Skip to content

eb-adapters

eb-adapters provides a small, focused collection of adapter classes that wrap external forecasting and regression libraries so they can be used interchangeably inside the ElectricBarometer ecosystem.

The goal of this package is interface normalization, not modeling innovation. Each adapter exposes a consistent, scikit-learn-like API:

  • fit(X, y, sample_weight=None) → returns self
  • predict(X) → returns a one-dimensional numpy array

This allows ElectricBarometer evaluation, selection, and cost-aware comparison utilities to operate uniformly across native scikit-learn estimators and non-sklearn forecasting engines.

What this package does

  • Wraps external libraries (Prophet, statsmodels, CatBoost, LightGBM, etc.)
  • Normalizes their interfaces to a common .fit / .predict contract
  • Preserves initialization parameters so models are cloneable
  • Keeps optional dependencies truly optional

eb-adapters does not: - Define loss functions or evaluation metrics - Perform model selection or ranking - Implement forecasting logic itself

Those responsibilities live in other ElectricBarometer packages.

Adapter philosophy

Adapters are intentionally thin:

  • Minimal logic beyond input normalization
  • No hidden state outside the wrapped model
  • Clear, documented assumptions about how X and y are interpreted
  • Safe to import even when optional dependencies are not installed

If a model cannot reasonably conform to this contract, it likely does not belong in this package.

Available adapters

Current adapters include:

  • Prophet (univariate time-series forecasting)
  • statsmodels ARIMA / SARIMAX
  • CatBoost regressor
  • LightGBM regressor

Each adapter is documented in the API section and generated directly from NumPy-style docstrings in the source code.

Relationship to other ElectricBarometer packages

  • eb-metrics: defines cost-aware error and loss functions
  • eb-evaluation: performs model evaluation, comparison, and selection
  • eb-adapters: makes external models compatible with those workflows

This separation keeps responsibilities clean and composable.

API documentation

See the API section for detailed adapter reference material generated automatically from the source code.