Virtually each seasonal adjustment software includes an ensemble of seasonality tests for assessing whether a given time series is in fact a candidate for seasonal adjustment. However, such tests are certain to produce either the same resultor conflicting results, raising the question if there is a method that is capable of identifying the most informative tests in order (1) to eliminate the seemingly non-informative ones in the former case and (2) to find a final decision in the more severe latter case. We argue that identifying the seasonal status of a given time series is essentially a classification problem and, thus, can be solved with machine learning methods. Using simulated seasonal and non-seasonal ARIMA processes that are representative of the Bundesbank's time series database, we compare certain popular methods with respect to accuracy, interpretability and availability of unbiased variable importance measures and find random forests of conditional inference trees to be the method which best balances these key requirements. Applying this method to the seasonality tests implemented in the seasonal adjustment software JDemetra+ finally reveals that the modifiedQSand Friedman tests yield by far the most informative results.
binary classification conditional inference trees correlated predictors JDemetra+ simulation study supervised machine learning