28.05.2013 Views

Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...

Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...

Etude des marchés d'assurance non-vie à l'aide d'équilibres de ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

tel-00703797, version 2 - 7 Jun 2012<br />

Chapitre 1. Sur la nécessité d’un modèle <strong>de</strong> marché<br />

η = g(π)<br />

-5 0 5<br />

Univariate smoothing<br />

Link functions<br />

logit<br />

probit<br />

compl. log-log<br />

0.0 0.2 0.4 0.6 0.8 1.0<br />

π<br />

π = g(η) (−1)<br />

0.0 0.2 0.4 0.6 0.8 1.0<br />

Inverse link functions<br />

inv. logit<br />

inv. probit<br />

inv. cloglog<br />

-4 -2 0 2 4<br />

Figure 1.4: Link functions for binary regression<br />

In this paragraph, we present gradually some classic smoothing procedures, from the simplest<br />

to more complex methods. Probably the simplest method to get a smooth function is to<br />

regress a polynom on the whole data. Assuming observations are <strong>de</strong>noted by x1, . . . , xn and<br />

y1, . . . , yn, a multiple regression mo<strong>de</strong>l is appropriate with<br />

Y = α0 + α1X + · · · + αpX p .<br />

Using f(x) = <br />

i αix i is clearly not flexible and a better tool has to be found. One way<br />

to be more flexible in the smoothing is to subdivi<strong>de</strong> the interval [min(x), max(x)] into K<br />

segments. And then we can compute the average of the response variable Y on each segment<br />

[ck, ck+1[. This is called the bin smoother in the literature. As shown on Hastie and Tibshirani<br />

(1990) Figure 2.1, this smoother is rather unsmooth.<br />

Another way to find a smooth value at x, we can use points about x, in a symmetric<br />

neighborhood NS(x). Typically, we use the k nearest point at the left and k nearest at the<br />

right of x to compute the average of yi’s. We have<br />

s(y|x) =<br />

1<br />

CardNS(x)<br />

<br />

i∈NS(x)<br />

where the cardinal CardNS(x) does not necessarily equal to 2k + 1 if x is near the boundaries.<br />

Again we do not show the result and refers the rea<strong>de</strong>r to Hastie and Tibshirani (1990) Figure<br />

2.1. This method, called the running mean, takes better into account the variability of the<br />

data. However we lose the smoothness of previous approches.<br />

An extension of this approach is to fit the linear mo<strong>de</strong>l y = µ + αx on the points (xi, yi)<br />

in the neighborhood (for i ∈ NS(x)). That is to say we have a serie of intercepts µ and slopes<br />

α for all observations. We called this method the running line, which generalizes the running<br />

mean, where α is forced to 0.<br />

76<br />

yi,<br />

η

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!