Creating an ELO Formula to Evaluate CS:GO Players After a Tournament

CS:GO Players

About 20 years ago, the Oakland Athletics baseball club had a man who was phenomenally calculating player utility.

After a long time, the technology that made the Athletics a phenomenal team has become cutting-edge.

In each discipline, there are more and more services with statistics: various ratings and indicators are trying to identify the strengths and weaknesses of the CS:GO players, “scatter” in terms of strength.

 In CS:GO, for the last ten years, has provided the fairest comparison of all players in the world, regardless of position (the so-called HLTV Top-20 of the Year).

The populist method of matching players “here and now” is based solely on K/D, ADR, KAST, Rating 2.0, and others.

As in mathematics, we lose the roots that lead further from the truth in this case. The idea of ​​the ELO formula came to me in the summer and was corrected experimentally (and will be for many more months) to the current, more stable state.

The bottom line is this: fit as many nuances as possible into a convenient calculation. The author of Sports recently proposed an exciting technology.

Still, the community remained dissatisfied with the absence/inclusion of certain values ​​(for example, there was no ADR role in that rating).

Our ELO formula will be based (don’t laugh) on the MVP formula for the PUBG Mobile championships.

True, from there, we take only one very important detail – the percentage. Questions about whether it is necessary to include certain types of numbers with HLTV should disappear if we proportionally lower their value (unnecessary – more, necessary – not so much).

The guys from PUBG made sure that only 40% of survivability and damage are considered in the formula and 20% of kills.

Full Version Of All Data For The Formula

Do not try to delve into it yet; we will analyze the scheme step by step. It consists of two parts – color (only the qualities of the player and teammates) and white (comparison with other players). 

Bold font – calculations already made for s1mple at the November IEM Beijing-Haidian; they are needed for clarity.

The first part of the big formula (which you don’t see yet) is the individual skill metrics: We pay attention only to the upper but multi-colored cap.

The total amount of these percentages is 100. This is the integer part, which we break down by Utility (25%) – controversial parameters taken at a small percentage in general terms. They may not give an iron impact for the player, but they should not be left without attention either;

 Shooting (40%) – the most common and most important K/D, adjusted to the realities of the formula (i.e., multiplied by 100, as you can see).

The overly confusing percentage of headshots was removed from the summer version; Clutches (20%) – the big picture above explains how the score for clutches is calculated after the tournament;

Stability (15%) – may be excluded in future versions because. It is very different for all players and vaguely represents the big picture.

Next, look at how the calculations of a real player are carried out in bold according to these instructions. Why multiply/divide the original values? To arrive at beautiful 142, 82, 50 and avoid unnecessary zeros or tenths.

 We pass to the lower cap, where the necessary shares have already been found for the obtained values ​​(20.5 from the utility, 56.8 from shooting, etc.).

 As you know, the individual game of the conditional s1mple is only 70% of the big formula indicated on the header. Therefore, summing up those same shares, we also lower their threshold to 70%, getting 60.5.

 IMPORTANT! We deliberately lower all the values ​​many times so that the players’ gap does not look critical. If any significant parameters prevail, they have a chance for a comeback in the final ranking.

 The same algorithm will work in terms that do not depend on the individual actions of the player:

The value of cards is a new thing that few people pay attention to. Do you remember the cases when a player who played well and was eliminated at the beginning of the tournament remained at the top?

Now they will be minimized thanks to the “card value.” Tier odds are also explained in the most extended screenshot of this thread;

Comparison with teammates is a very relative and controversial parameter, which depends on the player’s role in the team. Therefore, it was taken only under 40% out of a hundred.

 IMPORTANT! Our rating is individual in the first place. This explains dividing large amounts of data into two groups: significant for the formula (70%) and auxiliary (30%).

It’s Time To Put The Puzzle Together

This is a “paper” view of everything described above (except for the final APC). As a result of reduction according to all school rules, the ELO formula gets a compressed form:

From now on, everything except APC, we can calculate using this system, knowing the unknown notation:

  • CS – scores for clutches from the general scheme;
  • RD – stability, which came from a rating of 2.0;
  • M*TK/P – value of cards (number of played cards, multiplied by the tier-caf and divided by the place in the tournament);
  • PR-ATR is the difference between the 2.0 rating of the taken player and the 2.0 average rating of his team.

Knowing these notations, you can independently take the compressed formula into service.

IMPORTANT! Without APC, the formula loses its meaning. The essence of APC (Average player’s coef) is to show how much one player performs better than other teams’ players in the same role.

Step By Step:

  1. 16 teams are participating in the tournament (IEM). For s1mple, it’s him plus 15 snipers (for flamie, it’s him plus 15 preplant defenders, etc.)
  2. If among all snipers, s1mple turned out to be the best statistically, then he gets 16 APC points, if the second – 15, etc.
  3. For each side, as you can see, APC is taken separately. It’s no secret that the game’s roles and very essence change dramatically depending on the side, so it’s pointless to fit the conditional Nivera into one class for T and CT.
  4. We take the arithmetic mean from the APC for defense and attack, which will enter the compressed form of the formula (s1mple on IEM turned out to be the best terrorist sniper and the second special forces sniper. The average is 15.5).

What did I take into account when defining the roles? The table at hand changes during the competitive season (information is taken with an emphasis on Nuke – the most popular map of the second half of 2020).

Estimated Player Positions For Apc

It is more recent because. It was adjusted during the BLAST Premier Finals, considering the latest changes.

Are there any identified cons? Metrics such as “Stability,” “Utility,” and “Card Value” are still quite raw, which can spoil the truth of ELO.

Why is this technology fairer? Because the main role is played by comparison with peers, it does not require dividing the top into “snipers,” “riflers,” etc., adjusting the numbers to one competitive value.

How can it help in the analysis? It is easy to guess which indicator is dragging ELO to the bottom with manual calculation. If this is insufficient APC, the problem is somewhat more global than 70% and 30% of individual and near-individual values ​​with HLTV.

Very fast application in practice (in the future, we will expand separate materials on the results of tournaments):

Let’s look at completely different examples of the top three players in BLAST Premier Finals (according to HLTV) and give some explanations:

  • ZywOo broke away from s1mple due to 70% and 30% parameters. When multiplied by each other, these “caps” give a significant difference, although the APC of the eternally arguing heroes of the era is the same (7.5);
  • In the case of misutaaa harder, ELO messed up the “value of the cards” (namely, the detail responsible for his played encounters). Because  Kevin is a constant rotation player who misses half the cards, the formula “cuts off” a decent share of his ELO. If we calculated his competitors with a less cool APC (misutaa outstripped everyone on both sides), then the same dupreeh and tiziaN could get a lot more ELO and get into the top.

Although the system has many pitfalls due to its complexity, it is one of a kind. The community will only welcome suggestions for adjusting the device to get one of the most honest discoveries in the discipline in a year. Love CS!

Recently, the author launched a Telegram channel (, where he posts mini-posts of a similar style and the results of his observations.

Last time I reviewed in detail the IZA statistics, which I also invented myself. Soon there will be explanations about the roles in the ELO formula and answers to questions that have arisen.

Leave a Reply

Your email address will not be published. Required fields are marked *

You May Also Like