Nixtla    

Statistical ⚡️ Forecast

Lightning fast forecasting with statistical and econometric models

StatsForecast offers a collection of widely used univariate time series forecasting models, including automatic ARIMA and ETS modeling optimized for high performance using numba. It also includes a large battery of benchmarking models.

💻 Installation and Getting Started

pip install statsforecast

conda install -c conda-forge statsforecast

git clone https://github.com/Nixtla/statsforecast.git
cd statsforecast
pip install -e .

To get started just follow this guide.

🎉 New!

• ETS Example: 4x faster than StatsModels with improved accuracy and robustness.
• Complete pipeline and comparison: 20x faster than pmdarima and 500x faster than Prophet.

🔥 Highlights

• Fastest and most accurate auto_arima in Python and R.

• Fastest and most accurate ets in Python and R.

• New!: Replace FB-Prophet in two lines of code and gain speed and accuracy. Check the experiments here.

• New!: Distributed computation in clusters with ray. (Forecast 1M series in 30min)

• New!: Good Ol' sklearn syntax with AutoARIMA().fit(y).predict(h=7).

🎊 Features

• Inclusion of exogenous variables and prediction intervals for ARIMA.

• 20x faster than pmdarima.

• 1.5x faster than R.

• 500x faster than Prophet.

• 4x faster than statsmodels.

• Compiled to high performance machine code through numba.

• 1,000,000 series in 30 min with ray.

• Out of the box implementation of ses, adida, historic_average, croston_classic, croston_sba, croston_optimized, seasonal_window_average, seasonal_naive, imapa naive, random_walk_with_drift, window_average, seasonal_exponential_smoothing, tsb, auto_arima and ets.

Missing something? Please open an issue or write us in

📖 Why?

Current Python alternatives for statistical models are slow, inaccurate and don't scale well. So we created a library that can be used to forecast in production environments or as benchmarks. StatsForecast includes an extensive battery of models that can efficiently fit millions of time series.

🔬 Accuracy & ⏲ Speed

ARIMA

The auto_arima model implemented in StatsForecast is 20x faster than pmdarima and 1.5x faster than R while improving accuracy. You can see the exact comparison and reproduce the results here.

ETS

StatsForecast's exponential smoothing is 4x faster than StatsModels' and 1.6x faster than R's, with improved accuracy and robustness. You can see the exact comparison and reproduce the resultshere

Benchmarks at Scale

With StatsForecast you can fit 9 benchmark models on 1,000,000 series in under 5 min. Reproduce the results here.

🧬 Getting Started

You can run this notebooks to get you started.

• Example of different auto_arima models on M4 data

  • In this notebook we present Nixtla's auto_arima. The auto_arima model is widely used to forecast time series in production and as a benchmark. However, the alternative python implementation (pmdarima) is so slow that prevents data scientists from quickly iterating and deploying auto_arima in production for a large number of time series.

• Shorter Example of fitting and auto_arima and an ets model.

• Benchmarking 9 models on millions of series.

📖 Documentation (WIP)

Here is a link to the documentation.

🔨 How to contribute

See CONTRIBUTING.md.

📃 References

• The auto_arima model is based (translated) from the R implementation included in the forecast package developed by Rob Hyndman.
• The ets model is based (translated) from the R implementation included in the forecast package developed by Rob Hyndman.

Contributors ✨

Thanks goes to these wonderful people (emoji key):

fede 💻	José Morales 💻 🚧	Sugato Ray 💻	Jeff Tackes 🐛	darinkist 🤔	Alec Helyar 💬	Dave Hirschfeld 💬
mergenthaler 💻	Kin 💻	Yasslight90 🤔	asinig 🤔	Philip Gillißen 💻	Sebastian Hagn 🐛

This project follows the all-contributors specification. Contributions of any kind welcome!