> ML_LIBRARY // TSFRESH_v1.0
tsfresh
Blue Yonder / tsfresh developers — Automated feature extraction and hypothesis-driven feature selection for time series.
time-series-forecastingv0.20.2MITqualified
Model Training
Accelerators:
CPU
Distributed Training:Yes
Model Inference
Inference Accelerators:
CPU
Deployment Targets:server
What It Does
- +Automatic extraction of up to 800 statistical features per time series (Fourier coefficients, autocorrelation, entropy, peak counts)
- +Hypothesis testing (Benjamini-Yekutieli / Benjamini-Hochberg) to filter statistically significant features against regression/classification targets
- +Parallelized multi-core and distributed extraction via Dask
- +Seamless conversion of raw sensor waveforms into tabular feature matrices for scikit-learn and XGBoost
What It Does Not Do
- -Natively fit neural network forecasting models
- -Stream real-time updates at sub-second tick intervals
- -Operate on unstructured images or natural language
>Suitable Work Types
- Predictive maintenance identifying bearing failures from vibration waveforms
- Medical ECG and EEG classification
- Feature engineering before gradient boosting model training
>Unsuitable Work Types
- Real-time streaming microservices needing sub-10ms response times
- Pure text NLP modeling
Data Residency Implications
Completely local in-process execution. Zero telemetry.
Security Considerations
MIT license with permissive commercial use rights.
Operational Profile & Known Limitations
Maturity:mature
Learning Curve:low
Ops Complexity:low
Cost Tier:free-oss
> Known Limitations:
- Extracting 800 features on long high-frequency series is computationally expensive and memory-intensive; use extract_relevant_features with sample limits.
Associated Incident Patterns (Incidentpedia)
Enforce safeguards and monitoring to guard against these documented real-world failure modes:
> Primary Evidence & Benchmark Citations
tsfresh Documentationofficial-docs • >=0.19.0, <=0.20.x
2026-09-25HIGH