72% chance we lose 1st rd

[MadaboutWu]

thegalen, I don't understand the distinction between picking upsets and picking winners. What changes between the two tasks that would make this model more effective at the former than the latter?

The sample size of 15 of 21 is, of course, laughable. You probably need a sample size in the thousands (and certainly the hundreds) to get any sense of the model's accuracy.

[forevershockerfan]

That was Princeton.

This is the first time since 1946 that Harvard has been in the tourney.

You are right, my bad. I remember now. What a game. Thought Princeton had them.

[ShockTalk]

I find it interesting that many outsiders are looking at this game as an offensive vs defensive team matchup. I believe this thinking to be far from the truth. Coach Marshall is a hard nose defensive coach first, who believes that defense will lead to a successful offense. VCU leads in steals, but steals and TOs does not equal a solid defensive team on their own. To think WSU is not a strong defensive team is a huge error.

Secondly, I think one needs to look at "who has these numbers been built against", the competition faced. WSU overall SOS rank is 56, VCU is 154. The Valley worst SOS team is rated at .5095. The CAA has only 3 teams higher than this, one is last place 1-31 Towson and VCU isn't one of the other 2. Discounting Towson, ODU had the best SOS rating in the CAA of .5240. 6 teams in the Valley were higher including WSU who was at .5527.

Not saying VCU isn't good, if not real good. However, the question I'd ask is how good would their numbers have been playing in the Valley this year?

[SubGod22]

They probably would have finished 3rd in the Valley.

[thegalen]

thegalen, I don't understand the distinction between picking upsets and picking winners. What changes between the two tasks that would make this model more effective at the former than the latter?

The sample size of 15 of 21 is, of course, laughable. You probably need a sample size in the thousands (and certainly the hundreds) to get any sense of the model's accuracy.

That's not the sample size the model is built on. It's based on 256 observations (tourney matchups 2004-2010). The performance given is based on applying the model to an out of sample year (i.e., one not used in the model).

As for it being a useful tool, think of it this way. If you're a hospital and diagnose cancer as terminal correctly 70% of the time, but you want to know what's up with the 30% of the time you get it wrong, you build a tool to see what if any variables correlate with getting it wrong and how strong those relationships are. Then you incorporate that back in to your original model. This exercise stops at step 1, which for our purposes is fine because the seed serves as a proxy for expected win-loss. We know that VCU is supposed to lose (terminal diagnosis). We also know that upsets happen because seeding also serves as a proxy for composite strengths which leads to a non-negligible error rate aka relatively frequent upsets in early rounds (misdiagnoses). This simply tries to address the subset of composite based predictions and seed "errors" and uncover commonality among that population.

[thegalen]

Here's a bit more on the sampling: http://harvardsportsanalysis.wordpre...-and-improved/

[Kung Wu]

They probably would have finished 3rd in the Valley.

I see what you did there! LOL