StatsForecast

StatsForecast is an open-source Python library developed by Nixtla Inc. for large-scale time series forecasting. It is designed for data scientists and engineers who need to fit and predict millions of time series quickly and accurately in production environments. The library provides fast, numba-optimized implementations of classic statistical models including AutoARIMA, AutoETS, AutoCES, MSTL, and Theta, and claims to be 20x faster than pmdarima for ARIMA, 500x faster than Prophet, and 4x faster than statsmodels for ETS. It can fit 10 benchmark models on 1,000,000 series in under 5 minutes using parallel backends like Ray, Dask, and Spark. StatsForecast also supports probabilistic forecasting with confidence intervals, exogenous variables, static covariates, and anomaly detection, all accessible through a familiar sklearn-like .fit() and .predict() API.

The core workflow centers on the StatsForecast class, which accepts a list of models and a DataFrame with columns for unique_id, ds (timestamp), and y (value). After calling .fit(), users generate forecasts with .predict(h=horizon, level=[confidence_levels]). For example, AutoARIMA with season_length=12 on monthly data can be fit and predicted in three lines of code. The library includes automatic model selection (AutoARIMA, AutoETS, AutoCES, AutoTheta) and specialized models like MSTL for multiple seasonalities and AutoTBATS for complex seasonal patterns. All automatic models output both point forecasts and probabilistic forecasts with fitted values. Exogenous regressors are supported for AutoARIMA and AutoMFLES, and the library can detect anomalies by comparing in-sample prediction intervals against actual values.

StatsForecast competes directly with Prophet, Darts, and sktime, positioning itself as a faster drop-in replacement for Prophet (500x speedup claimed) and a more scalable alternative to pmdarima and statsmodels. Its benchmarks show it can handle 1,000,000 series in 30 minutes with Ray, and it is 1.5x faster than R's forecast package for ARIMA. While Darts and sktime offer broader model ecosystems (including deep learning), StatsForecast focuses exclusively on statistical models optimized for speed and scale. The library is part of the Nixtlaverse ecosystem, which includes tools like utilsforecast for plotting and evaluation. Its primary advantage is raw performance on large collections of univariate time series, making it suitable for production forecasting pipelines where speed and accuracy are critical.

However, StatsForecast has notable trade-offs. Its documentation and examples are less extensive than those of more established libraries like Prophet or statsmodels, which can make troubleshooting harder for new users. Some advanced features—such as distributed computing with Ray, Dask, or Spark—require additional setup and dependencies beyond a simple pip install. The library is narrowly focused on statistical models; it does not include deep learning or gradient boosting models like those in Darts or sktime. Additionally, while the core library is free and open-source under an Apache 2.0 license, commercial use of the broader Nixtla platform may require contacting the company for a tailored quote, which could be a barrier for enterprise teams needing support or advanced features.

Free and open-source under an Apache 2.0 license. Commercial use of the broader Nixtla platform requires contacting Nixtla Inc. for a tailored quote.