Maybe a Dumb Question - Predictors of Performance

[oljimmy]

I was thinking about the upcoming finals and about Stuart Skinner's weird record. Had a few disastrous games a while ago but has been lights out for a while. Then I thought: what will better predict his SCF performance, the whole playoffs or the recent string of games?

I realized that there is an important question here: if a player is playing game N, how many games do you go back (N-x) to achieve maximum predictive power, on average?

If a player is about to play a game, obviously it would be silly to go back 500 games, and average out that series of performances to predict what he will do next. It is equally obvious that you'd want to go back more than 1 game; having a brilliant or a terrible game means almost nothing(all on its own) when determining how someone will do in the following game. Somewhere between 1 and 500 there is a number that gets you maximum predictive accuracy for the average player. What is that number?

Obviously this is a crazy complex question but has anyone looked into it or something like it?

Click to read more...

 

[Hockey Outsider]

I'm not aware of anyone who looked into this, but I think it's an interesting question, and probably warrants a deeper dive at some point. (I might work on that, but unfortunately it won't be anytime soon).

One factor that would need to be considered - age. Let's say (and this is purely hypothetical) the "magic" number is 200 games. 200 games for a 22 year old would presumably underestimate their future performance (because they're still growing and learning); but 200 games for a 38 year would almost certainly overestimate their future performance. It might be literally true that 200 games is the best predictor on average, but even then, it could be misleading when applying it to older/younger players.

 

[oljimmy]

I'm not aware of anyone who looked into this, but I think it's an interesting question, and probably warrants a deeper dive at some point. (I might work on that, but unfortunately it won't be anytime soon).

One factor that would need to be considered - age. Let's say (and this is purely hypothetical) the "magic" number is 200 games. 200 games for a 22 year old would presumably underestimate their future performance (because they're still growing and learning); but 200 games for a 38 year would almost certainly overestimate their future performance. It might be literally true that 200 games is the best predictor on average, but even then, it could be misleading when applying it to older/younger players.

Right, exactly. After I posted this I realized that there will be too many individual factors that make a magic number kind of meaningless. There could be a magic number for some players though, maybe a relative measure of how consistent they are or how prone they are to wild swings in play.

 

[BlueSeal]

If they win or lose, it's because of Skinner.

 

[tinyzombies]

Skinner just has to stay in the paint- he’s so big that if he’s just “there” the d can take care of the rest

 

[JimmyH]

I was thinking about the upcoming finals and about Stuart Skinner's weird record. Had a few disastrous games a while ago but has been lights out for a while. Then I thought: what will better predict his SCF performance, the whole playoffs or the recent string of games?

I realized that there is an important question here: if a player is playing game N, how many games do you go back (N-x) to achieve maximum predictive power, on average?

If a player is about to play a game, obviously it would be silly to go back 500 games, and average out that series of performances to predict what he will do next. It is equally obvious that you'd want to go back more than 1 game; having a brilliant or a terrible game means almost nothing(all on its own) when determining how someone will do in the following game. Somewhere between 1 and 500 there is a number that gets you maximum predictive accuracy for the average player. What is that number?

Obviously this is a crazy complex question but has anyone looked into it or something like it?

Your x would not have durable predictive value because, as you mentioned, disastrous performances can interrupt a string of decent or otherwise great performances, or the other way around. Would likely be an unreliable method of analysis unless you smooth out the formula to account for game-to-game variance and outliers. I forgot all the symbols but a better formula is (N - median of x) where x = a range of past performances with commonality, then you can build multiple models where you expand or constrict the range based on whatever factors you believe to have impacted the performances within the given range.