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

---

Models

Neural Network (34)

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

Machine Learning (8)

Statistical (2+)

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

---

Documentation

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

Tutorials (Jupyter notebooks):

---

License

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