Modeling pairwise dependencies in precipitation intensities
Other literature type
(issn: 1607-7946, eissn: 1607-7946)
In statistics, extreme events are classically defined as maxima over a block
length (e.g. annual maxima of daily precipitation) or as exceedances above a
given large threshold. These definitions allow the hydrologist and the flood
planner to apply the univariate Extreme Value Theory (EVT) to their time
series of interest. But these strategies have two main drawbacks. Firstly,
working with maxima or exceedances implies that a lot of observations (those
below the chosen threshold or the maximum) are completely disregarded.
Secondly, this univariate modeling does not take into account the spatial
dependence. Nearby weather stations are considered independent, although
their recordings can show otherwise.
To start addressing these two issues, we propose a new statistical bivariate
model that takes advantages of the recent advances in multivariate EVT. Our
model can be viewed as an extension of the non-homogeneous univariate
The two strong points of this latter model are its capacity at modeling the
entire range of precipitation (and not only the largest values) and the
absence of an arbitrarily fixed large threshold to define exceedances. Here,
we adapt this mixture and broaden it to the joint modeling of bivariate
precipitation recordings. The performance and flexibility of this new model
are illustrated on simulated and real precipitation data.