> ML_LIBRARY // PMDARIMA_v1.0
pmdarima
Taylor G. Smith / alkaline-ml — Python's equivalent of R's auto.arima for automated seasonal ARIMA order selection.
time-series-forecastingv2.0.4MITqualified
Model Training
Accelerators:
CPU
Distributed Training:No
Model Inference
Inference Accelerators:
CPU
Deployment Targets:server
What It Does
- +Automated search over ARIMA(p,d,q) and SARIMA(P,D,Q) orders via Hyndman-Khandakar algorithm
- +Information criterion minimization (AIC, AICc, BIC, HQIC)
- +Unit root tests for stationarity and seasonal differencing (ADF, KPSS, CH, OCSB)
- +scikit-learn compatible estimator wrapper with fit/predict API
What It Does Not Do
- -Scale horizontally across Spark clusters (use StatsForecast for millions of series)
- -Train deep neural networks or transformers
- -Support multi-variate vector autoregression (VAR)
>Suitable Work Types
- Automating seasonal ARIMA parameter tuning for individual business KPIs
- Replacing legacy R auto.arima scripts with equivalent Python microservices
- Teaching and validating classical time series models in academic/financial settings
>Unsuitable Work Types
- Forecasting 500,000 retail SKUs overnight (pmdarima is too slow; use StatsForecast AutoARIMA)
- Real-time financial streaming execution
Data Residency Implications
Runs strictly in local memory. Zero network calls.
Security Considerations
MIT license with unrestricted commercial use.
Operational Profile & Known Limitations
Maturity:mature
Learning Curve:low
Ops Complexity:low
Cost Tier:free-oss
> Known Limitations:
- Stepwise optimization on long seasonal periods (e.g. m=365) is extremely slow and can freeze if bounds are not constrained.
Associated Incident Patterns (Incidentpedia)
Enforce safeguards and monitoring to guard against these documented real-world failure modes:
> Primary Evidence & Benchmark Citations
pmdarima Documentationofficial-docs • >=2.0.0, <=2.0.x
2026-09-25HIGH