2
$\begingroup$

Assume that I have two time series $Y_{1t}$ and $Y_{2t}$ that are sampled at the same frequency. Is there a way to quantify the uncertainty in their difference $Y_{1t} - Y_{2t}$? That is, can we get confidence bands on $Y_{1t} - Y_{2t}$?

My thought is that some sort of a dependent bootstrap should apply.

This question also seems related to this question.

$\endgroup$
2
  • $\begingroup$ Do you want to have a confidence interval at each time point, or over time points? $\endgroup$
    – amoeba
    Aug 29, 2014 at 11:32
  • $\begingroup$ @amoeba Your point and that of ssdecontrol are related. I want a confidence interval at every time point, which reflects the different uncertainty at each time point. $\endgroup$ Aug 29, 2014 at 11:33

2 Answers 2

1
$\begingroup$

This is kind of a non-answer, but it might invite its own discussion and is a bit long for a comment:

To get a confidence interval of the different at each time point, you'd need a point estimate of the difference at each time point. Bootstrapping only helps with standard errors. I have no idea how you'd compute such a point estimate unless you have many observations of each variable at each time point. Note that this is the situation in the question you linked. Otherwise, my impression is you might have to fit a model to get a prediction at each point. Then maybe you could construct a CI based on a difference-in-means t-test or something like that.

$\endgroup$
2
  • $\begingroup$ Does the needed number of samples at $t$ depend somewhat on whether the values of a time series are functions of previous values in the time series? $\endgroup$
    – Alexis
    Aug 29, 2014 at 15:21
  • $\begingroup$ I imagine they might, but imagination is about as far as I can get. Heuristically, this might be the same reason that your "effective sample size" in MCMC is not equal to the actual number of draws. Autocorrelated observations contain overlapping information, so conceivably you might need more of them. But this is just a guess and this topic is well outside my area of expertise. $\endgroup$ Aug 29, 2014 at 15:29
1
$\begingroup$

The difference $Y_{1t}-Y_{2t}$ is known with certainty; after all, you just measured it (unless you are measuring with uncertainty), so confidence bands around the difference seem to make little sense without an underlying model. I assume I may be misunderstanding your question - perhaps you want to edit it?

That said, you could fit, say, a state space model to the difference (e.g., with ets() in the R forecast package) and then simulate it to get the distribution of the underlying states at particular time points. Unfortunately, simulate.ets() only works for out-of-sample simulations, but you could probably extract what you need from the fitted model. Or you could fit the state space model to initial segments of the difference, then simulate out from there, sort of a "rolling origin" forecasting approach.

$\endgroup$

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.