Jackson
Using Recursive Variable Selection in Conditional Random Forests to Construct an Overall Seasonality Test (303717)
*Daniel Ollech, Deutsche BundesbankKarsten Webel, Deutsche Bundesbank
Keywords: conditional random forest; seasonality tests, JDemetra+, simulation study, supervised machine learning
Over the years, a large variety of statistical tests has been developed to check the presence of seasonality in a given time series. Using a conditional random forest based recursive variable selection scheme, we first identify the most informative seasonality tests. These tests then become the building blocks of an overall test that is designed to achieve the classification accuracy of the random forests whilst at the same time providing interpretable results. We show that a sensible decision rule can be derived by combining the Friedman test with the modified QS-test, when these are computed on the residuals obtained from fitting a non-seasonal ARIMA model to the input data. The resulting overall seasonality test has low type I and type II misclassification rates and evades excessive complexity.