Tag Archives: model

Improved Counterstrike Rating System

We’ve made some recent changes to our Counterstrike model to improve the way that ratings are updated. We are now accounting for the final score of each map to update each team’s rating, instead of simply looking at which team won and which team lost.

Note that when one team gains rating, their opponent loses the same amount of rating. This still applies in our new and improved rating system.

Under the new rating system, a team will gain more rating when they win by a higher margin – ie a 16-2 win will give a larger rating boost than a 16-12 win. In contrast, our old model would treat both of these wins equally. It makes sense to account for margin of win because the margin of win, on average, is indicative of the level of skill disparity between the teams. As such, factoring this into our rating system will give us more accurate and predictive ratings.

Note that there is still a jump that occurs when the winner changes from one team to the other. This is important, because although margin of victory can tell us something about the skill difference between teams, ultimately the most important predictive factor is the winner of the game. However, with the new system, there is proportionally less weight given to the match winner.

An interesting feature of the new system is that if an underdog puts up a strong performance against a much stronger team – but still loses – they will still gain rating (and conversely, the favorite loses rating). This makes sense, because the underdog exceeded expectations. We see this behavior reflected in the betting markets – and we now account for this in our model.

We can see here that if a 75% favorite wins 16-13, this will actually result in a small rating loss for the favorite, and a small rating gain for the underdog. In contrast, the old model would have given a rating increase to the favorite, which is counterproductive given their below-expectation performance.

In calibration, we are seeing that these changes give our model greater predictive ability. The changes are live in our online model and there are more improvements on the way. Feel free to swing by our discord if you’d like to discuss model features or chat with other esports enthusiasts and bettors.