Germany Regionalliga: prediction calibration, in public
Every number on this page is recomputed from settled matches only — aggregated, never cherry-picked, with under-powered buckets greyed instead of hidden. This is what our model actually did in this league, not what we wish it had done.
League profile — settled sample
Calibration: predicted vs realized
Before kickoff the model assigns a probability to its top 1X2 pick. Each row groups matches by that confidence. A calibrated model shows the two columns close together: when it says 55%, it should be right about 55% of the time — no more, no less. Rows with n below 30 are greyed: shown for completeness, too small to judge.
| confidence band | predicted (avg) | realized | n | |
|---|---|---|---|---|
| 30–40% | 38.2% | 35.3% | 266 | |
| 40–50% | 44.6% | 47.1% | 605 | |
| 50–60% | 54.5% | 51.1% | 393 | |
| 60–70% | 64.2% | 63.4% | 205 | |
| 70–100% | 75.5% | 62.9% | 89 |
Model vs market — reference facts
Top-pick hit rate is a reference fact, not the product. Closing markets are brutally efficient; beating the favourite's hit rate is not where value lives. Our terminal works through calibrated probabilities and extreme selectivity — the model abstains on roughly 99% of matches and flags only the rare case where its number and the market's number disagree enough to matter. We publish both columns anyway, including where the market wins.
The prediction industry sells certainty; we sell measurement. No profit promises, no highlighted winning streaks, no deleted losses. If a number here looks unimpressive, that is the point — it is real. The full instrument set (closing-line audit, distribution terminal, EDGE evidence board, replayable history) lives in the terminal.