An Optimal Design of Early Warning Systems: A Bayesian Quickest Change Detection Approach

This paper proposed a new optimal design of Early Warning Systems (EWS) to detect early warning signals of an impending financial crisis. The problem of EWS was formulated from a policy maker's perspective. Hence the probability threshold was obtained by minimizing the policy maker's welfare loss. This paper employed the state-of-the-art Bayesian Quickest Change Detection (BQCD) as the methodology to detect the early warning signals as soon as possible. We showed that the BQCD method outperformed the Logit model used in traditional EWS models based on results of simulation exercise and the out-of-sample predictions of the 1997 Asian financial crises. We found that not only early warning signals were stronger prior to a crisis, but also stronger warning signals appeared more frequently. The BQCD method was sensitive to the increase in frequency, hence out-performed the traditional Logit-EWS model.