How to choose the Elo K-factor

K is the most important setting in any Elo system. It caps how many points one game can move a rating. A high K reacts quickly to real changes in skill but also to luck. A low K is stable but slow.

What K does, in numbers

One game can change a rating by at most K. Between equal players, a win is worth K/2. An upset against a player rated 200 points higher is worth about 0.76 × K.

What the big systems use

FIDE (chess)

• K = 40 for players new to the rating list until they've completed events with at least 30 games, and for players until the end of the year of their 18th birthday while they're rated under 2300.
• K = 20 while a rating stays under 2400.
• K = 10 once a published rating has reached 2400, even if it drops back below.
• If K × games in a rating period would exceed 700, K is cut to the largest whole number with K × n ≤ 700.^[1]

The idea is simple: new and young players move fast, and established elite players move slowly.

US Chess

US Chess calculates K for each player and event as K = 800 / (N′ + m), where m is the games played in the event and N′ is the player's "effective" number of past games:^[2]

So K shrinks as a player builds a record, and lower-rated players keep a higher K for longer. The maximum N′ by rating, from the formula:

The specification's own example values: N′ = 6 and m = 4 give K = 80, N′ = 20 and m = 4 give 33.33, and N′ = 50 and m = 10 give 13.33.^[2]

FIFA men's world ranking

FIFA's version of K is called I, for "importance". It ranges from 5 for friendlies outside international windows to 60 for World Cup matches from the quarter-finals on, with values for Nations League, qualifiers and continental finals in between.^[3]

FiveThirtyEight's NFL model

FiveThirtyEight's published NFL Elo code used K = 20, multiplied by a margin-of-victory factor, with 65 points of home-field advantage.^[4]

Lichess and Chess.com

These sites don't use a fixed K. Their Glicko systems scale each update by how uncertain the ratings are, which works like a K that adjusts itself.^[5] See Elo vs Glicko-2.

Choosing K for your own league

These are our rules of thumb, not official rules. Check them against your own data.

• Few games per player (under 20 a season): K = 32 to 40. Ratings need to get somewhere useful quickly.
• Regular play (a few games a week): K = 20 to 32.
• Many games and a stable group: K = 10 to 20. Less noise from lucky streaks.
• Games with lots of luck, like dice or card games: use a lower K so one game doesn't swing the table.
• New players: copy FIDE's idea and give newcomers a higher K, such as 40, for their first games.
• Test it. Replay your past results with two or three K values and see which one would have predicted the next results best. With the league tool, you can change K and every match is replayed.

Related