 |
The Chairman
|
(address
is not clickable)7/06/05
Models of the Stars, by Victor
Niederhoffer
A query comes in regarding the methods used by star quants; whatever the methods are, I'm sure that any success is not due to the complexity involved. With successive models, it's essential to calculate the standard error of the estimate from the simplest model, i.e. to take the average error, adjusted for squares if you wish, but preferably just using the absolute errors about the mean. Once you have that average and total error, then fit the more complicated model. How much have the errors been reduced? By comparing the ratio of the reduction to the original error with an F-test, you can measure both the significance and magnitude of the improvement. Such a procedure will generally show that the simple variables used, i.e. the main effects, account for an order of magnitude more impact than the additional variables added. The significance of the initial variable will also be considerably greater. As for predictability, which is so much more important, see Bacon, and see Stigler's Statistics on the Table, especially the regression paper. Whatever mumbo the star quants purport to use, it isn’t complexity that gives them the edge, but more likely the niche, such as after-hours trading.