts of the method we are least sure about. Including where the method could be gamed, and what stops it.
The method is public, because otherwise none of this counts
A score nobody can check is a number you are being asked to trust. Ours is written out in full — what earns points, what earns zero, and why a forecast that took no directional view is recorded as an abstention rather than as a miss.
👉 Read exactly how a forecast is scored
The sample is still small and we will say so on every chart until it is not. Publishing the method before the data is the honest order to do it in.
The tools behind it
Every benchmark here is produced with ForecastNest, which is the same tool you can point at your own symbols and your own models. You bring whichever AI you already pay for — ChatGPT, Claude, Gemini, Grok, DeepSeek — and it keeps the record: the prompt, the answer, the reasoning, and the score when the window closes. Each forecast also generates a TradingView script, so you can see the range drawn where the price actually went.
👉 15 research prompts a month, free. $9.99 for 100.
Disclaimer: content published on ForecastNest Insights is for research, educational and informational purposes only. Nothing here is financial, investment, legal or tax advice, and nothing here is a recommendation to buy or sell anything. We publish how models perform on a disclosed task; we do not publish price targets or trading signals. Always do your own research.