Skip to main content
A ForecasterAdapter is the uniform interface a ConformalMethod talks to. It converts library-specific objects — Nixtla long-format DataFrames, for instance — into the numpy arrays a method’s array surface consumes, so the calibration math never sees framework types. Four adapters ship with the package:

​
Capabilities

Optional features are mixins. An adapter inherits only the ones it implements, a method declares what it needs through REQUIRED_CAPABILITIES, and the check runs where the forecaster is bound — at the top of calibrate / calibrate_cv. Two consequences worth knowing before you pick one:
  • CallableAdapter declares nothing, so it works with calibrate() and explicit arrays but not calibrate_cv(). It is the simplest way to wrap a model you already have, and the workhorse of the test suite. A quantile-capable variant lives in tests/helpers.py.
  • MLForecastAdapter has no quantile support, so it cannot back ConformalizedQuantileRegression. Calibration fails immediately rather than at predict time.
Capability errors surface at calibration, never in predict()
This is a deliberate contract: a fitted object is only ever built after the check passed, so the adapter it holds is known to support everything it will need.

​
Capabilities that depend on configuration

ForecasterAdapter.supports() defaults to an isinstance check, but an adapter may override it when a capability depends on runtime configuration rather than on the class. NeuralForecastAdapter does exactly this: it inspects the model’s loss, so a point-loss model fails ConformalizedQuantileRegression calibration while an MQLoss model of the same class succeeds. It also refuses the quantile and refit capabilities when the model has exogenous features. StatsForecastAdapter does the same when exog_features is set, and MLForecastAdapter, which has no quantile path, refuses refit when exog_features is set.

​
Lazy imports

StatsForecastAdapter, MLForecastAdapter and NeuralForecastAdapter are re-exported lazily from conformalforecast.adapters, so importing the package never pulls in statsforecast, mlforecast, neuralforecast or torch; the library is only imported when you first touch the adapter that needs it. Without the extra installed, that first access raises ImportError with the exact command:
CallableAdapter is always available; it is pure numpy.

​
Writing your own

Inherit ForecasterAdapter plus the capability mixins you implement, and implement every abstract method including the keyword-only X=None on predict/predict_batch/cross_validate and the aligned CV mode. Read adapters/callable.py first, it is the smallest complete example, and Exogenous features shows it wrapping a real model.

​
Next

  • StatsForecast — the most capable adapter; all seven methods work behind it.
  • MLForecast — a global gradient-boosted model, and the one capability gap.
  • NeuralForecast — horizon from the model, quantiles from the loss.
  • Embedding — skipping adapters entirely and using a method’s array surface inside your own library.