Backtest performance

Honest chronological holdout: each match is predicted using only ratings from earlier matches. Hyperparameters are grid-searched on the same post-cutoff reporting set for this first build, so treat these as baseline diagnostics, not a locked contest-grade out-of-sample score.

Out-of-sample Brier

0.2121

Accuracy

65.4%

Coin-flip baseline Brier

0.2475

Matches scored

1676

Data and tuned parameters

Source: Cricsheet normalized completed men's matches; chronological Elo holdout from 2024-01-01 onward
Holdout: chronological out-of-sample from 2024-01-01
K: 40
Home Elo: 0
Temperature: 0.9
Cross-format carryover: 0.08

Metrics

Format	Matches	Log loss	Brier	Accuracy
TEST	89	0.5878	0.2032	65.2%
ODI	222	0.7003	0.2480	55.9%
T20	1365	0.6040	0.2068	66.9%
ALL	1676	0.6159	0.2121	65.4%