Ranking random variables that show step change

Subject:Ranking random variables that show step change
Date:Sat, 26 Dec 2009 12:54:50 -0800 (PST)
I have a time series of a large number of random varaibles
(statistics). I'd like to find the ones that show a large and abrupt
change at time T. This may be a simple type of data mining, but as I
can't find a data mining newsgroup, I thought I'd post to some related
groups. There is a fair amount of noise in the data, and in addition
to the step change I am looking for, the variables tend to contain
other slower moving changes (so far I have been avoiding the last
issue by looking at a relatively short time windows).

What I have done so far is this:
- Correlate each varaible with a step function at time T. This does a
good job of finding variables with the abrupt change, and I can use
the correlation coefficient (R^2) to rank them.
- Calculate an average value before/after time T, and rank based on
the percent change.

My questions are:
- Is this a known problem that has some clever, well thought out
solution?
- Any ideas on other ways to do my ranking?
- Any ideas on good ways to combine various rankings to produce a
composite ranking? The motivation is I'm most interested in varaibles
that show an abrupt AND significant change in value.

Gene



Other posts:
How is MSE = Variance + bais^2
What a photon really is:
Math Researcher: Increase Your Research Productivity with the Leading Web 2.0 Research Portal
Negative r, positive B? regression
For a fixed entropy, which distirbution is closest (in KL sense) to any possible distribution
• Ranking random variables that show step change
how do i know that data lie in linear subspace or not?
Standard error of the ARIMA constant
check for Gaussianity in high dimensional data sets
Testing the difference between betas within a regression model
Entropy of a function of a random variable

generated at 21:53:21