Regime Shift Detection Project 

This site tracks the development of the software for regime shift detection in time series known as STARS. Originally, the acronym stood for the Sequential T-test Analysis of Regime Shifts, and the software contained only one module for detecting regime shifts (or change points) in mean. Later on, I added two other modules designed to detect shifts in variance and correlation, which were based on the sequential F-test, rather than the t-test. It was decided, however, to keep the original acronym, which now stands for the Sequential Three-part Analysis of Regime Shifts. All these parts can be used independently; however, it is better to do it step by step, from part 1 to  part 3, using the output from the previous step, as an input for the following one. Since the variance module works with anomalies (deviations from mean), it is recommended to start the analysis with the detection of regime shifts in mean (part 1). The deviations from the mean values of those regimes will serve as an input for part 2, i.e., the detection of regime shift in variance. However, you can use your own trend function (for example, a linear trend) instead of the step function calculated in part 1. Similarly, the input for the correlation module (part 3) is a pair of normalized time series, i.e., anomalies divided by standard deviation. Again, it is recommended to use part 2 to detect regime shifts in variance and divide anomalies by the standard deviations for those regimes. However, any other normalization technique can be used.

The STARS software has been used in a wide variety of applications, ranging from oceanography to climate research to economics (see Overview and references in Google Scholar). It has a number of advantages over other existing methods: 

• Automatically detects regime shifts (or change points) in mean, variance, and correlation coefficient;

• Better performance at the end of time series, indicating a possibility of a shift in real time;

• Can be tuned up to detect shifts on different time scales;

• No need to calculate anomalies prior to applying the method;

• No a priori information about shifts in the series is required;

• It can handle time series with autocorrelation, thus separating regimes due to persistence from the true regimes with different statistics.

Sergei Rodionov, Ph.D.
Climate Logic, LLC
Profile on LinkedIn
List of publications