Early online detection of high volatility clusters using Particle Filters. (15th July 2016)

• Record Type: Journal Article
• Authors: Mundnich, Karel; Orchard, Marcos E.
• Is Part Of: Expert systems with applications. Volume 54(2016)
• Page Start: 228
• Page End: 240
• Publication Date: 2016-07-15
• Subjects: Bayesian inference -- Risk-sensitive particle filters -- Stochastic volatility estimation -- Event detection; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2016.01.052
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: High volatility cluster detectors based on particle filters and hypothesis testing. Detection uses prior-posterior probability estimates in asymmetric hypothesis tests. Risk-sensitive particle filters used to track and detect greater financial risk. Scheme tested and validated using both simulated and actual IBM's stock market data. Abstract: This work presents a novel online early detector of high-volatility clusters based on uGARCH models (a variation of the GARCH model), risk-sensitive particle-filtering-based estimators, and hypothesis testing procedures. The proposed detector utilizes Risk-Sensitive Particle Filters (RSPF) to generate an estimate of the volatility probability density function (PDF) that offers better resolution in the areas of the state-space that are associated with the incipient appearance of high-volatility clusters. This is achieved using the Generalized Pareto Distribution for the generation of particles. Risk-sensitive estimates are used by a detector that evaluates changes between prior and posterior probability densities via asymmetric hypothesis tests, allowing early detection of sudden volatility increments (typically associated with early stages of high-volatility clusters). Performance of the proposed approach is compared to other implementations based on the classic Particle Filter, in terms of its capability to track regions of the state-space associated to a greater financial risk. The proposed volatility cluster detection scheme is tested and validated using both simulated and actual IBM's daily stock market data.