Robustness Check

The Warriors are 13-0. Normally, I would ecstatic but it revealed a big problem in my model.  My model now says the Warriors have a 0 percent chance of repeating their performance from last year, no matter your prior beliefs. Some people would say that Steph Curry is the ultimate heat check, but in this case he is the robustness check (I know, that was extremely lame... I'm just gonna own it).

The only team in my data set to start 13-0 was the 2002 Mavericks. They then went a "disappointing" 47 and 22 after the hot start. The reason I call it disappointing is that it is decidedly middle of the pack for teams that won 12 of their first 13.  Here is the final record of all teams that won 12 of their 13 games between 1996 and 2014.

	season*	team	wins	losses
0	1996	Bulls	69	13
1	2008	Lakers	65	17
2	2013	Spurs	62	20
3	1997	Lakers	61	21
4	2010	Spurs	61	21
5	2002	Mavericks	60	22
6	2001	Lakers	58	24
7	1996	Rockets	57	25
8	2013	Pacers	56	26
9	2006	Jazz	51	31

Using the Bayesian formulation of model, the probability of winning at least 67 game given 13-0 start is 0, but the probability of doing so given a 12-1 start is greater than 0. This seems pretty flawed to me; winning the extra game shouldn't hurt the probability winning a lot of games.  This result is driven by the fact that 2002 Mavs are the only 13-0 team, and they only made it 60 wins. 

Damn. My model is overfit. Honestly, it's not too surprising. What I published was a first cut. I didn't do any cross-validation. Importantly, I just used counts of raw data, rather than fitting distributions to the results. I can't fix it tonight, but you can expect some updates soon. 

*Note, all seasons are demarked by the year the season began. The 2010 Spurs played in 2010 and 2011.