Laboratoire d'Océanographie et du Climat: Expérimentation et Approches Numériques
Meeting room nr. 417, corridor 45-55, 4th floor
4 place de Jussieu
75252 PARIS CEDEX 05
Seminar: A Bayesian homogenization method: Sequential detection based on prior characterization of change points
by Alexis HANNART , Laboratoire d'Océanographie et du Climat: Expérimentation et Approches Numériques (LOCEAN)
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 detection problem, are dealing with this issue.
We take a bayesian approach to this problem. The bayesian framework enables to take advantage of previous homogenization results data, used here to quantify both prior distribution of jump amplitude and prior probability of jump occurence. We use these assumptions in a gaussian single change point model and combine them with an adequate cost function to successively decide upon the existence of a jump, and infer its characteristics in case it exists. Both can be done explicitly.
We then generalize to the multiple change point situation practically faced in homogenizing, now assuming that jump occurence a priori follows a stochastic renewal process, with distribution of time between jumps again characterized from past results data. This assumption is combined to previous single change point model and decisioning criterion, to isolate subsequences most likely to contain a unique, detectable jump. This provides the basis for sequential detection on the entire sequence. The iterative algorithm thus obtained is finally implemented on real and simulated series resulting in similar or improved performance level, but at much lower computational cost, than state of the art multiple change point detection methods.
For more information: A. HANNART ()
Link to presentation: