Darts vs sktime

Verdict

Both are free, open-source Python libraries that give forecasting a consistent fit/predict-style interface across many underlying models, and both let a team compare and ensemble models rather than commit to one library's house approach. Darts's emphasis is breadth: classical statistical models (ARIMA, exponential smoothing, Theta), gradient-boosting models, deep-learning architectures (N-BEATS, Temporal Convolutional Networks, Transformers) and pretrained foundation models all sit behind the same API, plus built-in probabilistic forecasting, covariate support and anomaly detection.

sktime's emphasis is rigor and interface consistency with the broader Python ML ecosystem: it extends scikit-learn's familiar estimator conventions — pipelines, cross-validation — to forecasting, alongside time-series classification, regression and clustering in the same toolkit. Its particular strength is validation tooling: temporal train/test splitting, rolling-window cross-validation, and pipeline composition with detrending and deseasonalizing transformers, aimed at teams that want a leakage-free evaluation workflow as much as a good forecast.

Choose Darts if

• You want deep-learning and foundation-model forecasters available in the same library as classical and gradient-boosting methods, without switching tools.
• Probabilistic forecasting and covariate (external variable) support out of the box matter to your use case.
• You're benchmarking many model types against the same dataset and want a single consistent interface to do it.

Choose sktime if

• Your team already works in scikit-learn and wants forecasting to feel like an extension of that, not a new paradigm.
• Rigorous, leakage-free evaluation — rolling-window cross-validation, proper temporal splitting — is as important as the model itself.
• You want time-series classification, regression or clustering available alongside forecasting in one toolkit.

What they share

Both are Apache/BSD-licensed open source with no commercial tier, both are Python-native and integrate with pandas, and both are explicitly designed to let a team compare multiple model families rather than lock into one. Either can serve as the "benchmarking harness" a team uses before picking a final model.

The honest caveat

Darts' breadth means some of its included deep-learning models require real GPU time and ML expertise to tune well — a wide model menu doesn't remove that cost. sktime's scikit-learn-style rigor is valuable but means less first-class support for the newest deep-learning or foundation-model approaches compared to Darts; if state-of-the-art neural forecasting matters more than validation tooling, Darts (or Nixtla's ecosystem) is the more direct route.