What is going to happen, with a number and a date.
A question stated plainly, an argued probability, the sources underneath. And when reality settles it, the result is recorded — whether it proves us right or wrong.
The data bank
Where progress actually stands — measured, not narrated. Each block queries a free public database live and shows its source. Nothing here is typed in by hand: if a number moves tomorrow, the page moves with it.
What cannot be measured automatically yet
Robotics has no usable public database. There is no Hugging Face equivalent for robots: production announcements, order books and deployments live only in press releases. That ranking will be kept by hand, from the daily watch, with a verification date on every line — and saying so is more honest than manufacturing a league table.
Chipmakers' accounts are public but heavy. Nvidia, AMD and Intel file their figures with the US regulator, reachable by API without a key. That is the next source to wire in: it will give a ranking by actual revenue rather than by press estimate.
Private lab valuations cannot be verified. No open database publishes them. They will stay displayed as reported figures, with their source and their date — never as facts.
The feed
Every article the pipeline actually put on file, grouped by source tier. These are the pieces that moved the probabilities — not a separate news selection.
The deadline calendar
Every question sorted by date: what has to be settled, and when. Each row opens its page.
Were we right?
The only page that counts. A good forecaster is not the one who is always right: it is the one whose “70%” comes true about seven times out of ten.
| Domain | Resolved | Brier | Mean error |
|---|
The Brier score is the mean squared error: 0 is perfect, 0.25 is what answering “50%” to everything gets you.
The mean error is the average distance between the stated probability and the outcome, in points. More intuitive than Brier, but it forgives large misses: read the two together.
None of these questions exists on a prediction market — that is the point of the site. So the comparison isn't a price: it's a reference score. 0.25 is what answering "50%" to everything gets you; a "reference" row also appears once a question whose base rate is computed from data (FDA registry, Fed futures) has resolved — the first is due on 26 October. Until then, with this few questions, nothing here proves anything.
Calibration by model
The model used for every estimate is recorded and published. Read the count before the score: three questions are an anecdote, not a track record.
| Model | Resolved | Brier | Mean error |
|---|
Calibration by evidence tier
Grouped by the strongest source on file when the question was last estimated (see "How this is made"). A pattern here would say whether weaker sources are correctly discounted — read the count before the score, same caveat as above.
| Evidence tier | Resolved | Brier | Mean error |
|---|
Most recently resolved
| Question | Resolved on | Stated | Error | Outcome |
|---|
How this is made
Five steps. Wording and estimation are always machine; for finance, crypto/ETF and mergers, spotting is now adopted automatically once a candidate clears a mechanical filter and an adversarial AI review — no systematic human step there. Everything else is a scheduled script, readable, with every decision traced. Quantum and deep tech are the exception: there is no reliable public calendar of dated, checkable events for either, so nothing there is auto-discovered — new questions are added by hand, occasionally, only when a genuinely verifiable one comes up; existing ones still update and resolve like any other.
For science and health, the watch reads international press in the language of the country concerned, Hacker News and lab announcements. For finance, crypto/ETF and mergers, it reads official filings on a fixed schedule instead (SEC filings, FTC merger notices) — no press. Either way, it proposes candidate questions.
Every question gets a deadline, an objective criterion and a designated resolution source before the estimate. Without those three, it does not ship.
The model gathers the sources, produces a probability and the reasoning behind it. The reasoning is what makes the page worth reading.
One page per question, and a snapshot of the sources archived on the day of the estimate — so that a dead link six months later does not make the forecast unverifiable.
At the deadline the designated source is consulted, the verdict recorded, the calibration recomputed. This is where a forecast becomes measurable.
What this site is not. Not financial advice, not medical advice. Health questions are about agency decisions and clinical-trial endpoints — never about what a treatment will do for any particular person.
What we commit to. Publishing the misses as visibly as the hits, never changing an estimate without timestamping the change, and showing the score even when it is bad.
Where the numbers come from. Press: RSS feeds, queried in the language of the country concerned, each article carrying a source tier. Company figures: official filings where they exist, otherwise press, cited and dated. No prediction-market odds: none of these questions has a market.
Data and credits. Official records used for resolution belong to their publishers: U.S. Food and Drug Administration (openFDA), U.S. National Library of Medicine (ClinicalTrials.gov), U.S. Securities and Exchange Commission (EDGAR), U.S. Bureau of Labor Statistics, U.S. Employment and Training Administration and U.S. Treasury series republished by FRED, Federal Reserve Bank of St. Louis (not endorsed or certified by the Federal Reserve Bank of St. Louis), NASA GISS (GISTEMP), NSIDC (Sea Ice Index) and Epoch AI (CC BY 4.0). The probabilities on this site are PolySignal's own estimates, not those of these organisations.
Source monitoring
Recent publications flagged as relevant by PolySignal's automated monitoring agent, across all tracked domains.