Answers are grounded in this paper and cite their sources.
Extracting features from time series data is a crucial step in time series analysis. As literature defines hundreds of features, various Python packages have been developed in the last decade that automatically extract feature sets. Several comparison studies have been published recently, highlighting computational differences and analyzing the extracted feature sets using small-scale or synthetic datasets. This paper benchmarks six Python packages for automatic feature extraction – tsfresh, TSFEL, seglearn, pycatch22, tsfeatures and Kats – on four real-world manufacturing datasets, jointly evaluating computational footprint and downstream classification performance. Computational effort, peak memory usage, and runtime are measured for single-threaded and multithreaded extraction. Downstream performance is assessed under five classifier settings: four fixed classifiers (Random Forest, XGBoost, logistic regression, and SVM) with search-free defaults, and TPOT as an AutoML scenario. The results show that the packages differ heavily in computational cost, with seglearn offering the most favorable profile under tight resource budgets and TSFEL achieving the lowest computational effort among packages supporting multithreading, albeit at the cost of reduced predictive reliability on more complex datasets. A mixed-effects analysis reveals a statistically significant package effect on balanced accuracy in all five classifier settings, separating a strong top tier (tsfresh efficient, seglearn, and pycatch22) from weaker TSFEL subsets. Based on these findings, we derive practitioner-oriented recommendations: seglearn default for resource-constrained settings, tsfresh efficient when broader feature coverage is warranted, pycatch22 when feature compactness and interpretability are important, and TSFEL when extraction cost is the dominant constraint and some reduction in predictive performance is acceptable.
No takes yet. Share an insight, caveat, or question.
Seidler et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: