Post by eric on Sept 27, 2018 21:48:06 GMT
It turns out I've posted five years of preseason standings / title odds predictions, so let's give them a glance.
Since exactly one team wins the title every year it's pretty difficult to do any real analysis on actual champions. On the other hand, we have win loss records for every team every year, so how do those stack up from preseason to actual?
Real nice, real nice indeed! On average they're off by six wins, which isn't too shabby considering they can't take into account injuries, trades, signings, players who haven't played before, and can only take a guess at TC. It's usually a good idea to see if the errors have any underlying function, which we can see by graphing the gap against the total:
Nope! Absolutely no correlation - high wins are as likely to be on the money as low wins.
.
Let's look at the two biggest outliers and see if there's a pattern there we can learn from:
The 2002 Spurs were predicted to win 42 games and in fact won 24. They also traded future Yes as a Knick Hall of Famer Neon Boudeaux halfway through the season. Pretty easy to suss out that correlation, and obviously impossible to predict what trades will be made.
The 2006 Hornets were predicted to win 47 games and in fact won 65. It would be easy to say "well, Yao!" but that turns out to not be the real story. Let's look at the Win Shares by position predicted and actual:
So yes, certainly Yao getting 15 instead of 7 wins helped, especially to help make up the drop off from Shareef to the two headed Mo Bamba - Halawi Winabi monster, and let's briefly examine why that gap existed. For rookies (or other players who've never played before) the Machine uses their preseason PER to estimate a WS/48 number, so Yao's decent 20ish PER only worked out to a decent .13ish WS/48, far below his actual .25. The famous weakness of PER is of course that it does a terrible job measuring defense, which is doubly an issue here since Yao's excellent defense also helps everyone else on what was before his arrival a very poor defensive unit.
But the big story is on the wings. Again a midseason acquisition tells part of the story, since RJ Barrett ain't got NOTHING on Jake Shuttlesworth, but that only accounts for 2 wins. The other 10 included massive improvements in Vince Carter and Giannis Antetokounmpo on the offensive end, and it's probably not the case that Yao swatting shots turned Vince Carter from a below to above average scorer, or Giannis from a good to great one. Predicting TC is of course a fool's errand, so the Machine goes with a baseline that estimates modest improvement for young players and modest decline for old ones, which is what happens on average, and in this case that resulted in a year over year estimate of +1 win.
There's certainly no better way to guess at TC, and I doubt there's a good way to more strongly emphasize defense since on the perimeter invisible attributes are just as if not more important than bleals. Instead, I think the best course is to imagine larger error bars on young and old teams, since they are more likely to blow up or be blown up respectively.
.
.
Let's take a quick gander at the relationship of preseason to end of season odds.
As we'd expect from how well it does with wins, pretty well correlated! Of the 119 team-seasons the Machine gave less than a 5% preseason shot at winning the title, only three broke 10% in end of season odds (02 pistons and suns, 06 hornets). Looking at it the other way, of the 116 that ended up with less than a 5% end of season shot at the title, only five had a 10% or better shot in preseason.
A team with no shot almost certainly doesn't have a shot, a team that has a shot almost certainly does; in each case about 1 in 20 blow it.
We can also look at aggregate odds against actual titles:
Not bad in either case! The Pacers are the only real surprise, the Knicks and Suns are extremely high on both lists and the Globetrotters do well for themselves too. Overall correlation favors day 120 over preseason, which it ought to, but the gap isn't as big as I would have guessed.
Since exactly one team wins the title every year it's pretty difficult to do any real analysis on actual champions. On the other hand, we have win loss records for every team every year, so how do those stack up from preseason to actual?
Real nice, real nice indeed! On average they're off by six wins, which isn't too shabby considering they can't take into account injuries, trades, signings, players who haven't played before, and can only take a guess at TC. It's usually a good idea to see if the errors have any underlying function, which we can see by graphing the gap against the total:
Nope! Absolutely no correlation - high wins are as likely to be on the money as low wins.
.
Let's look at the two biggest outliers and see if there's a pattern there we can learn from:
The 2002 Spurs were predicted to win 42 games and in fact won 24. They also traded future Yes as a Knick Hall of Famer Neon Boudeaux halfway through the season. Pretty easy to suss out that correlation, and obviously impossible to predict what trades will be made.
The 2006 Hornets were predicted to win 47 games and in fact won 65. It would be easy to say "well, Yao!" but that turns out to not be the real story. Let's look at the Win Shares by position predicted and actual:
pg wing big
11.1 act 26.6 act 29.5 act
10.4 pred 14.4 pred 22.4 predSo yes, certainly Yao getting 15 instead of 7 wins helped, especially to help make up the drop off from Shareef to the two headed Mo Bamba - Halawi Winabi monster, and let's briefly examine why that gap existed. For rookies (or other players who've never played before) the Machine uses their preseason PER to estimate a WS/48 number, so Yao's decent 20ish PER only worked out to a decent .13ish WS/48, far below his actual .25. The famous weakness of PER is of course that it does a terrible job measuring defense, which is doubly an issue here since Yao's excellent defense also helps everyone else on what was before his arrival a very poor defensive unit.
But the big story is on the wings. Again a midseason acquisition tells part of the story, since RJ Barrett ain't got NOTHING on Jake Shuttlesworth, but that only accounts for 2 wins. The other 10 included massive improvements in Vince Carter and Giannis Antetokounmpo on the offensive end, and it's probably not the case that Yao swatting shots turned Vince Carter from a below to above average scorer, or Giannis from a good to great one. Predicting TC is of course a fool's errand, so the Machine goes with a baseline that estimates modest improvement for young players and modest decline for old ones, which is what happens on average, and in this case that resulted in a year over year estimate of +1 win.
There's certainly no better way to guess at TC, and I doubt there's a good way to more strongly emphasize defense since on the perimeter invisible attributes are just as if not more important than bleals. Instead, I think the best course is to imagine larger error bars on young and old teams, since they are more likely to blow up or be blown up respectively.
.
.
Let's take a quick gander at the relationship of preseason to end of season odds.
As we'd expect from how well it does with wins, pretty well correlated! Of the 119 team-seasons the Machine gave less than a 5% preseason shot at winning the title, only three broke 10% in end of season odds (02 pistons and suns, 06 hornets). Looking at it the other way, of the 116 that ended up with less than a 5% end of season shot at the title, only five had a 10% or better shot in preseason.
A team with no shot almost certainly doesn't have a shot, a team that has a shot almost certainly does; in each case about 1 in 20 blow it.
We can also look at aggregate odds against actual titles:
Not bad in either case! The Pacers are the only real surprise, the Knicks and Suns are extremely high on both lists and the Globetrotters do well for themselves too. Overall correlation favors day 120 over preseason, which it ought to, but the gap isn't as big as I would have guessed.
