Predicting HLTV's top 20 players of 2025

HLTV's top 20 players lists round out every year of Counter-Strike. Whether you agree or disagree with the ranking philosophy behind them, they inarguably hold a lot of cultural weight when thinking about the greatest players in the past and present. Many newer or more casual viewers may use them as their only source of information in discussions about the greatest players of all time. For that reason, it's crucial that the metholodogy they employ is transparent, principled, and consistent. To their credit, we think the analysts over at HLTV do a great job of this, which raises the question: just how replicable is their process? Is it so predicatable that we can accurately guess which players will be in this year's top 20, and where? Let's find out!

Our goal here is simple — we are testing how accurately our machine learning models can predict this year's top 20. To start, we will collect a bunch of (potentially) useful metrics that we might use as regressors. According to HLTV themselves, statistics like KPRW (kills per round won), their own HLTV Rating 3.0, and MVP/EVP awards are heavily factored into how they judge players and their performances. But high stats alone are not enough to earn high positions; they are weighted by how prestigious the event is that the stats are from, and they reward longevity and dominance over a purple patch of play. By training supervised learning models on historical rankings, plus the data that helped produce them, we can try to mathematically replicate HLTV's values and biases.

Two goals, two models

This challenge is best broken down into two parts. The first: which players are most likely to be in the top 20? The second: how are they likely to be ordered? In our experimenting, we found that two different model architectures with slightly different feature sets are optimal for each.

Model A — a binary (yes/no) classifier for "will this player be in the top 20?" — prioritised EVPs, K/D Differential, HLTV Rating, MVPs, and total rounds, out of the dozen metrics we provided. This paints a clear picture; to get into the top 20, you must have a large enough sample size and have either won some tournaments or been one of the best players in some tournaments. Makes sense, and aligns with everyone's intuition. This model was scoring precision and recall ~0.91, meaning it was able to consistently identify 18 or 19 players out of a possible 20 on the historical data. Not too shabby.

Model B — a pairwise learning-to-rank (LTR) model — was much more interesting. HLTV Rating was still a considerable factor, but team ranking, number of prestige tournaments won (Katowice, Cologne, Majors), and playoffs K/D suddenly flew into scope. This is a curious discovery: getting into the top 20 seems to be more about individual merit, while your placement within the top 20 depends on how well your team did throughout the year, which trophies you lifted, and how well you performed under the brightest lights. This means that a player like donk (or s1mple in the past), who has arguably the greatest skill ceiling but might lack the silverware to support it, is predisposed to a lower ranking than players like ZywOo, who are still incredibly talented but have a far stronger supporting cast. This model averaged a Spearman's Rank value of ρ = 0.882 and MAE ≈ 1.8 on the historical data. This is a very strong performance, and in layman's terms, when it does get the order wrong, it's off by one or two positions at most.

Predictions

Given the theoretical strengths of this model, we're excited to see how it matches up with reality over the next few days. It was very confident in its selected top 19, giving them all a >74% of making the cut. That 20th spot, however, it wasn't so sure about — 910 got the final nod, with a 51.50% chance of claiming it, barely ahead of yuurih, iM, kyousuke, and YEKINDAR, who all averaged ~47%. Honestly, we're pretty torn on that too.

The other point of contention is clearly the number one spot. We spoke earlier about the model's learned biases, and it has demonstrated them perfectly here; donk out-performs ZywOo in every relevant stat at big events, including in playoffs and grand finals. Yet, it still thinks ZywOo has a better chance at claiming the throne, due to his team making deeper runs and lifting more trophies this year. People are calling them the greatest team of all time, and surely the best player on a team like that must be the best player in the world, right?

Well... device would like a word.

If you weren't already aware, you can also play along at home. The deadline is December 24th, so make sure to lock in your picks before then.