Signal to decision
Case study #1
From the round of 32 onwards, each prediction went out on a public feed before kickoff and was graded once the result came in. The group stage ran privately: those predictions were made and graded on the same rules, but nobody could see them in advance, so treat them as a private track record rather than something you can verify. This page is the final audit of all 104, the wrong calls kept alongside the right ones.
The same three steps that now run on business data.
Before kickoff the model produces a probability for each outcome and a scoreline. In the knockouts it also gives each side a chance of going through, and that is the call it gets graded on.
Before kickoff I screenshotted the dashboard's prediction, built a prediction board from that screenshot, and posted it to LinkedIn. The board stayed unchanged until kickoff, when the dashboard switched over to tracking the match live. So the record is a public feed post made before kickoff, not a timestamp on the dashboard itself.
Once the result was in, the call was marked correct or incorrect on the LinkedIn card. The dashboard holds the full audit of all 104 matches, and it keeps the wrong ones.
Both are in the record. Figures exactly as published before kickoff.
Loading...
Football was the proving ground: a fixed schedule and an unarguable answer at the end of every match, so a forecast either survives contact with reality or it does not. The method transfers directly. Forecast the next period, publish the range before the answer exists, grade it when it lands, and keep the misses on the record.
Follow the build on LinkedIn; every forecast on the live demo gets graded in public.
See it running on live UK retail data Explore the full dashboard