timeseries

TimeSeries scorer contract, distinct from legacy scored-table components.

score consumes observations or residuals. score_from_prediction aligns actual and predicted timestamps, then scores their absolute residuals unless a concrete scorer specifies another transformation. All methods accept one series or equal-length sequences of series.

Cauchy negative log likelihood using robust median and half-IQR estimates.

Signed per-component difference actual - predicted (no fitting).

Exponential negative log likelihood for nonnegative prediction samples.

Gamma negative log likelihood with method-of-moments sample parameters.

Gaussian negative log likelihood using sample mean and standard deviation.

Euclidean distance to the closest fitted normal-window cluster center.

Laplace negative log likelihood using sample median and absolute deviation.

Negative log likelihood fitted separately to prediction samples per point.

Requires deterministic observations and stochastic predictions (at least two samples). Degenerate scales are floored by min_scale; observations outside a distribution's support receive the finite cap -log(tiny). window averages successive point scores over a causal trailing window.

Absolute component residuals or their vector norm at each timestamp.

component_wise=True returns one absolute score per component. Otherwise ord chooses the numpy vector norm (default Euclidean).

Structural contract accepted by the optional PyOD adapter.

Poisson negative log likelihood for nonnegative integer observations.

Adapt a PyOD-compatible estimator's fit/decision_function to windows.

model must implement fit(X) and decision_function(X) with larger values meaning more anomalous. A deep copy is fitted per component so that models never share fitted state. Import/install PyOD only when choosing a PyOD model; this adapter itself has no PyOD dependency.

One-dimensional Wasserstein distance from each window to training values.

Components are flattened into one empirical distribution unless component_wise=True. This is not multivariate optimal transport.

Fit and score sliding windows without crossing series boundaries.