3
$\begingroup$

Specifically, I have a time series with fixed $\Delta t$ of the following form: $x_0, x_1, x_2, ... x_n$ and $t_0, t_1, t_2, ... t_n$ where $t_{i+1}-t_i =\Delta t $.

I'm interested in the improved efficiency for large data sets, if it is possible to simplify linear regression in this context.

Edit: I editet the title since equispaced is an important property.

$\endgroup$
1
  • 1
    $\begingroup$ It is likely won't do you a whole lot of good. All you gain are closed formulas for the moments of $t$: the mean is $n(\Delta t)/2$ and the variance is $((n+1)^2-1)(\Delta t)^2/12$. You still have to compute moments of the dependent variable $y$ as well as the covariance of $t$ and $y$. And for rolling regressions, which are updated quickly in constant time, even those formulas accomplish nothing. The only situation I have seen that benefits is where calculations are limited by software (in)capabilities or there are a huge number of dependent variables. $\endgroup$
    – whuber
    Jan 20, 2015 at 16:03

1 Answer 1

5
$\begingroup$

In fact in linear regression it's already assumed that the predictor variable (IV) has no error.

If not for that, you'd have errors-in-variables regression (though it has various other names). OLS is biased in this case.

So there's no opportunity for efficiency when you impose a condition that's already assumed.

There is some opportunity for some efficiency gain because the $t$'s are equispaced.

However, if you need to refit these models each time a new data point arrives, then there are larger opportunities for gain.

A warning, too, about fitting regression to time series; it must be done with care.

$\endgroup$
2
  • $\begingroup$ Thanks for clearing this up for me! As it happens the refit you've spoken of when a new data point arrives is of great interest - could you elaborate or point me to a relevant source? $\endgroup$
    – JHK
    Jan 20, 2015 at 10:15
  • 1
    $\begingroup$ Search our site for rolling regression. $\endgroup$
    – whuber
    Jan 20, 2015 at 13:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.