<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN">
<HTML><HEAD>
<META http-equiv=Content-Type content="text/html; charset=iso-8859-1">
<META content="MSHTML 6.00.5730.13" name=GENERATOR>
<STYLE></STYLE>
</HEAD>
<BODY><FONT size=2></FONT>
<DIV align=center><STRONG><FONT size=2>CICLO DE CHARLAS EN CIENCIAS DE LA 
ATMOSFERA</FONT></STRONG></DIV>
<DIV align=center><STRONG><BR></STRONG><FONT size=2><FONT 
size=3><STRONG>Seminario en Ciencias de la Atmosfera:<BR><BR>"Detección de 
saltos artificiales en series climáticas: una nueva técnica<BR>estadística 
bayesiana".<BR><BR><BR>Alexis Hannart, CNRS (LOCEAN-IPSL), Francia.<BR>Lugar de 
trabajo acutal: Departamento de Ciencias de la Atmosfera y los<BR>Oceanos 
(DCAO)<BR><BR><BR>Viernes 12 de Diciembre- 13:30 horas.<BR>&nbsp;Aula 8, 
DCAO.<BR></STRONG></FONT><BR>**************************************************************************<BR></DIV></FONT>
<DIV align=justify><FONT size=2>Abstract<BR>In climatology, long instrumental 
records are often affected by artificial<BR>shifts due to changes in the 
measurement conditions. As these<BR>inhomogeneities usually have the same 
magnitude as the signal studied, a<BR>direct analysis of the raw series can lead 
to wrong conclusions.<BR>Statistical objective homogenization procedures, mostly 
deriving from the<BR>so-called change point problem, are dealing with this 
issue.<BR>We propose a new multiple change point detection technique. Our method 
is<BR>based on the identification through bayesian decisioning of 
subsequences<BR>with a unique change point. This approach enables to maintain a 
low<BR>complexity while simultaneously leveraging advantages of the 
bayesian<BR>framework. In particular, appropriate prior distributions enable 
to<BR>introduce available empirical results on jump amplitude and frequency 
in<BR>the decisioning.<BR>Technically, we assume jump occurrence a priori 
follows a stochastic<BR>renewal process, with distribution of time between jumps 
fitted on past<BR>observations. We use these assumptions in a Gaussian single 
change point<BR>model and combine them with a quadratic cost function to 
successively<BR>decide upon the existence of a jump, and infer its 
characteristics in case<BR>it exists. Both can be done explicitly. Results on 
simulated series lead<BR>to similar or improved performance level compared to 
state of the art<BR>multiple change point detection 
methods.<BR><BR>*************************************************************************<BR><BR>Los 
esperamos a 
todos!<BR><BR><BR>Gracias!<BR><BR>Natalia<BR><BR></FONT></DIV></BODY></HTML>