Adaptive Bayesian Changepoint Analysis and Local Anomaly Detection

• Activity Number: 356 - Recent Advances in Change-Point Analysis for Business and Economics Data
• Type: Invited
• Date/Time: Thursday, August 12, 2021 : 12:00 PM to 1:50 PM
• Sponsor: Business and Economic Statistics Section
• Author(s): David S Matteson* and Haoxuan Wu
• Companies: Cornell University and Cornell University
• Keywords: Bayesian; Outliers; Time Series; Break Points

Abstract

This paper incorporates global-local shrinkage priors into a dynamic linear model to create an algorithm that can adaptively detect both change points and local outliers. The algorithm utilizes a state-space approach to separate the true signal from the impact of outliers and measurement errors with stochastic volatility. Within the state equation, a global parameter is used to reduce noise on a global scale and local parameters are used to track noise at each time-step. This set-up provides a flexible framework that allows the algorithm to detect unspecified change points in sparse, noisy data sets. The algorithm is effective at detecting changes in presence of significant outliers and noise with stochastic volatility. We additionally include an extension for dynamic regression analysis. We compare our algorithm against several other change point algorithms to demonstrate its efficacy in several simulation scenarios and two empirical settings.