Layer 1: Time Series Anomaly Detection

PyOD ships 5 stable time series detectors plus 2 experimental (SAND, AnomalyTransformer), all using the same fit/predict/decision_function API. The stable five are what ADEngine.list_detectors(data_type='time_series') returns; the experimental pair is available by class import but not routed by ADEngine yet. Rankings from TSB-AD benchmark [ALP24] (NeurIPS 2024, 1,070 datasets, 40 algorithms).

Input format: numpy array of shape (n_timestamps,) for univariate or (n_timestamps, n_channels) for multivariate.

Output: decision_scores_ of shape (n_timestamps,).

from pyod.models.ts_kshape import KShape
clf = KShape(window_size=20)
clf.fit(X_train)
scores = clf.decision_scores_

---

Detectors

Type	Detector	Year	Ref
Windowed	TimeSeriesOD: any PyOD detector on sliding windows	2026
Subsequence	MatrixProfile: STOMP, transductive	2016	Yeh et al.
Frequency	SpectralResidual: FFT saliency	2019	Ren et al.
Clustering	KShape: shape-based, #2 TSB-AD overall	2015	Paparrizos et al.
Streaming	SAND: drift adaptation	2021	Boniol et al.
Deep Learning	LSTMAD: LSTM prediction + Mahalanobis	2015	Malhotra et al.
Deep Learning	AnomalyTransformer: attention discrepancy	2022	Xu et al.

---

Transductive vs Inductive

Most time series detectors support both fit() and decision_function(X_test). One exception:

• MatrixProfile is transductive: use decision_scores_ after fit(), no out-of-sample predict().