polars-timeseries 0.9.0

Documentation • Source Code • PyPI

Distance Metrics • Clustering • Forecasting • Imaging • Changepoints • Preprocessing

polars-ts is a batteries-included time series toolkit built on Polars. It gives you Rust-accelerated distance metrics, 10+ clustering algorithms, a full forecasting stack (classical → deep → foundation-model → agentic), Bayesian and causal inference, streaming/online learning, and multi-agent domain applications — all from a single pip install, no heavyweight frameworks required.

Why polars-ts?

At a glance

import numpy as np
import polars as pl
import polars_ts as pts
from datetime import datetime, timedelta

# Build a small panel of 20 hourly series
dates = [datetime(2023, 1, 1) + timedelta(hours=i) for i in range(100)]
df = pl.concat([
    pl.DataFrame({
        "unique_id": [f"s_{i}"] * 100,
        "ds": dates,
        "y": (np.sin(np.linspace(0, 4 * np.pi, 100)) + np.random.normal(0, 0.1, 100)).tolist(),
    })
    for i in range(20)
])

# Cluster by shape similarity
result = pts.auto_cluster(df, methods=["kmedoids", "spectral"], distances=["sbd", "dtw"])
print(result.best_labels)  # DataFrame[unique_id, cluster]

# Forecast with a full ML pipeline
from sklearn.linear_model import Ridge
pipe = pts.ForecastPipeline(Ridge(), lags=[1, 7, 14], rolling_windows=[7], calendar=["day_of_week"])
pipe.fit(df)
forecasts = pipe.predict(df, h=7)

# Detect changepoints
breaks = pts.pelt(df, cost="meanvar", penalty=10)

Column defaults are unique_id, ds, y throughout. Pass id_col=, time_col=, target_col= to override.

Installation

pip install polars-timeseries

Extras for optional features:

pip install "polars-timeseries[clustering]"     # HDBSCAN, DBSCAN, spectral (sklearn + scipy)
pip install "polars-timeseries[forecast]"       # SCUM, auto_arima (statsforecast)
pip install "polars-timeseries[decomposition]"  # Fourier decomposition (polars-ds)
pip install "polars-timeseries[all]"            # Everything

Requires Python 3.12+ and Polars 1.30+.

Quick start

Pairwise DTW distance

import polars as pl
import polars_ts as pts

df = pl.DataFrame({
    "unique_id": ["A"] * 5 + ["B"] * 5,
    "y": [1.0, 2.0, 3.0, 2.0, 1.0,
          1.0, 3.0, 5.0, 3.0, 1.0],
})

result = pts.compute_pairwise_dtw(df, df)

Auto-cluster time series

result = pts.auto_cluster(
    df,
    methods=["kmedoids", "spectral", "kshape"],
    distances=["sbd", "dtw"],
    k_range=range(2, 6),
)
print(result.best_method, result.best_k, result.best_score)
print(result.best_labels)  # DataFrame[unique_id, cluster]

End-to-end forecast pipeline

from sklearn.ensemble import GradientBoostingRegressor
import polars_ts as pts

pipe = pts.ForecastPipeline(
    GradientBoostingRegressor(),
    lags=[1, 2, 7],
    rolling_windows=[7],
    calendar=["day_of_week", "month"],
    target_transform="log",
)
pipe.fit(train_df)
forecasts = pipe.predict(train_df, h=7)
# forecasts: DataFrame[unique_id, ds, y_hat]

Forecast with covariates

pipe = pts.ForecastPipeline(
    Ridge(),
    lags=[1, 2, 7],
    past_covariates=["temperature"],       # lagged automatically
    future_covariates=["is_holiday"],       # looked up from future_df
)
pipe.fit(train_df)  # train_df has temperature + is_holiday columns
forecasts = pipe.predict(train_df, h=7, future_df=future_df)

ARIMA forecasting

# Fit ARIMA(1,1,1) and forecast 12 steps ahead
fitted = pts.arima_fit(df, order=(1, 1, 1))
forecast = pts.arima_forecast(fitted, h=12)

# Or use automatic order selection
forecast = pts.auto_arima(df, h=12, season_length=12)

Changepoint detection

# PELT — multiple changepoints with mean/variance cost
breaks = pts.pelt(df, cost="meanvar", penalty=10)
# breaks: DataFrame[unique_id, changepoint_idx, ds]

# Bayesian Online Changepoint Detection
probs = pts.bocpd(df)

Exponential smoothing

# Holt-Winters seasonal forecast
result = pts.holt_winters_forecast(df, h=12, season_length=12, seasonal="additive")

Mann-Kendall trend test

import polars as pl
import polars_ts as pts

df = pl.DataFrame({
    "group": ["A"] * 10 + ["B"] * 10,
    "y": list(range(10)) + [10 - x for x in range(10)],
})

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

Seasonal decomposition

df = pl.DataFrame({
    "unique_id": ["A"] * 48,
    "ds": list(range(48)),
    "y": [10 + 5 * (i % 12 > 5) + 0.5 * i for i in range(48)],
})

result = pts.seasonal_decomposition(df, freq=12, method="additive")

Features

Distance metrics _{Rust, parallelized via Rayon}

All distance functions return a tidy DataFrame with columns [id_1, id_2, <metric>]. A unified compute_pairwise_distance(method=...) API lets you swap metrics with a single string.

Clustering & classification

Evaluation: silhouette_score, davies_bouldin_score, calinski_harabasz_score

Classification: knn_classify (distance-based k-NN), TimeSeriesKNNClassifier (OOP), KShapeClassifier (centroid-based), RocketClassifier, InceptionTimeClassifier, ResNetClassifier (deep)

Trend & changepoint detection

• Mann-Kendall test — non-parametric trend detection (Rust)
• Sen's slope — robust trend magnitude estimation (Rust)
• CUSUM — cumulative sum changepoint detection (Rust)
• PELT — multiple changepoints with mean/variance/meanvar cost functions
• BOCPD — Bayesian Online Changepoint Detection
• Regime detection — Hidden Markov Model state inference

Decomposition

• Seasonal decomposition — additive or multiplicative (classical)
• Fourier decomposition — harmonic decomposition with configurable frequencies
• Decomposition features — trend/seasonal strength extraction (simple or MSTL)
• Anomaly flagging — residual-based anomaly detection from any decomposition

Feature engineering

• Lag features — create lagged versions of a target column per group
• Rolling features — rolling window aggregations (mean, std, min, max, sum, median, var)
• Calendar features — extract day_of_week, month, quarter, is_weekend, etc.
• Fourier features — sin/cos pairs for seasonal modelling
• Target encoding — smoothed categorical encoding by target mean
• Holiday features — binary holidays + distance-to-holiday (requires holidays package)
• Interaction features — cross-term column generation
• Time embeddings — cyclical sin/cos encoding for time components

Target transforms

• Log transform — log1p / expm1 with automatic validation and lossless inversion
• Box-Cox transform — parametric power transform with configurable lambda
• Differencing — configurable order and seasonal period with metadata for lossless inversion

All transforms are group-aware, invertible, and accessible via the df.pts namespace.

Data preprocessing

• Missing value imputation — forward/backward fill, linear interpolation, mean, median, seasonal
• Outlier detection — z-score, IQR, Hampel filter, rolling z-score
• Outlier treatment — clip (winsorize), median replacement, interpolation, null
• Temporal resampling — downsample/upsample with configurable aggregation

Validation strategies

• Expanding window CV — growing training window cross-validation
• Sliding window CV — fixed-size training window cross-validation
• Rolling origin CV — general rolling-origin with configurable initial/fixed train size

Forecasting

• SCUM — ensemble model combining AutoARIMA, AutoETS, AutoCES, and DynamicOptimizedTheta
• ARIMA/SARIMA — explicit (p,d,q) order via statsmodels (arima_fit/arima_forecast) or automatic selection via statsforecast (auto_arima)
• Baseline models — naive, seasonal naive, moving average, and FFT-based forecasts
• Exponential smoothing — SES, Holt's linear, Holt-Winters (additive/multiplicative, Rust-accelerated)
• Multi-step strategies — RecursiveForecaster and DirectForecaster
• ForecastPipeline — end-to-end ML pipeline with feature engineering + transforms
• GlobalForecaster — cross-series panel model with optional ID encoding
• N-BEATS / PatchTST — native deep learning forecasters (NBEATSForecaster, PatchTSTForecaster, requires torch)
• Native multivariate deep — MultivariatePatchTST, iTransformerForecaster for jointly forecasting correlated series
• Time-LLM / LLM-PS — LLM-reprogrammed forecasting adapters (requires torch)
• Foundation models — Chronos, TimesFM, Moirai zero-shot forecasting
• Agentic forecasting — TimeSeriesScientist multi-agent pipeline (Curator → Planner → Forecaster → Reporter)

Probabilistic forecasting

• QuantileRegressor — one model per quantile level with CRPS-compatible output
• Conformal prediction — distribution-free intervals with coverage guarantees
• EnbPI — Ensemble Batch Prediction Intervals with adaptive online updates

Ensembling

• WeightedEnsemble — equal, manual, or inverse-error-optimized weights
• StackingForecaster — meta-learner trained on out-of-fold predictions

Forecast evaluation & diagnostics

• Metrics — MAE, RMSE, MAPE, sMAPE, MASE, CRPS
• Kaboudan metric — model robustness evaluation via block-shuffle backtesting
• Bias detection & correction — mean, regression, quantile mapping
• Calibration diagnostics — calibration table, PIT histogram, reliability diagram
• Residual diagnostics — ACF, PACF, Ljung-Box test
• Permutation importance — model-agnostic feature importance

Multivariate & hierarchical

• VAR — Vector Autoregression with OLS fitting and multi-step forecasts
• Granger causality — F-test for causal relationships between series
• GARCH — volatility modelling and conditional variance forecasting
• Forecast reconciliation — bottom-up, top-down, middle-out, MinTrace (OLS/shrink/covariance/CV), and PERMBU
• Grouped / cross-sectional hierarchies — reconcile across multiple non-nested grouping dimensions

Bayesian methods

• Kalman Filter / RTS Smoother — linear state-space models
• BSTS — Bayesian Structural Time Series (local level/trend + seasonality + regressors)
• Bayesian ETS — exponential smoothing with posterior predictive
• Bayesian VAR — vector autoregression with Minnesota/Normal-Wishart priors
• Gaussian Process regression — non-parametric Bayesian forecasting
• MCMC wrapper — adapter for NumPyro/PyMC backends
• Unscented / Ensemble Kalman Filter — nonlinear state-space models
• Particle Filter / SMC — fully nonlinear/non-Gaussian
• Bayesian anomaly scoring — posterior predictive p-values, Bayes factors

Causal inference

• CausalImpact — Bayesian structural time series counterfactual analysis
• Synthetic control — donor-pool weighted counterfactual estimation
• Placebo tests — significance testing via counterfactual permutation

Backtesting

• Unified backtest framework — backtest and compare_models for fit → predict → score across rolling windows with multiple models

Streaming / online learning

• StreamingETS — online exponential smoothing (SES, Holt, Holt-Winters) with single-observation updates
• StreamingKalmanFilter — incremental Kalman filtering for state-space models
• StreamingGlobalForecaster — incremental panel model for SGD-compatible estimators
• SlidingWindowManager — memory-efficient windowed state management

Anomaly detection

• Decomposition-based — residual threshold anomaly flagging
• Isolation Forest — unsupervised anomaly detection on engineered features
• RL anomaly agents — multi-agent consensus (AnomalyOrchestrator over Z-score, rolling-std, and MAD agents) for adaptive thresholding

Reinforcement learning & multi-agent

• Portfolio MARL — cooperating risk, return, and allocation agents (MARLOrchestrator, PortfolioEnv) for decision optimization
• ForecastEnv / PortfolioEnv / AnomalyEnv — Gymnasium-compatible environments for RL over time series

Domain agents

Hierarchical multi-agent systems for specific verticals, built on the core forecasting stack:

• Energy / demand — EnergyGridOrchestrator: regional → grid → household hierarchy with weather, renewable, and demand-response agents
• Supply chain — SupplyChainOrchestrator: demand sensing, promotion effects, inventory, and multi-echelon coordination
• Healthcare — ClinicalOrchestrator: EHR/vital monitoring, sepsis warning, escalation, and treatment agents (with federated averaging)
• Industrial IoT — MaintenanceOrchestrator: spectral features, health index, RUL estimation, and predictive maintenance scheduling

Model registry & experiment tracking

• ModelRegistry — version, persist, and load models with metadata
• Experiment / Run — track parameters, metrics, and artifacts across runs

Time series imaging

• Imaging transforms — GASF/GADF (to_gasf, to_gadf), Markov Transition Field (to_mtf), recurrence plots (to_recurrence_plot), spectrogram/scalogram, and path-signature images
• RQA & signature features — recurrence quantification (rqa_features) and signature features for downstream models
• Vision embeddings — extract_vision_embeddings for pretrained-CNN features over imaged series

Integration adapters

• NeuralForecast — convert to/from N-BEATS, PatchTST, N-HiTS format
• PyTorch Forecasting — convert to/from TFT, DeepAR format
• HuggingFace — convert to Dataset for Chronos, TimesFM, Lag-Llama
• Chronos / MOMENT embeddings — foundation model feature extraction for clustering
• Foundation forecast — ChronosForecaster, TimesFMForecaster, MoiraiForecaster (zero-shot)
• LLM forecast — TimeLLMForecaster, LLMPSForecaster (reprogrammed LLM)
• ForecastEnv — Gymnasium-compatible RL environment for decision making

Tutorials

The notebooks/ directory contains 14 end-to-end tutorials:

Development

git clone https://github.com/drumtorben/polars-ts.git
cd polars-ts
uv sync
uv pip install -e .
uv run pytest

Code quality

Pre-commit hooks run via prek (Rust reimplementation of pre-commit) or standard pre-commit — both read .pre-commit-config.yaml:

# Option A: prek (faster)
uv tool install prek
prek run --all-files

# Option B: standard pre-commit
pre-commit run --all-files

Type checking

# mypy (authoritative)
uv run mypy polars_ts/

# ty (fast, informational — beta)
uvx ty check polars_ts/

License

MIT