Overview

In climatology regime shifts are defined as rapid reorganizations of the climate system from one relatively stable state to another. These regimes may last for several decades, as it is clearly seen in the Pacific Decadal Oscillation. Perhaps the first reported climate shift in instrumental records occurred in the 1920s (Rodionov 1985, Drinkwater 2006). Since then a number of decadal climate regimes has been observed (Rodionov 2002). The concept of abrupt shifts in the climate system attracted much attention after the Great Pacific Climate Shift of 1976/77, which also marked the beginning of a warming trend in the global climate. Interestingly, that shift became apparent only many years after it had actually occurred (Kerr 1992). Detecting regime shifts as soon as possible is critically important in climate forecasting.

There are a number of methods designed to detect regime shifts in both individual time series and entire systems (Rodionov 2005a). For the overwhelming majority of these methods, however, their performance deteriorates toward the ends of time series. Rodionov (2004) developed a new method, called STARS for short based on a sequential t-test analysis that can signal a possibility of a regime shift in real time. Later, the method was expanded to include shifts in variance using the sequential F-test analysis (Rodionov, 2005b).

Rodionov et al. (2004) and Rodionov and Overland (2005) discuss an application of the method to the Bering Sea ecosystem. Examples of other applications of the method can be found in Litzow (2006), Gergis et al. (2006), D'Arrigo and Wilson (2006), Wilson et al. (2006), Daskalov et al. (2007), Lo and Hsu (2007), Gardner and Sharp (2007). For more references click here and here.

More recently (Rodionov 2015), a third module was developed capable of detecting regime shifts in correlation, and thus opening a new way to analyze and model the relationships between time series. The correlation coefficient is a principal statistical tool and the most widely used measure of a relationship between two variables. When using this tool, it is assumed that the nature of the relationship is linear and can be modelled by a simple linear regression. Large natural systems, however, such as the climate system, exhibit behaviors that are far more complex and seem to require equally complex, possibly nonlinear, models to simulate them adequately. An alternative view is that such systems may be governed by simple rules, but the parameters of those rules are changing, possibly in an abrupt fashion. In this paradigm, a major task would be to determine the timing of those abrupt changes that now can be done with STARS. For example, the applications of STARS to the Arctic Oscillation and the ENSO effect on the northern Tropical Atlantic reveal that the climate system experienced significant structural changes in the late 1990s. It becomes increasingly clear that while the climate regime shift in the late 1970s was primarily about the change in the mean level of fluctuations, the transformations in the climate system that started in the late 1990s had more to do with discontinuities in many well-established relationships between macroclimatic variables.

The STARS software program is written in Visual Basic for Application (Excel). It can detect shifts in three statistics - mean, variance, and correlation - of the time series. To reduce the effect of autocorrelation (or red noise) it is recommended to use prewhitening of the time series as discussed in Rodionov (2006).

Software requirements: STARS requires Excel 2007 or higher for Windows (does not work with Excel for Mac).