numeraire.core.engine

Walk-forward OOS engine. The most-reused, most-bug-prone, method-agnostic core.

The driver is deliberately small: for each (train, test) fold it fits the estimator on the train view and asks the fitted model for its capability output on the test view, then computes realized P&L from the original full view so the model never touches future returns. Output is one tidy container carrying the preprocessing/vintage provenance every result row needs (config_hash + data_vintage).

Naming convention: *Output classes (WeightsOutput, ForecastOutput, PricingOutput, PanelWeightsOutput) are the engine’s capability artifacts — the OOS panels an Evaluator consumes to produce result rows. *Result classes elsewhere (SimulationResult, SortResult, the stats *Result records) are the return values of one-shot computations. If it feeds an evaluator it is an Output; if it is a computed answer it is a Result.

backtest	Backtest estimator over view, dispatching to the right typed driver by capability.
backtest_weights	Run a walk-forward OOS backtest of a to_weights estimator over view.
backtest_panel	Walk-forward OOS backtest of a cross-sectional to_weights estimator over a ragged panel.
backtest_forecast	Walk-forward pseudo-OOS forecast (forecast-origin convention; GW2008 / 1-A / VoC).
backtest_pricing	Walk-forward OOS pricing of a to_pricing estimator: pooled predicted vs realized panels.
backtest_pricing_in_sample	In-sample pricing: one full-sample fit, expected returns over the whole view (in_sample).
config_hash	Stable short hash of a JSON-serializable config dict (preprocessing provenance).
WeightsOutput	OOS output for a to_weights method: realized weights aligned with realized returns.
PanelWeightsOutput	OOS output for a cross-sectional to_weights method over a ragged panel.
ForecastOutput	OOS output for a to_forecast method: per-origin forecast, realized return, benchmark.
PricingOutput	Output for a to_pricing method: predicted expected returns vs realized, on test assets.