How OddsIntel's WC predictions are doing

Methodology. For every settled WC2026 match, we record our model's 1X2 probability triple (sources national_team_v1 and national_team_v1_blended — the blend with market wins when both exist), the market consensus distribution from wc_market_consensus (avg of free public sources, vig-removed), and the actual outcome.

Brier score is the mean squared error between the predicted probability vector and the one-hot actual outcome (e.g. (0.55, 0.27, 0.18) vs (1, 0, 0) for a home win → 0.45² + 0.27² + 0.18² = 0.30). Lower is better. 0 means perfect; a uniform 1/3 guess scores ~0.667.

Log-loss is −log(p) on the eventual outcome, averaged over matches. Penalises confident wrong calls more than Brier does — a model assigning 1% to the eventual winner gets ~4.6 in that row alone.

CLV vs market is the average difference between our probability and the market's probability on the actual outcome. Positive = we leaned the right way harder than the market. Computed only on matches where both rows exist.

Calibration plot buckets every predicted probability (3 per match — home / draw / away) into 10pp bins, then plots the bucket's mean prediction next to the bucket's actual hit rate. When the model is well-calibrated, the amber bars track the dashed diagonal.