Analysing the efficiency of deseasonalisation methods in removing seasonal bias in outliers

R. Netshiomvani${}^a$, S. C. Liebenberg${}^a$ and G. L. Grobler${}^b$

${}^a$Focus Area for Pure and Applied Analytics, North-West University ${}^b$ Unit for Data Science and Computing, North-West University

Deseasonalisation is a common preprocessing step in climate time series analysis, particularly when the focus is anomaly detection, where seasonal patterns can mask or bias extreme events. Most deseasonalisation approaches focus on removing the seasonal mean, implicitly assuming that this is sufficient to eliminate seasonal patterns. However, many meteorological variables exhibit seasonality in both mean and variance, which may bias the detection of outliers. In this talk we will show whether deseasonalisation methods effectively remove seasonal bias in outliers from bivariate climate time series obtained from weather station and ERA5 data. Daily wind speed data from the South African Lowveld are analysed using paired observations from both sources. Standard deseasonalisation methods are used to remove seasonality in the mean, while less well-known approaches also account for seasonal variance. In addition, the effects of serial correlation and conditional heteroscedasticity are removed by applying an ARMA–GARCH model and analysing the resulting standardised residuals. To assess the presence of seasonal bias, the monthly distribution of detected outliers is evaluated using a chi-square goodness of fit test. The results suggest that reliable anomaly detection in bivariate climate data may require deseasonalisation methods that account for seasonality in both the mean and the variance.

Keywords: Deseasonalisation, Seasonal bias, Wind speed.