Forecasting players' career stats

[Sidney the Kidney]

Overall good stuff, but the late career stuff (ie. players when they're 31 and beyond) seem to almost universally expect dramatic drop off in production for guys who, up until that point, were putting up top 10 or better production league-wide.

I know that injuries seem to be the biggest culprit, but at the same time if you look at most of the "elite" or "generational" players from the past, they were very productive into their mid to late 30s. I see no reason why guys like Crosby, McDavid, Ovechkin, and Malkin wouldn't be the same way.

IMO, one major flaw as the model currently is built is too much expectation for injury and too much reliance on comparing these "generational" types to peers who were more in the good to very good range.

 

[Hockey Outsider]

I don't think I ever wrote back to @Zuluss re his thoughtful comments.

I think the biggest takeaway for me is it was probably inappropriate to apply the model to a player currently in his 30's. The "input" players have a wide range of career arcs. Most of them had long, full careers (Gretkzy, Jagr, Messier, Sakic, etc) - but the model also implicitly assumes that there's a chance that the player ends up burning out young like Jason Allison, Barry Pederson or (especially) Jimmy Carson.

I think that approach generally makes sense. We don't know for sure what will happen with, say, Nathan MacKinnon. I think it's highly unlikely he'll age like Jimmy Carson - but it's still within the realm of possibility.

In the case of an established player like Alex Ovechkin - we know that's not going to happen. There's zero chance that his career path will be like Allison's (or Pederson's, or Carson's), because he's already playing well past the age at which they retired.

If I were to forecast Ovechkin's career arc (a challenging task - he's practically unprecedented as a goal-scorer this late in his career), I'd need a separate set of input players - excluding those who retired before age 35.

What does this all mean? The case studies that I did for McDavid, Matthews, Kovalchuk (hypothetical - had he not left the NHL), MacKinnon and Pastrnak are probably valid. Maybe Stamkos too. The ones I did for Crosby, Ovechkin and Malkin clearly aren't. As I said in the first post, I'm not taking this forecast system too seriously - but I figured that this "product recall" was worth disclosing.

(Now that I've written this post, I'm pretty sure I wrote this response to Zuluss elsewhere - not sure if I posted it in another thread or not. But I figured it makes sense to post it here, either way).

 

[Hockey Outsider]

I wanted to revisit my February 2020 post about career forecasts. I think it's interesting to look back on these projections, and try to figure out what caused the differences. Was the model deficient, or is there something that couldn't have possibly been forecasted (ie a player suffering a major injury)?

Obviously, the league shortened both the 2020 and 2021 seasons due to COVID-19. On the other hand, the scoring environment has risen slightly. These two factors have largely offset each other.

Here are the nine players from my original post (excluding Kovalchuk, because he was a hypothetical). I'm comparing their career totals through the end of the 2024 season. These are sorted from best to worst:

McDavid

	Goals	Assists	Points
Projected	336	647	983
Actual	335	647	982
Difference	- 1	-	- 1

I'm not going to lie. Half the reason I'm posting this is because I get to boast about how close my projection was for McDavid. Forecasting 4+ years into the future, I got McDavid's total to within a single point. Obviously there's some luck to this. I can do player forecasts for the rest of my life and I might not ever get as good a result again.

MacKinnon

	Goals	Assists	Points
Projected	342	547	889
Actual	335	564	899
Difference	- 7	17	10

Highly accurate. I got MacKinnon's total accurate to within 1%. In reality he passed a bit more, and scored a bit less than I projected, but even then, both totals are within about 3%.

Stamkos

	Goals	Assists	Points
Projected	552	558	1,110
Actual	548	584	1,137
Difference	- 4	26	27

Highly accurate. I got Stamkos's total accurate to within 2.5%. The goals total is extremely close. If you want to nitpick, the model was a bit low on assists (by about 5%) - but I would still consider this very close overall.

Kane

	Goals	Assists	Points
Projected	468	779	1,247
Actual	471	813	1,284
Difference	3	34	37

Highly accurate. I got Kane's total accurate to within 3%. Same comment as for Stamkos - the model is extremely close for goals, and a bit conservative for assists.

Matthews

	Goals	Assists	Points
Projected	389	313	702
Actual	368	281	649
Difference	- 21	- 32	- 53

Somewhat accurate. Matthews simply missed too much time. I didn't publish per-game stats but from 2021 to 2024, I forecasted (per 82 games) Matthews would average 55 goals, 45 assists and 100 points. In reality he averaged 62 goals, 45 assists and 107 points. Overall Matthews is scoring goals at a somewhat better rate than the model predicted (assists are bang on), but is missing enough time that he's fallen below his forecasted total.

Pastrnak

	Goals	Assists	Points
Projected	364	412	775
Actual	348	379	727
Difference	- 16	- 33	- 48

Somewhat accurate. Overall the model overestimated Pastrnak's production by around 7%. However, in the original post, I commented that Pastrnak is playing on a very strong team with excellent linemates, and it's unlikely that will persist throughout his career. So the model said he'd be at 775 points, but my commentary said that, in this particular case, the forecast is probably overstated. Anyone who builds models should understand how and when it might not work.

Malkin

	Goals	Assists	Points
Projected	449	727	1,176
Actual	498	798	1,296
Difference	49	71	120

Now we're starting to get into the misses. Malkin has exceeded the forecast by 10%. Can we figure out why? Given his history of injuries, the model predicted that Malkin would have retired after 2022. (The results are very close through the end of that season - 449 projected goals vs 444 actual, 727 projected assists vs 702 actual, and 1,176 projected points vs 1,146 actual). Accurate to within 3%, but the model didn't expect that a 35 year old player with such a long history of missing games would keep going. (Predicting when a player retires is a very difficult exercise).

Crosby

	Goals	Assists	Points
Projected	531	936	1,467
Actual	592	1,004	1,596
Difference	61	68	129

The model clearly underestimates Crosby. We're only off by about 9% comparing the career totals, which isn't too bad. But if we only look at the forecast period, the model expected him to score just 205 points from 2021 to 2024. In reality, he scored 333. The reason? The model is really designed to project young players over their prime years. The aging curve is based on data from numerous players, many of whom fell off a cliff around 30 (ie Jason Allison, Barry Pederson, Jimmy Carson, etc). The model doesn't update its aging curve (so there was still some Carson baked into the forecast, which is simply wrong, because Crosby is still playing at an all-star level, and Carson had already retired by that age). My conclusion is the model shouldn't be used to forecast stats for still-productive players in their 30's, as it will likely understate their output.

Ovechkin

	Goals	Assists	Points
Projected	764	618	1,383
Actual	853	697	1,550
Difference	89	79	167

Same commentary for Ovechkin. The model predicted that he'd retire after 2022, and even then, his goalscoring was forecasted to plummet. What Ovechkin has done over the past four years is, essentially, not forecastable. He's warped people's expectations. I've heard many fans say something to the effect of "Matthews is keeping pace with Ovechkin though age 26, so he has a good chance of scoring 700 or 800 goals too". That's not how this works. Ovechkin is a freak of nature (I mean this in the most complimentary way possible), and hoping that someone a decade younger will age the same way, is not a reasonable expectation.

 

[bobholly39]

I wanted to revisit my February 2020 post about career forecasts. I think it's interesting to look back on these projections, and try to figure out what caused the differences. Was the model deficient, or is there something that couldn't have possibly been forecasted (ie a player suffering a major injury)?

Obviously, the league shortened both the 2020 and 2021 seasons due to COVID-19. On the other hand, the scoring environment has risen slightly. These two factors have largely offset each other.

Here are the nine players from my original post (excluding Kovalchuk, because he was a hypothetical). I'm comparing their career totals through the end of the 2024 season. These are sorted from best to worst:

McDavid

	Goals	Assists	Points
Projected	336	647	983
Actual	335	647	982
Difference	- 1	-	- 1

I'm not going to lie. Half the reason I'm posting this is because I get to boast about how close my projection was for McDavid. Forecasting 4+ years into the future, I got McDavid's total to within a single point. Obviously there's some luck to this. I can do player forecasts for the rest of my life and I might not ever get as good a result again.

MacKinnon

	Goals	Assists	Points
Projected	342	547	889
Actual	335	564	899
Difference	- 7	17	10

Highly accurate. I got MacKinnon's total accurate to within 1%. In reality he passed a bit more, and scored a bit less than I projected, but even then, both totals are within about 3%.

Stamkos

	Goals	Assists	Points
Projected	552	558	1,110
Actual	548	584	1,137
Difference	- 4	26	27

Highly accurate. I got Stamkos's total accurate to within 2.5%. The goals total is extremely close. If you want to nitpick, the model was a bit low on assists (by about 5%) - but I would still consider this very close overall.

Kane

	Goals	Assists	Points
Projected	468	779	1,247
Actual	471	813	1,284
Difference	3	34	37

Highly accurate. I got Kane's total accurate to within 3%. Same comment as for Stamkos - the model is extremely close for goals, and a bit conservative for assists.

Matthews

	Goals	Assists	Points
Projected	389	313	702
Actual	368	281	649
Difference	- 21	- 32	- 53

Somewhat accurate. Matthews simply missed too much time. I didn't publish per-game stats but from 2021 to 2024, I forecasted (per 82 games) Matthews would average 55 goals, 45 assists and 100 points. In reality he averaged 62 goals, 45 assists and 107 points. Overall Matthews is scoring goals at a somewhat better rate than the model predicted (assists are bang on), but is missing enough time that he's fallen below his forecasted total.

Pastrnak

	Goals	Assists	Points
Projected	364	412	775
Actual	348	379	727
Difference	- 16	- 33	- 48

Somewhat accurate. Overall the model overestimated Pastrnak's production by around 7%. However, in the original post, I commented that Pastrnak is playing on a very strong team with excellent linemates, and it's unlikely that will persist throughout his career. So the model said he'd be at 775 points, but my commentary said that, in this particular case, the forecast is probably overstated. Anyone who builds models should understand how and when it might not work.

Malkin

	Goals	Assists	Points
Projected	449	727	1,176
Actual	498	798	1,296
Difference	49	71	120

Now we're starting to get into the misses. Malkin has exceeded the forecast by 10%. Can we figure out why? Given his history of injuries, the model predicted that Malkin would have retired after 2022. (The results are very close through the end of that season - 449 projected goals vs 444 actual, 727 projected assists vs 702 actual, and 1,176 projected points vs 1,146 actual). Accurate to within 3%, but the model didn't expect that a 35 year old player with such a long history of missing games would keep going. (Predicting when a player retires is a very difficult exercise).

Crosby

	Goals	Assists	Points
Projected	531	936	1,467
Actual	592	1,004	1,596
Difference	61	68	129

The model clearly underestimates Crosby. We're only off by about 9% comparing the career totals, which isn't too bad. But if we only look at the forecast period, the model expected him to score just 205 points from 2021 to 2024. In reality, he scored 333. The reason? The model is really designed to project young players over their prime years. The aging curve is based on data from numerous players, many of whom fell off a cliff around 30 (ie Jason Allison, Barry Pederson, Jimmy Carson, etc). The model doesn't update its aging curve (so there was still some Carson baked into the forecast, which is simply wrong, because Crosby is still playing at an all-star level, and Carson had already retired by that age). My conclusion is the model shouldn't be used to forecast stats for still-productive players in their 30's, as it will likely understate their output.

Ovechkin

	Goals	Assists	Points
Projected	764	618	1,383
Actual	853	697	1,550
Difference	89	79	167

Same commentary for Ovechkin. The model predicted that he'd retire after 2022, and even then, his goalscoring was forecasted to plummet. What Ovechkin has done over the past four years is, essentially, not forecastable. He's warped people's expectations. I've heard many fans say something to the effect of "Matthews is keeping pace with Ovechkin though age 26, so he has a good chance of scoring 700 or 800 goals too". That's not how this works. Ovechkin is a freak of nature (I mean this in the most complimentary way possible), and hoping that someone a decade younger will age the same way, is not a reasonable expectation.

Absolutely incredible how close some of these predictions were.