← Research

How predictable is a ranked game before champion select?

Published October 7, 2026 · NA ranked solo queue, 457,284 held-out games of patch 16.18 (September 2026); EUW and KR as an early read

What the model is allowed to see

Everything that exists when the queue pops, and nothing after it: no champion, no ban, nothing from the game itself.

• The ten players’ rank and LP, and our own rating built from their match history.
• Each player’s last 50 ranked games: won or lost, how they performed, how long the games ran, how their lane went, how strong the opponents were. Only games that ended before this one began.
• The role each player was given, how much of their recent play is in that role, and whether they were autofilled.
• The champions each player usually plays in that role: what they are likely to pick, never what they did pick.

The model was trained on earlier patches and tested on 457,284 NA games of patch 16.18 that it had never seen. Every NA number below comes from those games.

How often it is right

The same 457,284 test games, with more of each player’s history added one step at a time, in the five measures we report for every model (the accuracy page explains each). A coin flip is 50% accuracy and an AUC of 0.5.

Rank alone gets nowhere. The gap between the two teams in our own rating predicts the winner no better than a coin (AUC 0.502; all five measures are in the loser’s queue study), because that is what the matchmaker evens out. Nearly everything the model knows comes from the players’ recent games. The last row is the same model with its confidence corrected on games set aside for that purpose; the sections below use it.

Low elo is far more predictable

In Iron the model calls the winner 69.4% of the time before champion select. In Gold it is 61.6%, and from Platinum up about 60%. We did not test why. One reading that fits: a rank says less about a player the less settled their account is, and lower tiers hold more accounts whose rank is still catching up with how they play.

When it names a favorite, the favorite wins that often

A model can be right often and still overstate how sure it is. This one doesn’t: lobbies it calls 60–65% are won 62.3% of the time, and the 17.4% it calls 70% or more are won 77.5% of the time. Only 26.5% of lobbies are within five points of a coin flip.

Even on rank, uneven on form

We split every forecast into the parts it was built from. This is how much of the difference between one lobby and the next each part carries:

Rank explains almost none of it, because rank is what the matchmaker balances, and it balances it well. What is left is what a rank moves too slowly to show: who is on a run, who is in a slump, who has just come back.

Read “recent games” with care. Riot matches on a hidden MMR that we never see. A player’s recent games are exactly where a gap between that MMR and their visible rank would show, so part of this may be MMR differences the model recovers, not only hot and cold streaks. These data cannot tell the two apart.

Off-role players are not evened out

A player counts as off-role when the role they were given makes up under 20% of their recent games. When one team has more of them than the other, it shows in who wins:

Each extra off-role player costs a team about three points of win rate, and only 37.9% of lobbies have the same number on both sides.

Could the same ten players have made a more even game?

For every lobby we kept the ten players and their roles, and asked the model about every other way to split them into two teams. The two players in each role can trade sides, which gives 32 possible splits. Then we looked at where the real one falls among them.

The real split sits in the middle. It is more even than the average of the 32 in 55.3% of games, and its typical place is number 16 of 32. By what this model can see, the split a lobby got is about as even as a random split of the same ten players.

The table shows the share of NA lobbies that would still have a favorite at 60% or more if the most even option had been played. “Duos kept together” never separates two players who have queued together before; 71.4% of games hold such a pair.

On paper, trading a single pair takes the lopsided lobbies from 48.6% to 4.2%, or to 10.7% with duos kept together.

This is the model marking its own homework, so read it as a ceiling. The most even of 32 estimates is flattered, because some splits only look even where the estimate happens to be off. Those games were never played, so nobody can check them. And every extra condition costs a real matchmaker queue time, which this ignores. What the table does show is that the information needed to spot a lopsided lobby exists before champion select.

Half of all lobbies cannot put everyone in a role they know

We also asked whether the ten players could be seated so that everyone plays a familiar role, meaning at least 20% of their last 50 games. Only 20.3% of real NA lobbies already are. A fully familiar arrangement exists at all in 50.2% of lobbies; in the rest, the ten players’ roles simply do not cover every position twice. Everyone on their main role is possible in 5.7%. That limit is set by who is in the lobby, not by how it is split. Where a familiar arrangement does exist, choosing it costs almost nothing in balance.

EUW and Korea: an early read

The history model, trained on NA only, also scored 18 days of games from EUW and Korea. It knows those players far less well (8 to 10 earlier games each, against months in NA), so it finds fewer favorites there. That is a weaker instrument, not fairer matchmaking. What does carry over is the comparison inside each lobby: in all three regions the real split is typically number 16 of 32. The table uses the history model alone in every region, NA included, so its NA row is slightly below the full model’s.

Still no loser’s queue

We re-ran the loser’s queue study with this stronger model, and nothing changed. After five or more wins in a row, players won their next game 53.0% of the time; after five or more losses, 47.9%. The model expected both, and a player on a win streak still gets slightly stronger teammates, not weaker ones.

What this can’t tell you

• One region in depth, one patch. NA, patch 16.18. EUW and Korea are an early read from a model that was not trained on them.
• This is not Riot’s MMR, and not a look inside the matchmaker. It measures lobbies as they arrived, from public match history. The matchmaker balances something we cannot see, and by our own rating it balances it almost perfectly.
• The model reads every game up to the one it predicts, including earlier games the same day. A system refreshed once a day would know a little less.
• Role preferences are inferred from what each player has played. The roles they queued for are not public.
• No player is named or rated here. Every figure is an average over hundreds of thousands of games; your own lobby is one draw.