Monitoring for Disruptions in Financial Markets

Abstract

Nous étudions les tests CUSUM historiques et séquentiels pour des séries dépendantes avec des applications en finance. Pour les processus temporels, une nouvelle dimension se présente : l'effet du choix de la fréquence des observations. Un nouveau test est également proposé. Ce test est basé sur une formule de prévision locale d'un pont brownien.

Historical and sequential CUSUM change-point tests for strongly dependent nonlinear processes are studied. These tests are used to monitor the conditional variance of asset returns and to provide early information regarding instabilities or disruptions in financial risk. Data-driven monitoring schemes are investigated. Since the processes are strongly dependent several novel issues require special attention. One such issue is the sampling frequency. We study the power of detection as sampling frequencies vary. Analytical local power results are obtained for historical CUSUM tests and simulation evidence is presented for sequential tests. Finally, a prediction-based statistic is introduced that reduces the detection delay considerably. The prediction based formula is based on a local Brownian bridge approximation argument and provides an assessment of the likelihood of change-points.