How good are these forecasts?
Every forecast scored on seasons the model never saw while its settings were chosen.
Held-back seasons
| Forecast | Forecasts | Log loss | Favourite won |
|---|---|---|---|
| The model | 108 | 0.694 | 55% |
| Base rate | 108 | 0.692 | 48% |
| Recent form | 108 | 0.688 | 55% |
A team's home ground is the city it plays in most; it gets home advantage there and nowhere else. Matches abandoned without a result are left out.
A win by wickets is turned into runs (about eight a wicket) so the size of a win can move the rating.
Franchise T20 results are close to a coin flip before a ball is bowled: on the held-back seasons the model is about level with the base rate (see the track record).
Tested on the held-back seasons of the five franchise leagues together (566 matches) and left out, because none beat the ratings alone: player ratings from every men's T20 match (runs above average with bat and ball, or an Elo for each player's XI, which follows a player between leagues), the share of a side's XI kept from last season (what auctions and retentions change), one set of model settings for every league, and counting balls to spare as well as wickets in a chase's winning margin.
The toss isn't in the model: its winner went on to win 52% of these leagues' matches, but it is decided half an hour before the start, after the forecast is made, and on the held-back seasons it helped by less than the noise.
Cricsheet adds matches a few days after they're played, so the latest result can lag.