Performance measurement of Bayesian online changepoint detection and its application

Recently with the proliferation of open data, time series analysis has garnered significant attention. In particular, analyses exploiting complex anomalies identifiable through meta-analysis of data have advanced, leading to the proposal of various anomaly detection methods. Therefore, we focus on Bayesian Online Changepoint Detection (hereinafter referred to as BOCD), a method that endows detection results with quantitative reliability in the form of change point probabilities as well as enables real-time detection. In this study, we investigated the characteristics of BOCD using simulation data in which changes in the generative algorithm were explicitly manifested. The results demonstrated that the method achieved high detection accuracy regarding average changes. Furthermore, comparisons with other statistical methods and deep learning approaches revealed that BOCD exhibits superior detection accuracy. Subsequently to examine the applicability of BOCD, we performed anomaly prediction for river systems by defining anomaly labels and anomaly probabilities based on rainfall and water level data. The evaluation of prediction accuracy demonstrated that using the change point probability of water level data as an explanatory variable is more advantageous than employing the water level data directly. Moreover, incorporating change point probabilities into the training and testing of the predictive model proved beneficial for establishing timely river anomaly alerts. The multifaceted validation based on both simulation data and various real-world datasets is expected to contribute to the future development of time series analysis.