HOST: Harmonic oscillator seasonal-trend model for analyzing the reoccurring nature of extreme events

Abstract

The Harmonic Oscillator Seasonal-Trend (HOST) model allows for automated analysis and pattern recognition in time-series data with varying time domains. Developed as a Python package, the software performs the decomposition of data into short- and long-term components and uses a range of modified sine waves to model both behaviors. Waveform synthesis is then performed to compose the final model, incorporating both timeframes. The model allows for the extraction of n harmonics from the data, or signal (representing any time-series data) analysis, as well as parametric assessment, that includes: (1) occurrence analysis with related decision thresholds determined during topological analysis; (2) magnitude; and (3) extremes assessment. Calculations are performed automatically after the user specifies the study's needs. Performance varies depending on the dataseries used, with long-term patterns usually reaching a Kling-Gupta efficiency >0.9 and short-term patterns being around 0.5. A decrease in accuracy in the testing dataset is observed for binary occurrence classification, associated with low event occurrence during this period, which can be partially addressed by extending the test set length.