Every prediction is written down and archived before kickoff, then reconciled automatically against the final score. Nothing is deleted and nothing is edited — you can check the record yourself.
Every published prediction is archived before kickoff. When the match finishes, the result is reconciled automatically against the stored probabilities — nothing is deleted or edited.
Model performance
v4.0 benchmark — frozen evaluation (train 2011-2024, feature history from 2008 / val 2025 / one-shot test 2026+). The live accuracy above comes from real pre-kickoff predictions logged since 2026-10-18; this section is the model's reference score.
Confidence-gated accuracy
A match is flagged high-confidence when its top probability clears a validation-selected threshold. Accuracy is always paired with coverage.
Month by month
On targetWatch
Calibration
When the models put a match in a given probability band, how often that outcome actually happened.
Method: each row is written to the database before kickoff with the exact probabilities served on the site at that moment. When the match ends, the stored numbers are compared with the actual result by a scheduled job — there is no manual step and no way to revise a past entry.