 |
Center for Nonlinear Studies |
Traditional sequential analysis procedures for change point detection, such as CUSUM, are not well suited to the modern streaming data analysis problems. Such problems are characterised by an unending sequence of data arriving at high frequency that is not interrupted by change point detections. A number of problems arise in this context, including setting control parameters for the detector, and resetting the detector after a change. These problems cannot be addressed by human intervention in high-frequency contexts. We develop change point detectors from adaptive estimation procedures. These procedures are shown in simulation to be less sensitive to control parameters than standard approaches in the continuous monitoring setting. The methodology is with data from finance and cybersecurity.
Host: Melissa Turcotte