0
$\begingroup$

Spurious regression is generated when we regress time series data that are non-stationary. So is it possible to time lag the variable to further find a higher correlation?

$\endgroup$

1 Answer 1

1
$\begingroup$

Taking lagged non-stationary variable as a covariate in a regression model will result in spurious model as well, which will output an optimistically high biased value of t-statistic and r squared. This is because non-stationary time-series have very long memory as could be illustrated by an acf function plot.

$\endgroup$
8
  • $\begingroup$ What model can you recommend in these analysis of non-stationary variables so spurious model would be omitted? I am analyzing a commodity price and a stock price. $\endgroup$
    – chidori
    Nov 2, 2017 at 10:06
  • $\begingroup$ @chidori, a good choice is to take first differences of the time-series (returns). I specifically recommend log(xt) - log(xt-1) when it comes to stock data, which are unstationary in terms of variance of the returns. $\endgroup$ Nov 2, 2017 at 10:31
  • $\begingroup$ So according to my research, ways to omit spurious results is to make the time series data stationary. If it exhibits a random walk, I need to use differencing and when it has a deterministic trend, I can detrend it. But if it fails, I'll analyze the returns instead (which what you have said). So when I examine the returns, do I assume it to be stationary? And what is the difference in normal returns and log(xt) - log(xt-1) and if it is unstationary, how do I correlate it? $\endgroup$
    – chidori
    Nov 2, 2017 at 10:39
  • $\begingroup$ @chidori, it is not necessary to detrend by substracting a trend model. You can take differences and the trend goes away. However, you can try to detrend, but then you don't take differences. log(xt) - log(xt-1) helps a lot when your returns are not stationary in variance. $\endgroup$ Nov 2, 2017 at 10:44
  • $\begingroup$ What's the best correlation method for log returns? $\endgroup$
    – chidori
    Nov 2, 2017 at 10:53

Your Answer

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

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