Rating 3.0 receives small accuracy improvements

Rating 3.0 has received a small update to improve the accuracy behind its calculations.

Since it was released, 29,030 maps, including 9029 on LAN, have been played on HLTV’s database and that means we have a wealth of information and context to add to our underlying logic behind rating.

3.0, unlike previous versions, has specific averages for each map. That means that the win probability percentages for a player in the same 5v5 scenario is as different as nearly ten percentage points on Anubis compared to a CT-sided map like Overpass. For advanced statistics, this context is crucial in order to make round swing and eco-adjustment fair on all players.

New maps like Cache were therefore a problem for 3.0, since there was no data on it in CS2, and we had to fall back to global averages for its games. Cache is CT-sided in buy rounds (T 47.1%), but before the update, every round on the map was scored as if it were an average map.

The odds of the T side team winning in 5v5 situations per map

Because it is so new, we are still dealing with a fairly low sample for Cache. But when there are not enough examples of an exact scenario (e.g. full buy vs full buy, 3v2, postplant) we will once again revert to global averages. This is decided per round, so a round never switches between map-specific and global numbers midway through.

For something like round swing, the extra year of data is extremely important. To reliably know the win percentage of a 4v4 post-plant on Inferno, we want as many rounds as possible in that situation. By adding those extra games to our dataset, round swing and eco-adjustment will become more reliable.

This approach was also a problem because Valve tweaked the economy just as development finished for rating 3.0. Since July 2025’s changes made the game more CT-sided, our rating could not fully reflect that as the old averages contained only a few months of games played under the new economy.

We are also pushing some changes to how round swing works. After this update, rounds are considered over when the bomb cannot be stopped from exploding: This means kills when the CTs don’t have a kit, and the bomb has fewer than 10 seconds left, are considered the same as exit frags, and the same is true for CT deaths.

This means rounds like this one from Alistair “⁠aliStair⁠” Johnston no longer give disproportionate amounts of swing, since the round was decided before he was left last alive.

A bug has also been fixed where players’ round swing went down after a frag. In some niche scenarios, usually with low samples, win probability for a team actually reduced after a kill. But with extra data behind these scenarios, this is no longer as common.

In the unlikely scenario where win probability still decreases after a kill, a player now receives 0 swing rather than negative, as the before and after scenarios are merged into one win probability.

With these changes, it was also time to update the averages and standard deviations used to calculate the rating, to ensure an average of 1.00 per player over a single event.

Together, this will cause small changes to old ratings. Across all 6,687 CS2 maps played between July and September 2026, a single-map Rating 3.0 moved by only 0.021 on average. 9.4% of player-maps moved by 0.05 or more, 1% by 0.10 or more, and the largest move was 0.31.

Cache, naturally, is the most affected map now that we have data for it specifically, while Inferno’s post-plant heavy nature means it is the most impacted by our bomb timer changes.

Overall, though, we believe that this trade-off is worth it for greater accuracy ahead of the 2026 edition of the HLTV Awards and Top 20 Players of the Year.

For the full patch notes, read here: