The unglamorous side of forecasting, now absurdly fast

What it does

StatsForecast collects classical univariate time series models—AutoARIMA, ETS, Theta, MSTL, and a battery of naive baselines—behind a familiar fit/predict interface. It handles probabilistic forecasts, prediction intervals, exogenous regressors, and anomaly detection, with built-in parallelization through Spark, Dask, and Ray. The library is explicitly positioned as both a production forecaster and a benchmarking tool.

The interesting bit

The value is in the unglamorous optimization: rather than chasing neural architectures, the project squeezes classical econometric models until they run 500× faster than Prophet and 20× faster than pmdarima, turning old academic workhorses into something you can deploy on a million series without leaving your laptop.

Key highlights

• Claims 500× speed over Prophet, 20× over pmdarima, 4× over statsmodels, and 1.5× over R, citing experiments hosted in the repo.
• Scales to one million series in under five minutes for benchmarks, or thirty minutes with Ray, per the project’s own tests.
• Supports probabilistic forecasting, confidence intervals, exogenous variables, and static covariates where the underlying model allows.
• Includes adapters intended to let you replace FB-Prophet in two lines of code.
• Offers distributed backends: Spark, Dask, and Ray.

Verdict

Data scientists and engineers who need fast, interpretable baselines or production-scale univariate forecasts should look here. Anyone hunting for multivariate or deep-learning architectures should keep moving.