The distance
How close are two complete 1X2 forecasts?
Total Variation distance measures the minimum probability mass that must be redistributed to turn one home-draw-away split into another. Zero means identical probability shapes.
54 / 26 / 20 versus 51 / 29 / 20 = 3 probability points apart
The as-of boundary
Future results are impossible inputs.
The cohort is frozen to evidence that existed strictly before the target publication timestamp. A closer match that settles tomorrow cannot enter yesterday's comparison.
Selection uses only the probability vector, model version, league and settlement time. The actual result, hit flag and Brier score are revealed only after the neighbours are fixed.
Fixed expansion order
The search cannot widen because a result looks flattering.
- 01
Same league + model
Start with the narrowest operational context. Keep it only when at least eight eligible records exist.
- 02
Same model
If the first scope is too small, expand across leagues while preserving the model version.
- 03
All public records
Only then use every eligible model and league. The displayed scope always states which level supplied the sample.
The method first looks inside a fixed 10-point Total Variation window and returns at most 12 records. If fewer than eight exist, it expands deterministically to the nearest eligible records and exposes the furthest distance instead of hiding the expansion.
Read it honestly
Historical neighbours provide context, not a second prediction.
Outcome counts
Home, draw and away totals show what happened in the fixed sample. They do not replace the current locked probabilities.
Mean Brier
One shared scoring definition makes the selected historical records auditable, while small samples remain clearly marked.
Every source row
Each neighbour links to its permanent match record so misses cannot disappear behind an aggregate sentence.