2016-17 Team Win Projection Contest / Discussion

[AcrossTheCourt]

What's SMRE?

And how do the rankings change with Pythag wins?

[permaximum]

SMRE shouldn't have any weight. In doesn't even make sense especially in this "NBA team win projection contest". This is not dart.

Also, using actual wins is a must.

*It

Tarrazu won. According to RMSE or RMSE/MAE blend.

Kmedved and caliban are honorable mentions.

[permaximum]

What's SMRE?

And how do the rankings change with Pythag wins?

SMRE is apperantly a metric Mike G invented. Think about it as dart. Unlike RMSE, it rewards "right on target" predictions. However in our case (actually almost in every case), it's completely useless because of the nature and distribution of team wins.

[AcrossTheCourt]

So is it error^.5?

[EvanZ]

Never really feel like I learn much from these predictions since a) the totals are fairly close to each other and b) minutes projections could make a significant difference.

Would be interested to see how everyone's predictions would work as retrodictions given that we now know minutes played for every player.

[shadow]

The Bucks absolutely destroyed my entry. Only had them projected to win 29 games after Middleton injury. If I exclude them my RMSE jumps up to 4.74, good enough for 3rd in that category. Although I'm sure a lot of people could have improved similarly if they removed their worst projection. So it's kind of a silly exercise.

[permaximum]

Never really feel like I learn much from these predictions since a) the totals are fairly close to each other and b) minutes projections could make a significant difference.

Would be interested to see how everyone's predictions would work as retrodictions given that we now know minutes played for every player.

No surprise there since most participants use +/- based models. Especially RPM and BPM. I was expecting similar results. I think randomness is more in play than minute projections.

[permaximum]

So is it error^.5?

SMRE is squared mean root error, and I thought it could serve as a tie breaker, if one party takes the straight avg error while another wins in RMSE.
With tanking seeming to take the league in a big way, it may be that we benefit from shrinking the larger errors rather than squaring them. Under-performing teams are now racing to the bottom and fleeing from their expectations. Contending teams necessarily benefit from the easy wins, so keep on exceeding predictions.

[Mike G]

Sometimes a process is better predicted with an exponent greater than one applied to the erros. Traditionally, 2 is used (along with 1/2), since calculation was possible without computers.
Sometimes an exponent less than one is more predictive. So 1/2 (along with 2) seems the logical and symmetrical choice.

[sndesai1]

i'm not sure i follow...i don't really see this as something like pythagorean % where you tune the exponent for best performance
the rmse has various mathematical properties that lead to its common use in judging predictions

[sndesai1]

Never really feel like I learn much from these predictions since a) the totals are fairly close to each other and b) minutes projections could make a significant difference.

Would be interested to see how everyone's predictions would work as retrodictions given that we now know minutes played for every player.

i don't know if there is a lot to learn from this contest about metrics a this point, but there probably is value in exploring if/how people model things like fit, travel, record in close games, player development, etc.

[Crow]

Congrats to the top performers. If any of those want to say more about their methodology or retrodict with actual minutes that would be appreciated and perhaps useful.

[EvanZ]

Might be interesting to try the Huber loss:

https://en.wikipedia.org/wiki/Huber_loss

[Crow]

Sounds like a good suggestion, in principle.

[BasketDork]

Gotta say, I don't feel my 'victory' has been accepted among the community.