How the 5overs win chance works
Every grade in 5overs comes from a win chance. Here's how it's worked out, and where its limits are.
1. Real ball-by-ball data
The game uses ball-by-ball records of 1,243 IPL matches from 2008 to 2026, published by Cricsheet. For every bowler and batter, it builds a record of what happens in overs 16 to 20: dots, singles, twos, fours, sixes and wickets.
2. Small samples are treated with care
A bowler who has bowled three death overs might have gone for 40 or for 15 by luck alone. So each player's record is blended with a typical record for his role, and the fewer balls he's bowled or faced, the more the typical record counts.
3. Each ball combines bowler, batter and situation
Batters swing harder when they need 14 an over than when they need 7. The game uses how real chases went at each over and required rate, then adjusts for the bowler and batter.
4. Thousands of simulated finishes
To get a win chance, the game plays out the rest of the innings hundreds of times, ball by ball, and counts how often the fielding side wins. To grade your decision, it does this for every option you could have chosen.
5. Checked against what really happened
The model was tested on 1,042 real chases from the start of the 16th over. Its Brier score, a standard measure of prediction error where lower is better, was 0.102, against 0.245 for always predicting the average. The displayed chances are then calibrated so that "60%" really means about 60%.
6. The match setting and the batter's innings
Every ball's odds also take account of things the data does record, all fitted together with the player ratings so nothing is counted twice:
- Season. Death-over scoring changes from year to year, and has risen sharply since 2024. Each puzzle uses its own season, and player ratings are adjusted for the era they played in.
- Ground. Bengaluru, Dharamsala and the Wankhede give up the most boundaries at the death; Chennai, Dubai and Sharjah the fewest.
- How long the batter has been in. A batter on his first few balls hits about a third fewer boundaries than one who's set.
- How each batter handles the asking rate. Some go harder when 12 or more an over are needed; some don't.
Each of these was checked on matches held back from the fitting, and each one improved the predictions.
7. What's modelled, not measured
Public ball-by-ball data doesn't record line, length, shot direction or where fielders stood. So the deliveries are built from published analysis of IPL death bowling: the economy, boundary and dot-ball rates of each length actually bowled (yorker, good length, back of a length, full, short, full toss), and how often bowlers land what they intend. The best bowlers land their plans more often, judged on their control and how rarely they're hit for boundaries. How a wicket falls follows each bowler's real split of caught, bowled, lbw and stumped. Shot directions and field effects are still our model, and each bowler's usual mix still reproduces his real record. Each batter's record against right-arm and left-arm pace and spin, how he usually gets out and how often he clears the rope do come from the data, so matchups and his strengths and weaknesses are real. Batting hand and bowling style come from Cricinfo player lists, and left-handers and left-armers are shown the right way round.