PhD graduated
Departure date : 12/31/2012
Supervision : Thierry ARTIÈRES

Statistical learning for classification and detection of chart patterns

This thesis deals with financial stock market analysis and is especially focused on chart pattern recognition. A chart pattern is a particular shape which has a predictive power; it is defined by theoretical rules. Detecting such patterns is difficult. There is an important gap between theory and practice; real patterns do not perfectly respect the theoretical rules. Moreover, chart patterns definition seems subjective; it depends on the financial expert. Finally, there is no large labeled datasets of chart patterns. We study classification and detection of chart patterns using statistical markovian systems. We focus on generative (Hidden Markov Models) and discriminative (Conditional Random Fields, Hidden CRFs) approaches which are standard technologies for sequential data recognition. We propose various strategies to learn accurate systems with small training sets. The first one blends HMMs and HCRFs in such a way that the modeling ability of the generative models is used to limit the overfitting of the discriminative ones. The second strategy, is a semi-supervised approach which learns jointly a HMM and a HCRF systems; it has some similarity with the well-known co-training algorithm. To design an accurate detection system dedicated to a particular financial expert, we propose a two level system where candidate patterns are first extracted from the financial stock-market using HMMs, and then they are confirmed as chart patterns or rejected by a SVM which uses an enriched representation of patterns. While the HMM system is learn once for every expert, the SVM level is trained with an active learning strategy to take into account the expert’s own detection criteria.
Defence : 07/09/2013 - 10h - Site Jussieu 25-26/105
Jury members :
M. Paquet Thierry - Professeur-LITIS [Rapporteur]
Mme. Tellier Isabelle - Professeur-Laboratoire LATTICE [Rapporteur]
M. Main Bernard - Directeur informatique chez BBSP
M. Gallinari Patrick - Professeur-LIP6
M. Artières Thierry - Professeur-LIP6

2011-2014 Publications

 Mentions légales
Site map |