Estimation of extreme quantiles from

heavy-tailed distributions in a semi-parametric

location-dispersion regression model

Communication at the 4th ”Rencontres des jeunes

chercheurs africains en France”
by
Aboubacrène Ag AHMAD(1,4)
in collaboration with
Stéphane GIRARD , Antoine USSEGLIO-CARLEVE(3) , Aliou
(2)

DIOP(4) et El Hadji DEME(4)

(1) Université des Sciences, des Techniques et des Technologies de Bamako, FST.
(2) Université
Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK.
(3) Univ. Rennes, Ensai, CNRS, CREST - UMR 9194, F-35000 Rennes.
(4) LERSTAD, Université Gaston Berger de Saint-Louis.

Aboubacrène Ag AHMAD (FST, LERSTAD) Oral communication December 10, 2020 1 /5

 Motivation
Modeling of extreme events in finance, insurance, environmental
science, ...
A recurrent statistical problem: estimation of extreme
conditional quantiles.

Aboubacrène Ag AHMAD (FST, LERSTAD) Oral communication December 10, 2020 2 /5

 Location-dispersion regression model
Consider the following regression model between a random response
variable Y ∈ R and a deterministic covariate vector x ∈ Π ⊂ Rd , d ≥ 1:

Y = a(x) + b(x) Z,

where
a : Π → R : regression function (unknown);
b : Π → R+ \ {0} : dispersion function (unknown);
Z ∈ R : is a heavy-tailed random variable with tail-index γ > 0.

Aboubacrène Ag AHMAD (FST, LERSTAD) Oral communication December 10, 2020 3 /5

 Inference

Step 1: Estimation of the regression and dispersion functions

â n (x) = q̂ n,Y (µ2 | x) and b̂ n (x) = q̂ n,Y (µ3 | x) − q̂ n,Y (µ1 | x)

Step 2: Estimation of the tail-index

1 kXn −1
γ̂n = log Ẑ mn −i ,mn − log Ẑ mn −kn ,mn .
k n i =0

Step 3: Estimation of extreme conditional quantiles

q̃ n,Y (αn | x) = â n (x) + b̂ n (x)q̂ n,Z (αn ).

Aboubacrène Ag AHMAD (FST, LERSTAD) Oral communication December 10, 2020 4 /5

 Advantages
+ More flexibility than other estimation approaches;
+ Allows to circumvent the curse of dimensionality;

Aboubacrène Ag AHMAD (FST, LERSTAD) Oral communication December 10, 2020 5 /5