[Todos] Charla en Ciencias de la Atmosfera - ATMO-CHARLAS

[Ciencias de la Atmosfera & Oceanos]

CICLO DE CHARLAS EN CIENCIAS DE LA ATMOSFERA

Seminario en Ciencias de la Atmosfera:

"Detección de saltos artificiales en series climáticas: una nueva técnica
estadística bayesiana".


Alexis Hannart, CNRS (LOCEAN-IPSL), Francia.
Lugar de trabajo acutal: Departamento de Ciencias de la Atmosfera y los
Oceanos (DCAO)


Viernes 12 de Diciembre- 13:30 horas.
 Aula 8, DCAO.

**************************************************************************

Abstract
In climatology, long instrumental records are often affected by artificial
shifts due to changes in the measurement conditions. As these
inhomogeneities usually have the same magnitude as the signal studied, a
direct analysis of the raw series can lead to wrong conclusions.
Statistical objective homogenization procedures, mostly deriving from the
so-called change point problem, are dealing with this issue.
We propose a new multiple change point detection technique. Our method is
based on the identification through bayesian decisioning of subsequences
with a unique change point. This approach enables to maintain a low
complexity while simultaneously leveraging advantages of the bayesian
framework. In particular, appropriate prior distributions enable to
introduce available empirical results on jump amplitude and frequency in
the decisioning.
Technically, we assume jump occurrence a priori follows a stochastic
renewal process, with distribution of time between jumps fitted on past
observations. We use these assumptions in a Gaussian single change point
model and combine them with a quadratic cost function to successively
decide upon the existence of a jump, and infer its characteristics in case
it exists. Both can be done explicitly. Results on simulated series lead
to similar or improved performance level compared to state of the art
multiple change point detection methods.

*************************************************************************

Los esperamos a todos!


Gracias!

Natalia

------------ próxima parte ------------
Se ha borrado un adjunto en formato HTML...
URL: http://mail.df.uba.ar/pipermail/todos/attachments/20081202/739c5be6/attachment.html