Regime shift detection group

Report of the Regime Shift Detection Group

Leader and lecturer: Sergei Rodionov, Rapporteur: Abigail Mcquatters-Gollop and Violeta Velikova, Participants: Nesho Chipev, Snejana Moncheva, Galina Shtereva, Georgiy Daskalov, Vesselina Mihneva, Sonia Ouzunova, Violin Raikov, Elisaveta Peneva, Viktor Nikolsky, Laura Boicenko, Radka Mavrodieva, Svetla Bratanova, Alexandra Uzunova.

Theme:Overview of Regime Shift Detection Methods, regime shift detection software (author Sergei Rodionov), practical exercises.

Table of methods for detecting regime shifts and regime shift detection software are downloadable from www.beringclimate.noaa.gov .
Regime shift detection methods:
(http://www.beringclimate.noaa.gov/regimes/Regime_shift_methods_list.htm for a complete list of methods, contactperson Sergei Rodionov).
• Methods for detecting shifts in the mean:

o Students T test
▪ Most common, robust, well established but you need to test for a regime shift by picking a year manually which violates the assumption that the change point should not be pre-assigned (ie each year should be given equal weight) and is not automatic.
o Bayesiananalysis
▪ Accounts for uncertainty of estimating change points and can be used for predictions but requires a mathematical model of the data and works best with a single change point scenario.
o Mann-Whitney U test
▪ Non parametric and easy to use but the data needs to be detrended.
o Lepage test
▪ Non parametric and similar toMann-Whitney U and Pettitt tests. Can be used to detect different shaped trends as well as regime shifts but is a single change point scenario. Not automatic and a single change point scenario.
o Wilcoxon rank sum
▪ Non parametric and similar to Mann-Whitney U test.
o Pettitt test
▪ Non parametric and similar to Mann-Whitney U test. Can be used to detectdifferent shaped trends as well as regime shifts but is a single change point scenario.
o Mann-Kendall test
▪ Also non parametric but not automatic and data must be detrended.
o Standard normal homogeneity test (SNHT)
▪ Very popular and mostly used for meterological time series. Not good when change points are close together in time, there aremore than 4 change points, or methodology of sampling or analysis has changed over time.
▪ Free package available on net: OnClim. Contains the standard normal homogeneity test.
o Regression based approach
▪ Better than SNHT when multiple change points exist but not sensitive to small shifts or shifts that occur within 10 points of one another.
oCUSUM test
▪ Cumulative deviation test. Easy to use but works with anomalies and using different base periods may affect results.
o Oerlemans method
▪ Based on comparison of an a priori described typical change with an interval of a given time series. Can be applied to any time series and results are easy to compare but a statistical significance test can’tbe constructed and requires the best fitting of a curve that represents an idealized break in the data. No longer used.
o Signal-to-Noise ratio
▪ A regime shift is defined when the signal to noise ration exceeds 1. Simple and easy to use but only works for a single change point.
o Intervention analysis
▪ Extension of ARIMA method. Allows forquantitative estimate for statistical significance of step interventions while accounting for autocorrelation in the time series but, like with t test, the time and type of intervention should be specified in advance.
o Markov chain Monte Carlo
▪ Strong basis on Bayesian approach. Must find best model to describe your data which makes method complex and difficult to use....