CHANGELOG.md

v0.8.0 (2026-05-05)

Added

• LLM-based forecasting (#155): TimeLLMForecaster (patch → cross-attention with prototypes → decode) and LLMPSForecaster (multi-scale CNN pattern extraction → decode). 15 tests. PR #200.
• Agentic forecasting (#156): TimeSeriesScientist multi-agent pipeline with Curator, Planner, Forecaster, and Reporter agents. 87 tests with TDD edge-case coverage. PR #188, #197.
• KASBA Rust port (#189): Port of KASBA clustering algorithm into src/kasba/ — MSM distance, elastic k-means++ init, triangle inequality fast assignment, stochastic barycenter averaging, empty cluster recovery. 16 Rust unit tests. PR #198, #199.
• Contrastive clustering (#154): Self-supervised contrastive learning with NT-Xent loss. PR #187.
• Deep clustering (#157): DEC/IDEC autoencoder-based deep clustering. PR #187.
• Foundation model forecasting (#151): ChronosForecaster, TimesFMForecaster, MoiraiForecaster for zero-shot forecasting. PR #186.
• Native DL forecasters (#150): NBEATSForecaster and PatchTSTForecaster with fit/predict API. PR #186.
• Deep classifiers (#152): RocketClassifier, InceptionTimeClassifier, ResNetClassifier. PR #186.
• Causal inference (#153): CausalImpact (Bayesian structural counterfactual) and SyntheticControl with placebo tests. PR #183.
• Covariates support (#149): Past, future, and static exogenous variables in ForecastPipeline. PR #181.
• Backtesting framework (#167): Unified fit → predict → score across rolling windows. PR #184.
• Bayesian methods (9 modules): Kalman Filter (#141), BSTS (#142), UKF/EnKF (#143), Bayesian ETS (#169), Bayesian VAR (#170), MCMC wrapper (#174), GP regression (#175), Bayesian anomaly scoring (#176), Particle Filter (#180).

Infrastructure

• Registry-based __init__.py for lazy imports (#144).
• Shared lazy-import helper for submodules (#145, PR #172).
• Notebook normalization pre-commit hook (#146, PR #172).
• Consolidated test fixtures (#147, PR #173).

Documentation

• Notebooks 11 (imaging), 12 (advanced features), 13 (agentic forecasting).

---

v0.7.0 (2026-04-25)

Features

• Add HDBSCAN/DBSCAN density-based clustering with precomputed distance matrices (#106).
• Add agglomerative/hierarchical clustering with dendrograms (#107).
• Add k-means with DBA (DTW Barycentric Averaging) clustering (#108).
• Add CLARA/CLARANS scalable k-medoids variants (#109).
• Add ROCKET/MiniRocket feature extraction for clustering (#111).
• Add U-Shapelet clustering for time series (#112).
• Add Chronos/MOMENT embedding adapters for clustering (#113).
• Add spectral clustering with SBD kernel (K-Spectral Centroid) (#125).
• Add auto_cluster automated clustering pipeline selection (#126).

Fixes

• Bump rand 0.8.5 → 0.8.6 to resolve GHSA-cq8v-f236-94qc (#124).
• Upgrade rustls-webpki 0.103.12 → 0.103.13 for GHSA-82j2-j2ch-gfr8 (#110).
• Add scipy to clustering optional dependency (required by spectral clustering).
• Add pytest.importorskip guards for tests requiring scipy/sklearn.
• Fix mypy type errors in auto_cluster module.

Documentation

• Replace 7 API-demo notebooks with 10 theme-based tutorials (#89).
• Add polars-ts logo image for README banner (#88).

Tests

• Add SCUM ensemble model tests (#90).

v0.6.0 (2026-04-21)

Features

• Add ARIMA/SARIMA forecasting with explicit (p,d,q) order control via statsmodels.SARIMAX (arima_fit, arima_forecast).
• Add automatic ARIMA order selection via statsforecast.AutoARIMA (auto_arima).
• Both ARIMA backends raise helpful ImportError when optional dependencies are missing.

Performance

• Accelerate exponential smoothing (SES, Holt, Holt-Winters) with Rust implementation (#86).
• Accelerate k-medoids PAM clustering with Rust implementation (#84).

Improvements

• CI: replace pre-commit with prek (Rust reimplementation) for faster code-quality checks.
• CI: add ty type checker as non-blocking informational job alongside mypy.
• Add [tool.ty] configuration section in pyproject.toml.
• Document prek and ty local developer workflows in README.

Tests

• Expand coverage for HF adapter, clustering, calibration, and adapters (#85).
• Expand coverage for changepoint, clustering, adapters, and standalone modules (#83).
• Add 8 ARIMA tests (5 statsmodels, 3 statsforecast-gated).

v0.5.0 (2026-04-16)

Features

• Add KShape time series clustering using shape-based distance with centroid computation.
• Add KShape time series classifier.
• Add k-Medoids (PAM) time series clustering (kmedoids) using any of the 12 distance metrics.
• Add k-Nearest Neighbors time series classification (knn_classify) using any of the 12 distance metrics.
• Add SBD (Shape-Based Distance) metric (compute_pairwise_sbd).
• Add Frechet distance metric (compute_pairwise_frechet).
• Add EDR (Edit Distance on Real Sequences) metric (compute_pairwise_edr).

Improvements

• Add shared distance dispatch utility (_distance_dispatch) for reuse across clustering and classification.
• Upgrade Rust dependencies: pyo3 0.25, polars crate 0.49.1.
• Add py.typed marker for PEP 561 type hint distribution.
• Add lazy import system for optional dependencies (forecast, decomposition).
• CI: add coverage reporting, MkDocs deployment workflow, polars compatibility matrix (1.30–1.33).

Tests

• Add 85 tests for k-NN classification covering correctness, custom columns, and multiple metrics.
• Add 84 tests for k-Medoids clustering covering output, correctness, edge cases, and multiple metrics.
• Add 90 tests for KShape clustering.
• Add 85 tests for KShape classifier.
• Add tests for SBD, Frechet, EDR distance metrics.
• Add unified API tests for all 12 distance metrics.
• Add 52 lazy import tests for optional dependency handling.

v0.4.0 (2026-04-14)

Features

• Add Sen's slope estimator (sens_slope) — robust median-of-pairwise-slopes trend magnitude.
• Add CUSUM changepoint detection (cusum) — cumulative sum control chart for mean shift detection.
• Add anomaly flagging to decomposition methods via anomaly_threshold parameter.
• Add ERP (Edit Distance with Real Penalty) distance metric (compute_pairwise_erp).
• Add LCSS (Longest Common Subsequence) distance metric (compute_pairwise_lcss).
• Add TWE (Time Warp Edit Distance) distance metric (compute_pairwise_twe).
• Add FastDTW approximate algorithm (method="fast", param=radius).
• Add Sakoe-Chiba band constraint (method="sakoe_chiba", param=window_size).
• Add Itakura parallelogram constraint (method="itakura", param=max_slope).
• compute_pairwise_dtw now accepts optional method and param arguments (backward compatible).

Improvements

• Add unified compute_pairwise_distance(method=...) API — single entry point for all 9 distance metrics.
• Deduplicate Rust distance code into shared utils.rs (grouping, hashing, parallel pairwise).
• Optimize all distance metrics to O(m) memory with two-row DP.
• Add #[pyo3(signature)] annotations for Rust safety with pyo3 0.24+.
• Add pyarrow as core dependency for reliable Arrow interchange.
• Lightweight core: optional dependencies (forecast, decomposition) are lazy-loaded.
• CI: add minimal-deps job, polars compatibility matrix, clippy, and code-quality checks.

Fixes

• Upgrade Rust dependencies (pyo3 0.24, pyo3-polars 0.21, polars crate 0.48).
• Pin Python polars to >=1.30.0,<2.0.0 for ABI compatibility.
• Fix trailing space in freqs docstring causing mkdocs strict build failure.

Tests

• Add 286 tests covering all distance metrics, decomposition, trend, changepoint, and edge cases.
• Add input validation, edge case, and parallelism stress tests for distance metrics.

v0.3.0

Features

• Implement Seasonal Decomposition.
• Implement Fourier Decomposition.
• Implement Naive Dynamic Time Warping.

v0.2.0

Features

• Implement Mann-Kendall's Trend Statistic.

Chore

• Make library usable on PyPI with the Rust expressions.

v0.1.0

Features

• Implement Kaboudan metric.

Documentation

• Add automatic references to docstrings.
• Access docs under https://drumtorben.github.io/polars-ts/.

Chore

• Initialize Repo.