I have some data that has the pattern in the picture below (but little noisier than that).

I want to run a non-linear regression that tries to capture the dynamic of this data in the time-series where there is a jump that occurs at a specific time (in the picture below t=5) and with some exponential decay of the variable Y following the jump.

I don't have much of a clue on how to specify this regression. At first I was thinking to simply do:

$ln(y) = ln(a) + b*ln(x)+ln(u)$

But what about the jump? I am interested in the inference of all three parameters simultaneously (including the jump parameter).

enter image description here

  • $\begingroup$ Could you elaborate on the motivation for fitting a regression? Are you interested in inference about the size/location of the jump? Rate of decay? All three parameters simultaneously? $\endgroup$
    – Andrew M
    Feb 24, 2015 at 6:38
  • $\begingroup$ @AndrewM good point! All three parameters simultaneously. $\endgroup$
    – CharlesM
    Feb 24, 2015 at 6:41
  • $\begingroup$ Sorry, but I want to be quite sure -- the time of the jump is unknown? $\endgroup$
    – Glen_b
    Feb 24, 2015 at 7:14
  • $\begingroup$ @Glen_b I know when the jump occur because at time=5 (in my graph below) is when there is a very important macroeconomic announcement $\endgroup$
    – CharlesM
    Feb 24, 2015 at 7:24
  • $\begingroup$ Then I only count two parameters. How are there three parameters? $\endgroup$
    – Glen_b
    Feb 24, 2015 at 9:33

1 Answer 1


Your function appears like it could be parametrized as $y = Ce^{-A(x-B)}\mathbb{I}_{[x>b]}$, where $\mathbb{I}$ is 1 if $x>B$ and 0 else. I can't see anything stopping you from just fitting a non-linear regression to your hypothesized functional form, although you will need to be careful with your choice of an optimizer, since this function is not differentiable in $B$.

A cursory look at the literature$^1$ and extant software indicate that grid searches on $B$, and maximizing the remaining parameters conditional on $B$ are used, and that it can be shown that you only need to check the values $x$ that the data assume.

It's not clear to me, however, that standard asymptotic theory would apply to derive standard errors, owing to the lack of differentiability (w/r/t $B$) in the likelihood. I suppose bootstrapping could provide a way out.

$^1$ Muggeo. "Estimating regression models with unknown break-points" (2003)

  • $\begingroup$ ah I see. OK so when x is greater than the "known" jump time $B$, I have $y=Ce^{-A(x-b)}$ otherwise some constant $C$ plus noise. Let's say that I assume that my errors are normally distributed, can I estimate this using MLE? $\endgroup$
    – CharlesM
    Feb 24, 2015 at 7:43
  • 1
    $\begingroup$ Yes, and if you can assume that $B$ is known, then any old optimizer or solver for non-linear least squares will do. $\endgroup$
    – Andrew M
    Feb 24, 2015 at 7:47
  • $\begingroup$ Running an OLS like $ln(Y) = ln(C) -A*ln(x-B)*D + ln(u)$ where $D = 1$ if $x>B$ would be however problematic, correct? $\endgroup$
    – CharlesM
    Feb 24, 2015 at 7:50
  • $\begingroup$ Taking a log of $Y$ changes how the errors enter (multiplicatively vs additive), and more importantly, you can only log $Y$ if it is positive everywhere. Even if we define $\log(0) \times 0 =0$, I don't think we can write anything sensible. Formally, $y=Ce^{-A(x-B)}\mathbb{I}_{[x>b]} \Rightarrow \log(y) = (\log C -A(x-B)) + \log \mathbb{I}_{[x>b]}$, but still have a $\log 0$ appearing in the final term. $\endgroup$
    – Andrew M
    Feb 24, 2015 at 21:59

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.