On Tue, 23 Jan 2001 14:07:34 -0600, Bill Vedder <[EMAIL PROTECTED]>
wrote:
>
> I'm trying to better understand a statistical method which is vaguely
> outlined in a paper I'm reading and am hoping a kind soul here can help.
>
> The method is described as "variance reduction". The author uses it to
> decide whether an economic indicator truly has some forecasting ability
> over and above a simple "naive" forecast. The method consists of
> ordering values of the indicator (the independent variable) and the
> associated historical future return (the dependent variable) then
> grouping the observations into "cells". Each cell may have, for example,
>
> 200 observations of the ind var and the associated dep var. Averages are
> taken for the values of the dependent variable in each cell; these, then
> are the conditional forecasts ie. values of the dependent variable
> conditional to the value of the indicator. An average is also calculated
> over the entire set of dependent variable observations and is used to
> define a "naive forecast".
>
> The conditional forecasts are then compared to the naive forecast as a
> first step in determining whether the indicator truly has predictive
> significance. However, as when comparing two means from noisy data,
> there's always a chance that the predictive value you've calculated is a
> random result.
>
> And that's my question. How does one use "variance reduction" to
> determine whether two means are statistically different.
>
The idea is basically, "Analysis of variance."
You have your variance as a measure of fit.
If you measured CHANCE outcomes a few hundred times,
how many of those would do better than what *your*
model predicts? - If you can build a good enough
statistical model, you can do a test based on the numbers,
comparing Variance A to Variance B. If the one is a
subset of the other, you can sometimes subtract and
test the difference.
Economic data are hard to analyze because you don't
know how many pieces (of something) that the time-series
represents.
- hope this helps.
--
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
