Keep a real record of your AI investment research
You already ask ChatGPT, Claude or Gemini about the market. The answers are good — and then they are gone, scattered across old chats, impossible to compare, never checked against what actually happened.
ForecastNest gives you a structured way to ask, stores exactly what was asked and when, and turns the outcome into a scorecard once the horizon closes. Bring any AI you already pay for.
Free to start — no card, and no AI subscription of ours to buy. You use your own.
Research management for people who use AI
We do not sell you predictions, and we make no claim about how accurate your AI is. We give you a standard way to ask, a record of what was said, and an honest answer about whether it held up.
A consistent way to ask
Every prompt is built the same way and states the output it wants — the price it found, a range, a confidence, a direction per horizon. Your AI does the research: give it one that can search, and it reaches news, filings and positioning as of today. Ask the same question twice and the difference is the answer, not the question.
Or bring your own template
Prefer your own method? Use it. As long as the answer comes back in the same output format — a price range, a confidence, a direction per horizon — it is recorded and scored exactly like any other.
A record that cannot drift
The window is stamped when the prompt is built, so a forecast cannot be backdated into a period that has already played out. There is no editing a prediction once it is saved, and none of it can be quietly revised when it goes wrong.
See it on the chart
Every forecast generates a TradingView indicator — paste it into the Pine editor and the window and the range are drawn where the price actually went. One button charts all five models at once, each in its own colour, which is the quickest way to see whether they disagreed about direction or only about how wide to be.
Any AI, side by side
ChatGPT, Claude, Gemini, Grok, DeepSeek, Kimi, Perplexity — paste each one's answer against the same prompt and they line up next to each other. Models keep getting better; this is how you find out which of yours is.
You settle it, we keep the score
When the horizon closes you enter what the asset did — the chart script reads the figures back for you — and every model that answered is settled at once. From there the arithmetic is ours: did the close land inside the stated range, does a 70% confidence hold up 70% of the time, and how far was each model's own idea of the price from the truth. Abstentions are recorded as abstentions, not counted as misses.
Four standing experiments
Beyond your own questions there are four standard ones the site runs every trading day — AI Picked, Event Gaps, Macro Rotation and Extreme Movers. Each has a fixed prompt and a fixed way of scoring, so results are comparable across models, across members and across months. Subscribers can run any of them with their own AI against the same prompt and the same dates.
A room where the runs are shared
Every run a subscriber makes is visible to the other subscribers, credited to whoever made it. The point is not a leaderboard — it is that the same question, asked the same way on the same day, has been answered by a dozen different models and you can read all of them side by side. What the site itself runs is public once settled; what members run stays inside the room.
Everything stays searchable
Filter your history by symbol, horizon, model or prompt type. The exact prompt behind every forecast is kept alongside it, so you can always see what the model was actually looking at.
Three steps, then it keeps itself
The work is asking well and remembering honestly. Both of those are ours; the analysis stays yours.
1
Ask
Pick a symbol and a horizon — that is the whole form. We build the prompt, the same structure every time, stating exactly what to answer and in what shape. We do not tell the model what the asset is trading at: finding that out is part of the task, and it is the clearest sign of whether a model looked anything up.
2
Record
Paste the answer back and say which model wrote it. The window runs from the moment the prompt was built, so a forecast cannot be backdated into a period whose outcome is already known — and once saved it cannot be edited.
3
Settle
When the horizon closes you enter what the asset actually did — or read it off a chart, using the TradingView script we generate for you — and every model that answered is settled together. Did the close land inside the stated range? Do the confidence numbers hold up? Which of your models is doing better?
Your method
Use your own prompt if you prefer
Our templates are a starting point, not a requirement. If you have a way of asking that works for you, keep it — the only thing that has to match is the shape of the answer.
Anything that comes back in this format is parsed, stored and scored exactly like every other forecast, and lines up beside them in your history. Do include current_price: we never tell the model what the asset is worth, and how close its own figure lands is part of the score.
The price the model found for itself — it is scored on that
A price range in currency, so coverage can be measured
A confidence number, so calibration can be measured
A direction per horizon — NEUTRAL is a valid answer
Expected output
{ "signals": { "4h": "BULLISH | MILD BULLISH | NEUTRAL | MILD BEARISH | BEARISH", "1d": "...", "7d": "...", "30d": "...", "180d": "..." }, "current_price": "$214.30", "price_range": "$210.00–$219.50", "confidence": "58%", "data_access": "live | partial | stale", "model": "GPT-5.1", "answered_at_utc": "2026-09-01T14:32:00Z", "rationale": "why, in the model's own words" }
Scoring
Measured in the order the numbers can be trusted
Single-period direction is close to a coin flip for everyone, and we will not pretend otherwise. The measures that carry real information come first.
Range coverage
How often the close actually landed inside the stated range. Unambiguous, and the number worth quoting.
Confidence calibration
Do the forecasts labelled 70% come true about 70% of the time? Almost nobody measures this, and it is what separates a useful forecaster from a confident one.
Direction accuracy
Reported last, and honestly. Abstentions are recorded as abstentions rather than counted as misses, so an AI that says "not enough evidence" is not punished for being right about that.
Models keep improving. The point of keeping a record is that when yours gets better, you will be able to show it — and when it does not, you will know that too.
Nothing here is investment advice. It is your research, recorded and measured.
Start keeping a record of what your AI actually said
What it costs
One paid plan, billed monthly. It buys more room to work and the parts of the tool built for people who use it daily. The research itself, the scoring and the record are identical on both — paying does not buy a better answer.
Free
Enough to see whether this suits you
$ 0
15 research prompts a month
- Research prompts built for any asset you name
- Answers recorded and scored when the horizon closes
- Your own scorecard — hit rate, range accuracy, calibration
- Unlimited history and scoring, on both plans
- The four standing experiments, public once they settle
Member Room to use it
For research as a habit, not a trial
$ 9.99 /month
100 research prompts a month
- Everything in Free, at 7× the monthly allowance
- Answers recorded and scored when the horizon closes
- Your own scorecard — hit rate, range accuracy, calibration
- Unlimited history and scoring, on both plans
- The four standing experiments, public once they settle
- Run any of the four standing experiments with your own AI
- Read every other subscriber's runs, credited to whoever made them
- Follow the standing experiments before they settle
Billed monthly and cancellable at any time from your profile. Cancelling stops the next payment and you keep access until the month you have paid for ends. We do not refund payments already made — see the Cancellation and Refund Policy.
Prices are in US dollars and exclude sales tax, which Stripe calculates and adds at checkout for where you live. Stripe is the merchant of record and handles the payment; we never see your card details. Allowances run per calendar month and do not carry over.
Questions worth asking first
Still unsure about something? The contact form below reaches us directly.
Do I need to buy an AI subscription from you?
No. You use whatever you already pay for — ChatGPT, Claude, Gemini, Grok, DeepSeek, Kimi, Perplexity, or something else entirely. We build the prompt and keep the record; the thinking happens wherever you already do it.
How accurate are the forecasts?
That depends on your AI, not on us, and we will not put a number on it in marketing. What we will do is measure yours honestly and show you the result: how often the close landed inside the stated range, and whether the confidence numbers hold up. If your model is not good at this, your scorecard will say so.
Can I use my own prompt instead of your templates?
Yes. The templates are a starting point. The only requirement is that the answer comes back in the expected format — the price the model found, a price range, a confidence, and a direction per horizon. Anything matching that is parsed, stored and scored the same way, and sits beside the rest of your history. Keep the price in: we never tell the model what the asset is trading at, and how close its own figure lands is part of what gets scored.
What is free, and what needs a subscription?
A free membership runs 15 research prompts a month, keeps its full history and scorecard, and reads the four standing experiments once they settle. $9.99 a month raises the allowance to 100 and opens up the rest of the tool: running the standing experiments yourself with your own AI, and reading every other subscriber's runs while they are still open. More lands on that side over time. What does not change is the research itself — same prompt, same record, same scoring. You are not paying for a better answer. Billed monthly, cancel at any time, and you keep the month you have paid for.
How does a forecast get scored?
You settle it. When the horizon closes, enter what the asset was worth when you asked and what it closed at, and every model that answered that same question is settled at once — they shared a window, so they share an outcome. Optionally add the window's high and low. Every forecast generates a TradingView script that reads all four figures off the chart for you, which is both easier and more accurate than looking them up. Everything after that is arithmetic we do.
Can I edit or delete a forecast after saving it?
You can delete one that has not been settled yet — for a mis-selected model or a wrong paste. You cannot edit one, and you cannot delete one that has already been settled. A track record you can revise after the fact is not a track record, and every number on your scorecard is computed from these rows.
Is my research private?
Private by default. Sharing is an explicit act, one forecast at a time. Shared research shows only your display name.
Which symbols can I research?
Any listed stock, ETF or index, and any major crypto pair. Your AI does the looking, so you are not limited to a list we maintain. Search by ticker or company name.
Is this investment advice?
No. ForecastNest is a research and record-keeping tool. Nothing here is a recommendation to buy or sell anything, and the forecasts are produced by your own AI rather than by us. Treat it as a notebook that checks its own arithmetic.