: News Article:
SBD Interview with Pegula
View Single Post
06-07-2012, 10:34 PM
Join Date: Feb 2009
Location: Northern Virginia
Originally Posted by
You're right, one really should build a regression model that adjusts for confounding variables and then check for a significant p value.
I was wrong earlier when I said that there was correlation - and you're right, there might not be statistical correlation (unless we check for extraneous values).
Better point: correlation isn't causation. Regression models can indicate correlation. There may not be a correlation between man-games-lost and wins, but is there a causal relationship between man-games-lost and wins?
This difference is the mistake that the quants made on Wall Street. The PhD's Wall Street hired applied models in an attempt to predict an interactively complex system. The ones that AIG hired told them that there was a 99.85 percent chance of never having to pay out on credit-default swaps. Didn't work out so well. Did the .15 just happen, or were the models wrong?
Taking two data points (man-games-lost and wins/loses), using a regression model, and determining that because there's no correlation, they don't effect each other is like using a set of pliers to dissect a giraffe. The pliers work just fine. Wrong tool.
There are so many variables that aren't considered in that model. I'm not great at math, but I can tell where it applies, and where it doesn't apply well.
View Public Profile
Der Jaeger's albums
Find More Posts by Der Jaeger