I have a stochastic model used to simulate time series of some process. I am interested in the effect of changing one parameter to a specific value and want to show the difference between the time series (say model A and model B) and some sort of simulation based confidence interval.

I have been simply running a bunch of simulations from model A and a bunch from model B and then subtracting the medians at each time point to find the median difference throughout time. I used the same approach to find the 2.5 and 97.5 quantiles. This seems like a very conservative approach since I am not considering each time series jointly (e.g., each point is considered independent of all others at previous and future times).

Is there a better way to do this?

  • $\begingroup$ Why use the median, rather than the mean? Are the distributions not symmetric? $\endgroup$ – naught101 Apr 17 '12 at 10:44
  • $\begingroup$ Were you able to find an answer to this question? $\endgroup$ – tchakravarty Aug 29 '14 at 11:21
  • 1
    $\begingroup$ @TC, this question seems closely related. $\endgroup$ – Mars Feb 23 '15 at 18:32

If you can simulate from the two time series (let's call them $X_t$ and $Y_t$, where $t=1,2,...,T$), and if you simulate from both of them $S$ times so that you get the time series tuples $(\{X_t^s\}_{t=1}^T, \{Y_t^s\}_{t=1}^T)$ for $s = 1,2,...,S$, then instead of calculating the median difference throughout time as \begin{align}\Delta M = \text{median}(X_1^1-Y_1^1, X_2^1-Y_2^1,...,X_T^1-Y_T^1, X_1^2-Y_1^2,...,X_T^S-Y_T^S),\end{align} you could instead simulate from the median difference as a function of time. What I mean by this is that you can define \begin{align}\Delta M(t) = \text{median}( X_t^1-Y_t^1, X_t^2-Y_t^2, ..., X_t^S-Y_t^S),\end{align} so that you now get the median as a function of time. If you can assume that the median is the same across time, the estimates for $\Delta M(t)$ should coincide with the estimate for $\Delta M$ for large enough numbers of simulations $S$. But if the function $\Delta M(t)$ exhibits strong time-dependence (i.e. is very different for different values of $t$), you will be able to see this through simple means as for instance plotting.


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.