Player career projection modeling

[Lonewolfe2015]

Why is Stamkos never going to play more than 73 games in a season again when he's never missed a game with an injury until this year?

 

[Derrty]

I will bump this thread in 16 years. Setting an alarm on my phone for 2030.

You're going to look like Zack Morris in 16 years when you pull that dinosaur of a phone out.

 

[Bear of Bad News]

This left the first page on the main boards, so I'm moving it here.

 

[wgknestrick]

Why is Stamkos never going to play more than 73 games in a season again when he's never missed a game with an injury until this year?

A proper "projection" is NEVER going to assume a perfect season for anyone. While I can't defend the OP's method, your question is unfounded relative to statistical methods..

 

[ted2019]

I still see Stamkos as a 700+ Goals player and if Ovechkin doesn't head back to mother Russia, breaking the 700 Goals barrier also.

 

[Delicious Dangles*]

I get what you were trying to do, but you're comparing historical trends and historical players that may or may not have any relation to current players, assuming all players follow those trends, ignoring current trends, assuming equal risk of injuries, assuming lockouts, assuming MAJOR drop-offs that really make no logical sense on an average OR to the individual players. Not to mention things like playing 4 games per year for their last 4 years?

It seems you made these predictions off of "player types", but I don't know how you slotted these players into one specific player type or drew comparables.

This probably all makes sense in your statistical model, but that's the problem with using flawed statistical models to try and prove or predict everything.

I don't get the point of this when it is so obviously, wildly wrong.

 

[Bear of Bad News]

I don't get the point of this when it is so obviously, wildly wrong.

Perfect is the enemy of good. What's wrong with someone liking the numbers side of things, and wanting to give something a try?

We prefer constructive criticism in this subforum if you have any to give.

 

[Delicious Dangles*]

Perfect is the enemy of good. What's wrong with someone liking the numbers side of things, and wanting to give something a try?

We prefer constructive criticism in this subforum if you have any to give.

I gave plenty of constructive criticism in my previous post, thank you very much. Including numerous variables that were ignored, and numerous assumptions made that I don't think can be made in any proper model.

This isn't perfect or good. It's relatively useless. It's no more accurate than me picking random numbers and making a thread, but I'm sure you wouldn't be taking too kindly to that.

I wasn't aware this was the "you must love and agree with all forms of extremely flawed or worthless statistics and models" forum. If you're making a model, the BASE requirement should be that it actually makes sense in reality.

And I have every right to give that opinion, even if you don't interpret it as "constructive criticism".

 

[Bear of Bad News]

I wasn't aware this was the "you must love and agree with all forms of extremely flawed or worthless statistics and models" forum. If you're making a model, the BASE requirement should be that it actually makes sense in reality.

Read the two forum-specific rules here:
http://hfboards.mandatory.com/showthread.php?p=78276171

Like I said, feel free to offer constructive feedback.

 

[Delicious Dangles*]

Read the two forum-specific rules here:
http://hfboards.mandatory.com/showthread.php?p=78276171

Like I said, feel free to offer constructive feedback.

If you're talking about the special rules, I did read them, and I did follow them.

Respecting the science of statistics does NOT mean agreeing with every model that is presented. In fact, it is FOR the respect for the science of statistics that people BUTCHERING said statistics should be identified, and incorrect models and uses eliminated.

Like I said, I DID feel free to offer constructive feedback. It's not my fault you dislike my constructive criticisms. I'm not sure what you want out of me.

 

[Bear of Bad News]

So how would you improve upon it, with actionable items?

 

[Delicious Dangles*]

So how would you improve upon it, with actionable items?

Not do all the things I pointed out earlier, and then some?

He was wrong right from the beginning, since he took an average of 50 random players to try and project out for elite players, which would bring different variables and trends. He was also wrong right from the start because he tried to "average" out GP and injuries.

He also seemed to project based on the player's career production, but this ignores the whole development path. So somebody that took a different development path will have wildly different results than he should.

Which is probably why the ever-durable Kessel, projected to get 91 points this year and getting better every game, who took a slower development path and had valid reasons for his decreased production in earlier years, is projected to miss a ton of games and peak at 60 points for the rest of his career past 26.

Frankly, I would scrap the whole thing, because no matter what you do, it is not going to have any useful predictive power or be representative of anything.

 

[Bear of Bad News]

Thank you.

You had me 100% until the "scrap the whole thing" idea.

I disagree with the notion that "no matter what you do", you won't have any useful predictive power. Hockey Prospectus' VUKOTA method has pretty good predictive power, for one.

 

[Delicious Dangles*]

Thank you.

You had me 100% until the "scrap the whole thing" idea.

I disagree with the notion that "no matter what you do", you won't have any useful predictive power. Hockey Prospectus' VUKOTA method has pretty good predictive power, for one.

I suggested scrapping THIS model, not all models. I'm not aware of the method that you speak of, but I'm sure there are models that would have (not necessarily good, but still) SOME predictive power that people could find a use for.

It's just not this one, and that's the point I was making. Unless you change this one so much that it's essentially an entirely different model, it has no use.

 

[Bear of Bad News]

I suggested scrapping THIS model, not all models. I'm not aware of the method that you speak of, but I'm sure there are models that would have (not necessarily good, but still) SOME predictive power that people could find a use for.

It's just not this one, and that's the point I was making. Unless you change this one so much that it's essentially an entirely different model, it has no use.

I'd want to see a bit more under the hood before making that conclusion.

 

[Bear of Bad News]

Now as for how I'd do it (and roughly how I think that VUKOTA works)...

For a given player at age X, I would develop a similarity score that describes how well that player's data matches up against other players through age X. There would be an art to this, of course - for projecting the future, what "matters" in terms of similarity? Ignoring the fact that many of the things that would be important are likely not measured (or not measured well).

Suppose that the player matches up well against Player A (92% similar), Player B (89% similar) and Player C (85% similar).

I'd then take the composite of Player A, B, and C's age X+1 bodies of work (weighted by those percentages) and that would be my middle of the road projection. With enough players in the sample (and there would be more than 3 "similar" players for most comparisons), one could also estimate a variance and a confidence interval.

I've been working on something like this for goaltenders for awhile now, and it's actually what led me to develop my "opponent strength" and "game-to-game variability" metrics (described in other threads in this forum) as potential inputs for the similarity scores.

 

[Ohashi_Jouzu*]

When it comes to offensive number projections, I wonder if there's a way to factor shooting in there as a way to "tweak" our expectations in terms of future points.

I've been thinking about this kind of thing most recently because of Jagr's and Selanne's ability to remain productive at such an "advanced" age. When Jagr went goalless in last year's playoffs, people were quick to assume that Jagr had finally hit the "drop off", and would be less likely to be productive going forward ("he's done", was a common read around here). I believe he commented something to the effect that he's "not worried about capitalizing on chances until the chances are no longer there to capitalize on". Fact is, he was still generating chances then (almost 3 SOG/game credited, and no doubt more in terms of "missed shots" - more on that in a second) and continues to create chances now. Selanne also continues to generate chances, just slightly less and less over time, and a sort of correlation/pattern appears.

Now, "chances" are obviously subjective on some level, and shots on goal alone leave out a significant portion of the picture when it comes to my idea of generating offense, so I propose including "missed shots". Even though missed shots have no way of impacting advanced metrics like Corsi/Fenwick, etc, and are only available (afaik) for "recent" seasons, I think they might contribute to a better-rounded picture of how "potentially potent" a player is offensively, and help frame some of the player projections. Put together, I'd expect an exploration of shots/missed shots to reveal some sort of correlation with offensive role, opportunity on a team, AND ability (c.f. similar exploration via situational TOI, or whatever). It might even yield another layer of comparison (maybe an element those "similarity scores"), or reveal other "comparables" that might not be as obvious from simple comparison of point scoring. I'm just curious as to how it would take shape if/when broken down by age and more and more players get added to the mix.

This isn't a proposed "instead of", btw, but a possible contributing element to whatever model - maybe a correction factor for just the goal scoring side of it, or something. There are guys out there like Ryan Smyth and Shane Doan, for example, that might get overlooked in a conversation that tunnel-visions on point production, but when you look for underlying factors contributing to their consistent production in very much "2 way forward" roles, look no further than consistently getting enough chances/shots on goal to place them both top 10 all-time in that regard. The precipitous fall of Scott Gomez's offensive production (as a Habs fan, it's a pretty standard/classic example) can also be somewhat discerned from his increasing difficulty in generating shots - even at a relatively young age compared to the guys I've mentioned so far.

 

[Hockey Monkey]

He was wrong right from the beginning, since he took an average of 50 random players to try and project out for elite players, which would bring different variables and trends. He was also wrong right from the start because he tried to "average" out GP and injuries.

The initial population wasn't totally random, it was a group of "successful NHLers" which actually included some elite players (Yzerman, Gretzky, and Messier were all in there).

I'm not sure what your second point means.

 

[Czech Your Math]

Now as for how I'd do it (and roughly how I think that VUKOTA works)...

For a given player at age X, I would develop a similarity score that describes how well that player's data matches up against other players through age X. There would be an art to this, of course - for projecting the future, what "matters" in terms of similarity? Ignoring the fact that many of the things that would be important are likely not measured (or not measured well).

Suppose that the player matches up well against Player A (92% similar), Player B (89% similar) and Player C (85% similar).

I'd then take the composite of Player A, B, and C's age X+1 bodies of work (weighted by those percentages) and that would be my middle of the road projection. With enough players in the sample (and there would be more than 3 "similar" players for most comparisons), one could also estimate a variance and a confidence interval.

I've been working on something like this for goaltenders for awhile now, and it's actually what led me to develop my "opponent strength" and "game-to-game variability" metrics (described in other threads in this forum) as potential inputs for the similarity scores.

I think you would want way more than 3 "similar" players, if possible.

I would use similarity in categories like these:

- peak (top 2 or 3 seasons) point production
- peak PPG (probably need a minimum total games for each or a total of the peak seasons)
- G/A ratio in those peak seasons
- % GP in those peak seasons
- total games, points, PPG, G/A ratio, and % GP through current age

Another way might be to weight recent seasons more heavily and then base as similarity score off that, based on age (e.g. [3*X1 + 2*X2 + X3]/6... where X1 is most recent season, X2 is previous season, etc.).

A player with a higher peak is generally more likely to be able to play longer (because, e.g., 70% of their peak is still very valuable, while that 70% for a more typical player is of more marginal value).

 

[Bear of Bad News]

I think you would want way more than 3 "similar" players, if possible.

I agree, and didn't want to write out 20 or more players in my shortened example.

Presumably, one would select a similarity threshold, and take all players who met or exceeded that level of similarity.

I'm sure that some players would be truly unique (how comparable could one be to Wayne Gretzky, for instance).