Automatic threshold selection for POT analysis I have got > 100 river daily flow time series and I need to select a threshold for each of them in order to perform a Peak Over Threshold extreme value analysis.
I know that I can select a threshold manually with specific R packages but I would like to know if there exists an automatic method which will calculate for me the most suitable threshold for each of the 100 time series.
e.g. I could set up a 90% threshold based on the max daily flow value but it could be a very extreme one and therefore I will miss all the other ones.
Any suggestion?
Thanks
 A: You should have a look to this article by C. Scarrott and A. Mac
Donald A Review of Extreme Value Threshold Estimation and Uncertainty
Quantification. A simple and
reasonable idea is to combine two models: one for the bulk, one for the
tail. At least, this makes the threshold choice less crucial than it
is in standard POT. As you will see, some well-documented R packages
on CRAN relate to the methods described in the article: evmix by
C.J. Scarrott and Y. Hu and condmixt by J. Carreau.
Another reference is chap. 4 in the book Extreme Value
Modeling and Risk Analysis.
A: I suggest building a Transfer Function model (regression with steroids!) to predict daily values based upon memory and possible predictors like day-of-the-month , week-of-the-month , month-of-the-year et al . After forming a suitable model which of course would include possible intervention variables like pulses ,  level shifts , local time trends I would then forecast and obtain prediction intervals for each period in the future by re-sampling the model errors (allowing for anomalies in the future) and obtain a family of forecasts (frequency distributions) for each period out. I would then examine these monte-carlo distributions for their statistical properties (eg. max/min/median ) to assess threshold values. In this way you can manage your own analysis of the possible future values.
