numeraire.core.evaluators.OutOfSampleR2Evaluator#

class numeraire.core.evaluators.OutOfSampleR2Evaluator(benchmark: str = 'historical')[source]#

Bases: object

Out-of-sample R^2 of a forecast vs a benchmark, 1 - SSE_model / SSE_benchmark (percent).

Pooled across all origins and assets; positive => the model beats the benchmark OOS.

benchmark selects the yardstick:

  • "historical" (default) — the prevailing-mean benchmark carried in the output (Goyal-Welch 2008): the right metric for predictive-regression methods (e.g. 1/A’s dp).

  • "zero" — a zero forecast, SSE_benchmark = sum r^2. This is the Gu-Kelly-Xiu (2020) convention for the machine-learning cross-section (return predictability is measured against “no signal”, not against a fitted mean), and it materially changes the number.

__init__(benchmark: str = 'historical') → None[source]#

Methods

__init__([benchmark])

evaluate(oos_output)

Attributes

requires