numeraire.baselines

Universal benchmarks bundled in numeraire — the reference rules every method is compared to.

Four estimators, all framework citizens (they pass numeraire.testing.check_estimator) and all registered via the numeraire.methods entry-point group (dogfooding the open discovery):

• EqualWeight — 1/N to_weights (the naive benchmark).

• MinVariance — global minimum-variance to_weights (sample covariance + window).

• MeanVariance — plug-in mean-variance to_weights (S^-1 mu, explicit norm).

• HistoricalMean — prevailing-historical-mean to_forecast (Goyal-Welch benchmark).

The pure weight functions (equal_weights(), minimum_variance_weights(), mean_variance_weights()) are the single source of truth for these formulae — method packages (e.g. the naive-diversification reproduction) build on them rather than re-deriving the algebra.

Serious constrained optimizers are not re-implemented here; they arrive through the optional skfolio adapter (numeraire.adapters). These baselines are the always-available floor.

EqualWeight	The 1/N benchmark estimator: rebalances to equal weights over view.assets each period.
MinVariance	Global minimum-variance estimator: sample covariance from the (optionally windowed) view.
MeanVariance	Plug-in mean-variance estimator: sample mu/S → S^-1 mu, normalization explicit.
HistoricalMean	The prevailing-historical-mean forecaster (Goyal-Welch benchmark).
equal_weights	The naive 1/N portfolio: w_i = 1/n (needs no moment estimation).
minimum_variance_weights	Global minimum-variance weights S^-1 1 / (1' S^-1 1) — the FOC of min w'Sw, 1'w=1.
mean_variance_weights	Plug-in mean-variance (tangency) weights ∝ S^-1 mu, normalization made explicit.