Author(s)

M Leleu & J F Boulicaut

Abstract

We are studying new techniques for computing similarities between stock market situations. The situations are represented by means of event sequences (tens or hundreds of event types, hundreds or thousands of events). These events are obtained from available financial data (e.g., discretized financial times series, Reuters financial news). Interestingly, event sequences enable to mix quantitative and qualitative information. In order to analyse the situations and, e.g., to overcome direct similarity measure limitations, we compute signatures for the situations. The signatures are made of sequential patterns and we consider different possibilities for the selection/computation of characteristic patterns. Signatures contain not only the so-called frequent sequential patterns but also some infrequent ones (e.g., a sequential pattern that indicates a crash on some values). We discuss an ongoing application that concerns the study of \“market trends”. Experts already identified in a collection of situations three kinds of periods (trend to a raise, trend to a decline, stability) and we compute the signatures of these labelled situations, looking for characteristic and discriminant patterns. The ultimate goal is then to classify a new situation by looking at its signature and the similarities with labelled signatures. 1. Introduction Financial market evolutions are described by both qualitative and quantitative data with a temporal nature. Our data mining approach has to cope with large data streams while taking into account all these aspects. Therefore, we chose to represent financial market situations by means of event sequences. These events are obtained from market data (e.g. discretized financial time series) and from the financial news provided by specialized press-agency (e.g., Reuters or Bloomberg). We are then looking for signatures of sets of market

Keywords