Changepoint & Anomaly Detection Guide

polars-ts provides trend tests, changepoint detection, regime inference, and anomaly detection — all group-aware and Polars-native.

Trend detection

Non-parametric test for monotonic trend. Returns a struct with tau, p_value, and trend (increasing/decreasing/no trend). Implemented in Rust.

import polars as pl
import polars_ts as pts

result = df.group_by("unique_id").agg(
    pts.mann_kendall(pl.col("y")).alias("trend"),
)

Robust trend magnitude estimator (median of all pairwise slopes). Implemented in Rust.

result = df.group_by("unique_id").agg(
    pts.sens_slope(pl.col("y")).alias("slope"),
)

Cumulative sum control chart for detecting mean shifts. Implemented in Rust.

result = pts.cusum(df)

Pruned Exact Linear Time algorithm for finding multiple changepoints with configurable cost functions.

result = pts.pelt(df, cost="meanvar", pen=10)

Cost functions: "mean", "var", "meanvar".

Bayesian Online Changepoint Detection — detects changepoints in a streaming fashion using a Student-t model.

result = pts.bocpd(df, hazard=250, mu0=0, kappa0=1, alpha0=1, beta0=1)

Hidden Markov Model state inference via Baum-Welch EM.

result = pts.regime_detect(df, n_regimes=3)

Flag anomalies from decomposition residuals exceeding a threshold.

result = pts.seasonal_decomposition(df, freq=24, method="additive", anomaly_threshold=2.0)

Unsupervised anomaly detection on engineered features.

result = pts.isolation_forest_detect(df, contamination=0.05)