65 features, one config file: TSFEL does the boring part of time series ML

What it does TSFEL extracts 65+ features from time series across four domains: statistical, temporal, spectral, and fractal. You describe what you want in a configuration dict, point it at a signal (an ECG, an accelerometer stream, whatever), and get back a feature matrix ready for scikit-learn-style pipelines. The config file is storable, so the same feature set can be reproduced later — a small thing that saves real arguments.

The interesting bit Each feature ships with a computational cost rating, so you can estimate extraction time before running it on your 40-hour EEG recording. The fractal features (Hurst exponent, DFA, Higuchi dimension) are the unusual part — most feature libraries stop at mean and kurtosis.

Key highlights

• 65+ features across statistical, temporal, spectral, and fractal domains, including MFCCs and wavelet features
• Config-file-driven extraction for reproducibility
• Computational cost estimates per feature, rated 1–3
• Every feature is unit tested and documented individually
• Custom features are straightforward to add and contribute back

Caveats

• Fractal features are disabled by default — they need longer signals to be meaningful
• Lempel-Ziv complexity is also off by default due to slow execution

Verdict If you’re doing HAR, biosignal, or audio classification and don’t want to hand-roll your twentieth spectral feature, this is the library. If you need forecasting models rather than features, look elsewhere — TSFEL stops at the feature matrix.