Can I draw any proper conclusions about the linearity and strength of a relationship between two non-stationary time series (I'm considering two series of interest rates, series $A$ and series $B$) using simple linear regression when the residuals are serially correlated and heteroskedastic?


I see that computing a naive Pearson correlation coefficient (PCC) doesn't help since this calculation yields a random variable regardless of sample size. Since the $\beta$ in simple linear regression is related to the PCC, I think this would also yield a random variable for the coefficient in the model.


If I run a simple linear regression, and despite non-IID residuals, have an $R^2$ of $0.75$, can I say that "series $A$ explains $75%$ of the variability in series $B$"? Or do I have to qualify that statement by saying "series $A$ explains $75%$ of the variability in series $B$ in this particular sample"? The model is still consistent (p.19), but not unbiased?


Without stationarity (and therefore ergodicity) you can't assume that the linear regression is a consistent estimator. Basically your observations are not iid.

A linear regression at that point is just a geometric construct, the line that minimizes squared distance, but it doesn't really have statistical meaning.


I think you need to say "series A explains 75 of the variabilitiy in series B in this particular sample" and if you want to generalize this you need to do cross validation.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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