Is there a statistic that measures the stability of a set of coefficients from a rolling (i.e. last 100 days) multiple regression(s)?

I have not done econometrics for a while, so I am probably going to screw this up. Anyway, do a regression over the whole time range that includes dummy variables times the independent variables (1 during the period of interest, and zero elsewhere). That could be the unrestricted regression. Do another regression without these dummy variables. This would be the restricted regression. The resulting error sum of squares would be F testable. Something like ((ESSr - ESSur)/q) / ((ESSur/(N-k)) where F is dist with q degrees of freedom in numerator and N-k in denominator (and k are the number of vars including dummy, q is number of vars being tested). As I said, it has be a while since I did this so buyer beware but it is something like this.

For inference on the stability of moving-window or expanding-window regressions look at the work of Werner Ploberger and various co-authors. They've developed a bunch of tests, some kind of like CUSUM tests, to do exactly this kind of thing.richg.

maybe you can use chows breakpoint test for a quick check, just came into my mind, which method do you use for doing the regression?solong

Thank you for all the great responses.We apply a standard five factor multiple regression model to managers using 100 data points. Each day, 1 point enters and another point exits the 100 data points; thus, the five coefficients evolve over time. Some managers have relatively stable exposures while others vary. I was hoping to find a statistic that measured the overall stability of the five coefficients and then compare the stability of one manager’s coefficients to another’s coefficients.