2015-16 Team win projections

[Crow]

I believe Hollinger used PER, perhaps age curve for similar players and subjectivity to win the contest at least twice. I used multiple metric blending and subjectivity to win twice. If PTPM wins a third time, especially in a row, that would strength its claim. Via minutes, metric or conversion to wins method or the sum of the steps.

Could I beat pure PTPM with a metric blend of the 2016 top 5 that gives PTPM a real big weight? Maybe, if I put another entry together and leave out Vegas and maybe as much or all of my own subjective adjustments.

[EvanZ]

Minutes projections clearly have value. In fact, that alone is worthy of its own model and research. But for purposes of ratings, it would be nice to separate out the rating from the minutes. That way we know for sure what we are actually praising.

Congrats to Andrew, regardless. It's impressive to beat this group.

[permaximum]

I believe Hollinger used PER, perhaps age curve for similar players and subjectivity to win the contest at least twice. I used multiple metric blending and subjectivity to win twice. If PTPM wins a third time, especially in a row, that would strength its claim. Via minutes, metric or conversion to wins method or the sum of the steps.

Could I beat pure PTPM with a metric blend of the 2016 top 5 that gives PTPM a real big weight? Maybe, if I put another entry together and leave out Vegas and maybe as much or all of my own subjective adjustments.

The difference here, PTPM has been winning since it's development. It entered into 2 prediction contests or whatever you call it and it won both of them in a row. OFC there are lots of other components to consider but I'm trying to be optimistic here since it's new and a bit different than usual RAPM/SPM hybrids.

[Crow]

I understand and agree. I just figured a bit of context and caution might be worth mentioning and wondering if PTPM and blending might might be a worthy rival to pure PTPM. Wouldn't hurt to try to push further and see if pure PTPM is best or if some further improvement is possible. Or maybe go with blended minutes estimates and pure PTPM.

[Crow]

Congrats to the top, top 3, top 5 on this scorecard. Still have another bigger, perhaps a bit different? scorecard to come.

[kmedved]

Here are colorful standings, which also add record ATS: Sorry that my playoff-seed tie breakers don't work.



(A few entries cutoff at the end got their own screenshot).

I'm not so worried as to whether PTPM won because Andrew's minutes projections were a bit better. First, he won somewhat comfortably. Second, the big thing is that PTPM is clearly at least on par with the RPM-based methods which the other top entries used. Having a discrete metric generate those kinds of results is pretty important, regardless of whether it's 0.1 RMSE better or maybe only just as good once you strip out minutes. (I am basically disregarding the possibility that his minutes projections were wildly better than anyone else's, since that seems impossible).

A well-earned congrats.

[Crow]

Congrats to the 2016 top 5 blend for best performance, though it wasn't a stand alone, at the deadline entry of course.

[sndesai1]

congrats to andrew and thanks to a few of you for keeping us updated with the progress.

i was just wondering, doesn't andrew blend pt-pm with rapm in his entry? i'm not sure that he's purely using his pt-pm metric.

[permaximum]

You're probably right, it must be a blend of PTPM and RAPM.

[kmedved]

Congrats to the 2016 top 5 blend for best performance, though it wasn't a stand alone, at the deadline entry of course.

The 2015 Top 5 meanwhile did fine, but only marginally better than just the average of all the picks. That's ... disappointing?

[sndesai1]

not sure it would change your point but it would probably be better to compare with the average of only those who participated last year

[kmedved]

not sure it would change your point but it would probably be better to compare with the average of only those who participated last year

Why? I'm not following.

[Crow]

Yeah, calling PTPM "pure" was misleading and probably a mistake. It is a blend I believe, unless it changed from first year. I just meant just that (blend) instead of blended further.

Yeah km, I hadn't noticed that about 2015 top 5 and overall average. Always in a rush. Maybe there is some other option besides just top 5 predictors from 2015 or all of 2016 that would do even better reliably or at least a good chance at reliable. Variable weights based on performance in prior years. I was doing that on my projection blends for a handful or two but you could include more. Retrodiction on past results could help guide over freehand blending.

[sndesai1]

not sure it would change your point but it would probably be better to compare with the average of only those who participated last year

Why? I'm not following.

if the idea was that the top performers from one year don't predict the performance too much better than "average", it only seems fair to compare the top 5 and average of the same cohort.

for example, you and caliban participating is likely helping the average entry for this year. if you two had participated in the 2015 contest, your entries may have placed in the top 5, which would have improved the 2015_top 5 entry in this year's contest. given that there's also new entries dragging down the average, the net impact may not be very large, but i just thought it would make for a clearer comparison

[kmedved]

not sure it would change your point but it would probably be better to compare with the average of only those who participated last year

Why? I'm not following.

if the idea was that the top performers from one year don't predict the performance too much better than "average", it only seems fair to compare the top 5 and average of the same cohort.

for example, you and caliban participating is likely helping the average entry for this year. if you two had participated in the 2015 contest, your entries may have placed in the top 5, which would have improved the 2015_top 5 entry in this year's contest. given that there's also new entries dragging down the average, the net impact may not be very large, but i just thought it would make for a clearer comparison

Got it. That does make sense. I will try and match them up as best as possible sometime next week - will be interesting.