I'm using a bunch of techniques and methods from Extreme Value Theory to analyze my data. I have a time series representing the number of events happening in a given day. The time series is unequally temporally spaced, so that I have multiple observations for some days, and zero observations for other days. However, no missing values: if I do not have any observations for a given day, it means that no events happened that day.

The number of events distribution shows a Pareto behavior. I do not attach the plots, but I verified it with qqplots, mean excess plots, Zipf plot, etc. Moreover, I found out that the moments greater than 1 do not exist, so that I should expect that my distribution is in the Maximum Domain of Attraction of the Fréchet distribution, with $\alpha \in [1,2)$.

To estimate $\alpha$ and investigate the dynamics generating the number of events, I used Peaks Over Threshold (POT). Such an approach has two main consequences:

  1. the exceedances (when extreme events happen) are governed by a homogeneous Poisson process, so that the inter-arrival times between the exceedances are exponentially distributed;
  2. the number of events exceeding the threshold follow a GPD (Generalized Pareto Distribution).

One important assumption is that the observations are IID.

I know well the phenomenon under investigation: no way that the number of events happening on a given day can affect the number of events happening in other days. I'm sure. It would make no sense. However, I check for autocorrelation using ACF plot. I can safely conclude that there is no autocorrelation in my time series. See the figure below.

enter image description here

Next, in order to estimate the GPD parameters, I used shape function from evir (R package) to choose a proper threshold $L$.

enter image description here

$L = 5$ looks like a suitable choice, since the value of $\xi$ is somehow stable and the $95\%$ confidence intervals are narrow.

Using $L = 5$ I obtained the following (nice) estimates:

  • $2,430$ exceedances ($0.917\%$ of the total number of observations)
  • $\xi = 0.84 \pm 0.0376$
  • $\sigma = 5.08 \pm 0.1964$
  • $\mu = 6.30 \pm 0.1134$

Morover, since $\alpha = \frac{1}{\xi}$, $\alpha = 1.19$ (a value consistent with my exploratory analysis). The Pareto tail fit is shown in the figure below.

enter image description here

However, the inter-arrival times between exceedances are very far from being exponentially distributed. See figure below, where the concave departure indicates a Pareto behavior.

enter image description here

Such a behavior is confirmed by the empirical cumulative distribution function (ECDF) and from the fact that all the moments are infinite. I can safely conclude that the exceedances' occurrences are not governed by a homogenoeus Poisson process, since their inter-arrival times are Pareto distributed.

Moreover, when I look at the ACF plot of the inter-arrival times, I see lag-1 to lag-5 dependences. See figure below.

enter image description here

Indeed, if I look at the point process exceedances I see a sort of clustering of extreme events. See figure below.

enter image description here

Such results do not depend on the choice of the threshold (I obtain the same results for $L \in \{1,2,10\}$).

Such a behavior does not make sense, I'm not analyzing earthquakes or similar extreme events that tend to cluster. I was expecting anything but extreme events clustering.

Am I missing something? Any ideas?

However, how can I model that kind of inter-arrival times (autocorrelated and Pareto-distributed with infinite moments)?


Your Answer

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

Browse other questions tagged or ask your own question.