If a chess grandmaster beats a beginner 99% of the time, does that mean chess is 99% skill? And if an amateur in a coin-flipping tournament wins half their games, does that make coin-flipping 50% skill?
Every contest outcome sits at the intersection of three separate things: the skill gap between the competitors, the game’s intrinsic randomness, and contest length, meaning how many actions, rounds, or games dilute that noise.
Where Dota 2, CS2 Land
To measure where competitive games actually sit between pure chance and pure skill, I looked at real match records rather than simulated ratings: 78,722 ticketed Dota 2 maps from OpenDota (filtered to the 14,735 maps and 6,683 series across 114 pro-circuit leagues connected to The International) and 26,155 Counter-Strike 2 series from Valve’s official regional standings repo.
To put them on the same scale as traditional sports and board games, I fit the competition-depth framework of Jerdee and Newman (Science Advances, 2024). In their model, depth measures how reliably the stronger player wins when you draw two typical opponents from the competitive pool.
Here is what the real numbers show:
- Dota 2 (Single Map vs Series): In an isolated map between two teams at The International 2025, the stronger team wins 62.5% of the time (95% CI: 58.2% to 68.3%). Across the broader pro circuit (45 active teams with 20+ games), that rises to 65.2%. Expanding to a full Best-of-3 series filters out early map variance: the typical favorite win rate jumps to 67.6% among TI teams and 68.3% across the pro circuit.
- Counter-Strike 2 Series: Among Valve’s global top 20 teams, the typical favorite wins 66.8% of series (CI: 62.2% to 71.9%). Across all 176 ranked teams in Valve’s ecosystem, that rises to 73.4% (CI: 71.7% to 76.1%), because the open standings pit regional qualifier teams against world-class rosters.
- Reference Benchmarks: Evaluating Jerdee and Newman’s model with its fitted posterior-mean luck term yields favorite win rates of 61.5% in Scrabble tournament matches, 65.3% in NBA regular-season games, 68.3% in online chess on Lichess, 71.9% in ATP tennis matches, and 74.1% in competitive Super Smash Bros. Melee sets. Factoring in fitted luck moves the NBA from level with Dota 2 series down to level with Dota 2 single maps, and every real competitive game clusters tightly between 61% and 75%.
The Parity Paradox
If professional play has so much strategic depth, why do championship matches between elite teams feel so unpredictable?
This is the Parity Paradox: the higher the competitive level, the closer match outcomes contract toward a 50/50 coin flip, precisely because both competitors play near perfection.
In my out-of-sample Elo evaluations across the 2025–26 season (tuning the update rate strictly on prior years):
- In Dota 2 single maps, the pre-game favorite won 61.4% of 4,490 matches.
- In Dota 2 series, the favorite won 65.9% of 1,628 series.
- In CS2 series, the favorite won 62.8% of 11,450 series.
In both games, about 44% to 45% of all professional matches featured a pre-game favorite below 60%. Outside of heavy mismatches, elite esports operate in a narrow band between 50% and 65%. When two top-5 teams play each other, single tactical moments, such as a pistol conversion or a Roshan fight, swing the outcome.
Past work
In 2018, an MIT group led by A.E. Hosoi published Luck and the Law in SIAM Review (Getty, Li, Yano, Gao, and Hosoi), about which I found out while listening to a podcast. They introduced a split-half persistence metric to quantify skill in professional sports and fantasy contests for legal proceedings.
The authors pooled five seasons of real sports data (2010 to 2014), giving each team five points on their plots:
- NHL Hockey: 0.54 (82-game season)
- MLB Baseball: 0.63 (162-game season)
- NFL Football: 0.68 (16-game season)
- NBA Basketball: 0.86 (82-game season)
When computed on the 2025 esports season for teams with at least 30 games, Dota 2 maps score 0.61, and CS2 series score 0.69.
It is tempting to look at those numbers and conclude that CS2 series (0.69) involve more skill than NFL football (0.68) or MLB baseball (0.63), but it’s wrong. The authors explicitly noted that “the measured value of depends on the number of games associated with each player.” Because binomial sampling noise shrinks with sample size, true talent variance becomes progressively easier to detect as more games are played.
Yet in their cross-sport comparison (Figure 6), the authors left the metric unadjusted for season length. They recognized that the 16-game NFL scored unexpectedly high relative to leagues with 82 or 162 games, calling it “somewhat surprising since there are only 16 games per team in the NFL season” and leaving it as “a puzzle for future investigation.”
If you normalize by season length to calculate the per-game skill signal , the sports ranking reverses entirely:
- NFL Football: 0.133 per game (16 games)
- NBA Basketball: 0.075 per game (82 games)
- NHL Hockey: 0.014 per game (82 games)
- MLB Baseball: 0.011 per game (162 games)
NFL football has the highest per-game skill signal, but its short 16-game schedule kept its R^_ at 0.68. MLB baseball has a tiny per-game skill signal, but its massive 162-game schedule inflated its R^_ to 0.63.
To prove this on real esports data, I took the exact same teams in the 2025 season and recomputed after thinning each team’s schedule to a random subset of games:
- Dota 2 Maps: At 30 games per team, is 0.38. At 60 games, it reaches 0.52. At 160 games, it climbs to 0.69.
- CS2 Series: At 30 games per team, is 0.51. At 60 games, 0.62. At 80 games, 0.66.
measures the statistical power of a specific schedule to separate a specific population. It cannot measure the intrinsic randomness of a rulebook.
Citations
- Getty, Li, Yano, Gao, and Hosoi (2018): “Luck and the Law: Quantifying Chance in Fantasy Sports and Other Contests,” SIAM Review, 60(4), pp. 869–887.
- Jerdee, M. and Newman, M.E.J. (2024): “Luck, skill, and depth of competition in games and social hierarchies,” Science Advances, 10(45), eadn2654. DOI: 10.1126/sciadv.adn2654.
- Valve Corporation (2013): “Matchmaking,” Dota 2 Blog, December 6, 2013. Archived capture.
- Valve Corporation (2024–2026): “Counter-Strike Regional Standings,” GitHub Repository.