// the take
mlforecast (Nixtla) turns time series forecasting into a tabular ML problem: it builds lag, rolling-window, and date features from long-format series and feeds them to any sklearn-style regressor (LightGBM, XGBoost, linear models, etc.). It's aimed at people forecasting many series at once who want gradient boosting or other ML models instead of classical stats methods, and who may need to scale past a single machine.
The feature engineering core (lags, expanding/rolling transforms, date features) is implemented efficiently rather than looped in pandas, which is the actual expensive part of this kind of forecasting at scale. Model-agnostic design means any estimator with fit/predict works without rewriting - you're not locked into Nixtla's own models. Distributed training is real, with separate backend modules for dask, ray, and spark rather than a thin wrapper claiming to scale. Conformal prediction is built in for interval estimates instead of users having to bolt on ad hoc quantile tricks.
Multi-step forecasts are generated recursively, feeding each prediction back in as a lag input, so errors compound over longer horizons - that's inherent to the approach and something new users won't realize until their 30-day-out forecasts degrade. Supporting three distributed backends (dask/ray/spark) triples the surface area that can break on version mismatches with the underlying ML libraries (e.g. xgboost-ray compatibility), and you're on the hook for keeping all three working if you touch that code path. There's no library-level support for per-series model selection or hyperparameter search across your series - the hyperopt guide shows a manual workflow, it's not something `MLForecast` does for you. It's also part of a larger Nixtla ecosystem (statsforecast, neuralforecast) with its own schema conventions (unique_id/ds/y), so there's a bit of lock-in to those conventions if you're mixing it with other tooling.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →