6 Comments
User's avatar
TheMouth's avatar

How is margin of victory (mov) implemented in your model? Is (3 + s)^.85 applied as a scalar to your k value or as an addition to post-game elo change?

In other words,

If the standard elo change is: +- k * (1 - expected_win)

is the mov applied like:

1. +- k * (1 - expected_win) * (3+s)^.85 (would be a large number)

2. +- k * (1 - expected_win) + (3+s)^.85

Both seem larger than I'd expect. In the scenario where a 2000 elo team beats a 1500 elo team by 20, the expected point spread, the winning team would normally gain ~2 elo points (k factor of 38). With the MOV adjustment (as addition, option 2 above), this would be boosted to ~16. (as mult, option 1 above, this would be ~30).

Intuitively, this feels high for a team meeting expectations. Am I missing something? Or is this meant to add volatility?

"Margin of victory matters. Although there is some degree of diminishing returns — in basketball particularly, teams often send in the scrubs in lopsided games, and free-throw strategy can turn narrow wins into games that were closer than they appear from the final score — larger margins of victory get a higher multiplier. Specifically, the margin of victory factor is calculated as (3 + s) ^ .85, where s is the scoring differential"

Expand full comment
Daniel JC's avatar

Based on notes on some old 538 models, I think there is another term in the +- update which normalizes so that the expected post-game rating is the same as the pre-game rating. This is akin to dividing the `k * (1 - expected_win) * (3+s)^.85` by the expected value of `(3+s)^.85` conditioned on s being positive. If the expected post-game rating is not the same as the pre-game rating then there is something the model "knows" that is nor reflected in the current ratings.

Expand full comment
Tom Adams's avatar

The LRMC (Baysian) pi factors are posted here:

https://www2.isye.gatech.edu/~jsokol/lrmc/lrmcratings.html

So there is a "rating" of sorts, not just a ranking

Don't use the conversion factors from Sokol's early papers because they are for classical LRMC. But the two conversion *methods* in those papers might be useful.

Expand full comment
Tom Adams's avatar

PS: a linear conversion of the pi's to spreads is not very good in my opinion.

Expand full comment
Abe's avatar

Can the outcome of a game between A and B change C’s rating? For example, if C beat B last week and then B upsets A so B’s rating rises, does C’s go up too?

Expand full comment
Daniel JC's avatar

Does your Elo update rule satisfy the martingale property? I.e. that the expected ratings after a game are the same as the ratings before the game.

Expand full comment