Ball-by-ball records for the 5 grounds this competition uses, 103 matches in total, plus live AI win probability updated after every ball once a match starts.
Across 103 matches at these grounds a first innings averages 154.6 and a chase 138.2, at 7.32 runs an over. On first-innings scoring that places the SA20 4th among the 10 T20 leagues we hold ground data for, against PSL at 172.4 and BPL at 139.2.
The league average is the least interesting number on this page. Chasing costs 40.4 runs at The Wanderers, Johannesburg and only 13.0 at Newlands, Cape Town. The highest-scoring ground on the list is The Wanderers, Johannesburg at 8.17 an over. A total that wins comfortably at one of these grounds loses at another.
| Ground | Matches | 1st inns | Chasing | Gap | RPO |
|---|---|---|---|---|---|
| The Wanderers, Johannesburg | 5 | 183.6 | 143.2 | +40.4 | 8.17 |
| St George's Park, Gqeberha | 10 | 150.8 | 127.2 | +23.6 | 6.95 |
| SuperSport Park, Centurion | 24 | 163.8 | 148.5 | +15.3 | 7.82 |
| Kingsmead, Durban | 33 | 149.2 | 134.6 | +14.6 | 7.09 |
| Newlands, Cape Town | 31 | 149.9 | 136.9 | +13.0 | 7.17 |
These are T20 records for each ground rather than for this competition alone, so treat them as ground character. Sample sizes vary a lot — a ground with five matches behind it deserves far less weight than one with a hundred, and both are in the table rather than hidden behind an average. The live model does not use these figures: once a ball is bowled it reads the actual over-by-over run and wicket pattern instead.
It starts when the match does. There is no pre-match tip here. Win probability is recalculated after every ball from the score, wickets in hand, balls and runs remaining, and how the pitch is behaving in this innings rather than on average.
Measured, not claimed. 81.5% win-probability accuracy on a true holdout: trained on 2017-2024 matches only, then scored on 2,546 matches from 2025-26 the model had never seen. Accuracy rises through an innings, from 79.2% at over 6 to 82.2% by over 15.
Next-over forecasts too. Expected runs for the coming over land inside the published range 72.3% of the time, measured across 46,718 over-boundary snapshots.