Build context for metrics computation

Precomputes predictions, log_lik, and loo based on metric needs and retained predictions. Only computes requested context that the fitter supports.

Usage

build_metric_context(
  fit_result,
  fitter,
  data_bundle,
  metrics,
  seed = NULL,
  retain = character()
)

Arguments

fit_result

A bayesim_fit_result object from a successful fit

fitter

S7 Fitter object

data_bundle

A validated data bundle list

metrics

List of S7 Metric objects

retain

Retention specification. Requesting "predictions" computes predictions even when no metric needs them.

Value

A named list containing any of:

• predictions: Prediction results from the fitter

• log_lik: Pointwise log-likelihood matrix (S x N)

• loo: LOO-CV results

• loo_psis, loo_psis_ll, loo_epred: PSIS object, pointwise log-lik, and posterior-expectation predictions backing the LOO prediction metrics

Details

The function inspects the needs property of each metric to determine what context elements are required. It then checks if the fitter supports each capability and computes them. Any errors during computation result in NULL values for that context element.

Evaluation data: predictions and log_lik are computed on the TEST set when data_bundle$test is present, otherwise on the training set. Every built-in metric that consumes them (pred_*, elpd_test, r2_test) compares against the test response, so the predictions must be for the test rows. The LOO context is always built on the training set — leave-one-out is in-sample by construction. loo_epred is likewise a training-set matrix; when no metric needs "loo" (or the fitter lacks LOO support) it is built directly via predict_epred() rather than through the LOO context, so declaring needs = "epred" alone still delivers it. The PSIS machinery (loo_psis, loo_psis_ll) rides along with loo_epred: it is computed only when some metric declares "epred"; a run whose LOO metrics need only the elpd summary (needs = "loo" alone, e.g. elpd_loo_metric()) pays for the loo_fit() summary alone (#69).