Twiga Forecast#
Twiga Forecast
Point & probabilistic time series forecasting -
one interface, every model.
pip install twiga
Most forecasting libraries force a choice between ML and deep learning, between point forecasts and probabilistic ones, between training from scratch and leveraging pre-trained models, or between flexibility and production readiness. Twiga removes those trade-offs.
One interface, every model
A single TwigaForecaster trains, tunes,
evaluates and backtests any model through an identical API. From
parameter-free baselines and pre-trained zero-shot models to gradient-boosted
trees, neural networks and weighted ensembles, switching architectures means
changing one config object.
Zero-shot forecasting, no training required
Chronos2, pre-trained on 2M+ diverse time series, generates 21-quantile probabilistic forecasts with no domain data, no fitting step and no calibration. Use it as a fast baseline, on small datasets, or blended in an ensemble with learned models.
Explore foundational modelsComposable probabilistic architecture
Any backbone pairs with any distribution head without modification, giving a full matrix of probabilistic models from a small set of reusable building blocks. Adding a new architecture or head is a focused, isolated change.
Explore probabilisticCalibrated uncertainty on any model
Conformal prediction wraps any trained model, including plain point-forecast models like LightGBM, with finite-sample intervals that carry a formal coverage guarantee. One call after training, no retraining and no distributional assumptions required.
Explore conformal predictionStructured experiments and ablations
A single ExperimentEngine.run(spec) call
runs backbone HPO once per dataset, then evaluates every condition across all
folds with automatic MLflow tracking and result aggregation. Pydantic configs,
Optuna integration, time-based backtesting and checkpoint-based resume are
first-class throughout.
Statistical analysis and feature intelligence
Eight feature-target association measures are computed
in a single call and fused with Borda-count rank aggregation and random-forest
importance to produce a robust ranked feature list. Stationarity tests, entropy
measures and residual diagnostics are included, all returning tidy DataFrames.
SHAP attribution is available for any fitted model via the optional
twiga[explain] extra.
Integrated visualization toolkit
A Lets-Plot suite covers the full workflow: forecast
vs. actual plots, quantile fan charts, reliability diagrams, residual
diagnostics and experiment result heatmaps. All functions return composable
ggplot objects with no Matplotlib boilerplate.
Production MLOps out of the box
An optional five-module stack covers experiment tracking, versioned checkpoints, drift monitoring, a production REST API and automated retraining as a schedulable workflow. All modules install together and operate without external infrastructure.
Explore MLOpsContributors
Thanks to the people who built Twiga.
