PipelineTS

⚡ Simplest and most professional TimeSeries forecasting + AutoML.
One line to forecast. Professional when you need it.

---

Installation

pip install PipelineTS          # core (ML + stats)
pip install PipelineTS[torch]   # + PyTorch NN models

Python ≥ 3.9 required.

---

Quick Start

One-line forecast

from PipelineTS import forecast

pred = forecast("sales.csv", n=12)   # auto-infers columns, cleans data, runs AutoML

Diagnose → Forecast → Backtest

from PipelineTS import diagnose, forecast, backtest

report = diagnose("sales.csv", horizon=12)   # data readiness check
pred   = forecast("sales.csv", n=12)         # AutoML prediction
score  = backtest("sales.csv", n=12)         # walk-forward evaluation

print(report["status"])      # READY / WARNING / NOT_READY
print(score["summary"])      # mean ± std of MAE across folds

Progressive professional control

from PipelineTS import AutoForecast

model = AutoForecast(
    time_col="date", target_col="value",
    horizon=12,
    preset="high_quality",   # fast | medium_quality | high_quality | best_quality
    quantile=0.9,            # 90% prediction intervals
)
model.fit(data)
pred = model.predict()

print(model.leader_board_)   # ranked model comparison
print(model.strategy_)       # routing decisions

model.save("forecaster.pts")
loaded = AutoForecast.load("forecaster.pts")

ModelPipeline — full control

from PipelineTS.pipeline import ModelPipeline, PipelineConfigs
from sklearn.preprocessing import StandardScaler

pipeline = ModelPipeline(
    time_col="date", target_col="value",
    lags=12, quantile=0.9,
    include_models=["d_linear", "catboost", "prophet"],   # or 'all', 'nn', 'ml', 'light'
)
leaderboard = pipeline.fit(data)
pred = pipeline.predict(n=12)

# Per-model configuration
configs = PipelineConfigs([
    ("d_linear", "d_linear_std",  {"pipeline_configs": {"lags": 24, "scaler": StandardScaler()}}),
    ("catboost", "catboost_diff", {"pipeline_configs": {"differential_n": 1}}),
])
pipeline = ModelPipeline(..., configs=configs)

SmartRouter — intelligent AutoML

from PipelineTS.pipeline import SmartRouter

router = SmartRouter(
    time_col="date", target_col="value",
    n_predict=12,
    preset="medium_quality",          # auto-configures models, CV, search, ensemble
    ensemble_strategy="auto",
    quantile=0.9,
)
router.fit(data)
pred = router.predict(n=12)

# Multi-quantile intervals
result = router.predict_quantiles(n=12, levels=[0.5, 0.8, 0.9])

# Panel data (multiple series)
router = SmartRouter(time_col="date", target_col="sales", id_col="store_id")
router.fit(panel_data)

# Save / load
router.save("router.pts")
router = SmartRouter.load("router.pts")

---

Features

Capability	Details
50+ built-in models	34 NN · ML · statistical · optional foundation models
Zero-friction API	forecast() auto-infers columns, cleans data, runs AutoML
SmartRouter AutoML	Profiles data → selects models, lags, scaler, hyperparams, ensemble
Conformal intervals	Distribution-free CQR for NNs + conformal prediction for all models
Multi-quantile	predict_quantiles(levels=[0.5, 0.8, 0.9]) in one call
Panel data	id_col for native multi-series with per-series scaling
Covariates	Known future & past covariates for GBDT, Prophet, AutoARIMA
Multivariate	ITransformer / SRSNet: multi-input → multi-output
Incremental learning	update(new_data) warm-starts NNs, refits ML/stat models
HPO	Built-in Optuna search inside SmartRouter (hpo_strategy='quick')
Preprocessing toolkit	Missing values, outlier detection, stationarity tests, frequency detection
Feature engineering	Fourier, calendar, holiday, lag rolling features
Evaluation	Walk-forward backtesting, residual analysis, multi-model comparison
Visualization	ACF/PACF, decomposition, forecast plot — with Chinese font auto-detection
Save / load (.pts)	Binary format with SHA-256 checksums and atomic writes

---

Models

Neural Network (34)

Key	Model
d_linear	DLinear — decomposition linear
n_linear	NLinear — simple linear mapping
n_beats	N-BEATS
n_hits	N-HiTS — hierarchical interpolation
tft	Temporal Fusion Transformer
transformer	Transformer encoder
tide	Time-series Dense Encoder
gau	Gated Attention Unit
stacking_rnn	RWKV linear RNN + gated residual
time2vec	Time2Vec + RWKV
patch_rnn	Patch-based RNN
tcn	Temporal Convolutional Network
itransformer	Inverted Transformer (multivariate)
srs_net	Selective Representation Space Net (multivariate)
deepar	DeepAR — RWKV encoder + Gaussian head
timexer / time_mixer	Modern modular TimeXer / TimeMixer variants
timesnet / msgnet	Frequency and multi-scale convolutional variants
pyraformer / informer / reformer	Efficient transformer-style variants
etsformer / autoformer / fedformer	Decomposition and frequency-aware variants
lightts / tsmixer / wpmixer	Lightweight mixer variants
patchtst / multi_patch_former	Patch transformer variants
nonstationary_transformer	Non-stationarity-aware transformer variant
timefilter / seg_rnn / tirex	Filtering, segmented RNN, and trend-residual variants

Modern NN variants share modular reusable layers and a central spec registry to avoid duplicated wrapper/backbone boilerplate.

Machine Learning (8)

Key	Model
torch_boosting_forest	GPU gradient boosting (MART/DART)
torch_bagging_forest	GPU bagging forest
deep_forest	gcForest cascade ensemble
wide_gbrt	Wide-table GBRT with 40+ lag features
multi_output_model	Multi-output sklearn regressor
multi_step_model	Multi-step sklearn regressor
regressor_chain	Chained regressors
gc_forest	Classic gcForest

Statistical (2+)

Key	Model
prophet	Custom Prophet-like with ridge regression
auto_arima	Auto ARIMA — native implementation, no statsmodels
theta / ets / stat_ensemble	Fast lightweight baselines

Foundation — zero-shot (6, optional)

pip install chronos-forecasting      # Chronos-2
pip install tirex-ts                 # TiRex foundation
pip install transformers==4.40.1     # Sundial / Time-MoE

Key	Model
chronos_2	Amazon Chronos-2 (120M, covariate support)
chronos_2_small	Chronos-2-Small (28M)
chronos_2_synth	Chronos-2-Synth
tirex_foundation	NX-AI TiRex zero-shot foundation model
sundial	THUML Sundial / Timer 3.0
time_moe	Time-MoE mixture-of-experts foundation model

---

Documentation

Full API reference, guides, and tutorials live in the docs/ directory and the tutorials/ folder.

Resource	Link
Quick Start	docs/quickstart.md
All Models	docs/models.md
Pipeline & PipelineConfigs	docs/pipeline.md
SmartRouter	docs/advanced.md
Preprocessing	docs/preprocessing.md
Feature Engineering	docs/feature_engineering.md
Evaluation & Metrics	docs/evaluation.md
Visualization	docs/visualization.md
Zero-Friction API	docs/api_reference.md#zero-friction-api
API Reference	docs/api_reference.md
Changelog	CHANGELOG.md

Tutorials (Jupyter notebooks):

#	Topic
00	Zero-Friction API — load_data, diagnose, forecast, AutoForecast, backtest
01	Quick Start Guide
02	All Models
03	Multivariate Prediction
04	Advanced Pipeline
05	Preprocessing & Data
06	Hyperparameter Tuning
07	Benchmarks
08	Visualization
09	Multi-Quantile Intervals
10	Multi-Series & Covariates
11	Incremental Learning
12	SmartRouter & Pipeline Deep Dive

---

License

This project is licensed under the Apache 2.0 License. See LICENSE for details.