← Signal Before Signal
Track record
Every cluster the system flagged is reviewed once its signal window has been closed for 21 days, against whatever became public afterwards. Hits and misses are on the same page. Where a call was right, the link is the evidence — the dates are checkable without taking our word for anything.
Not enough reviewed yet
No clusters have been reviewed yet. Percentages appear once there are at least 5, because below that a hit rate describes the sample rather than the system. This page fills as clusters age past their review window rather than being written in advance.
How this is counted
- Hit rate is confirmed divided by confirmed + partly right + false alarm. Unclear cases are excluded, not counted as misses: plenty of private companies never announce anything either way, and calling that a failure would be as wrong as calling it a success. Wherever a rate is shown, the excluded count is shown with it.
- Lead time is measured from the start of the signal window to the date the news became public — not from the date the review was recorded. It is only measured where the call was at least partly right, and negative values are kept in the median.
- Reviews are manual and made by the operator of this instance. That is a real limitation: the same person runs the system and grades it. The evidence links exist so the grading can be argued with.
- Demo data is excluded everywhere on this page. The illustration on the home page is invented and never reaches these numbers.
Watch a project for a week and judge the next one yourself.