SKREAM Prediction System
Messy Fractals’ SKREAM (Simulating Kabaddi by Regressive Estimations for Adaptive Models) develops probabilistic projections for every match in a season based on individual player performances and their effective contribution to their teams. The system uses a unique player rating system, where each player’s raiding and tackling abilities are calculated based on a variety of metrics.
Each raid, each tackle, and each minute spent on the bench will change the way the game unfolds. To predict match win probabilities we need to understand how players perform head to head. In Kabaddi, this means understanding how a raider performs under different defense scenarios. For instance it might be much harder for raiders to perform with 3-4 tacklers in the defense compared to a 6-7 member defense.
Using 2000+ raids and tackles digitized from Season 2, we predicted the relationship between Raid Strength, Tackle Strength and consequently the points scored in that raid.
Check out win and loss projections and playoff odds for the ProKabaddi league.
When we looked at the data available to us, provided by the ProKabaddi League and some data that we collected by ourselves, it was messier than what we expected and realized that a single overall player rating was not enough to make holistic and accurate predictions.
For a raider, we used the following data:
- RPPM Raid Points/Total number of Matches played
- PPR Raid Points/Total number of Raids
- RPM No. of Raids/Total number of Matches played
- ERTR No. of Empty Raids/Total number of Raids
- SRTR No. of Successful Raids/Total number of Raids
- URPM No. of Unsuccessful Raids/Total number of Matches played
For a defender, we used the following data:
- TPPM Tackle Points/Total number of Matches played
- PPT Tackle Points/Total number of Tackles
- TPM No. of Tackles/Total number of Matches played
- STPT No. of Successful Tackles/Total number of Tackles
- UTPM No. of Unsuccessful Tackles/Total number of Matches played
Based on this, SKREAM evaluates raiding and tackling ratings for the players using multi-linear regression models and they were significantly consistent with their real life performances. These ratings also account for the Form factor of the players. That is, the performance of a player (both defensive and offensive) in his most recent match will carry more weight when compared to his performances in the previous matches.
The system takes these ratings to the next level by running several thousand simulations accounting for match conditions, team specific strengths and weaknesses etc., and provides probabilistic outcomes for the matches.
The matches were simulated with these assumptions:
- Raiders with higher Raid Strength have a higher probability of conducting raids.
- Defense players out in successful raids are selected randomly.
- On winning a match, the player strengths for all players in the team increased by 1 percentage point
In the end, match results are decided by the interaction between 14 players across 80 raids. But over a period of 56 matches, predictions are bound to make sense