The Forecast Combination Lab compares and combines completed forecast runs. It does not re-estimate the source models. The host first checks that candidate runs share the target series, horizon, vintage treatment, transform, and evaluation window.
Point Combination
The point-combination engine supports:
- Equal weights: each model receives weight 1/N.
- Bates-Granger: weights are based on inverse holdout mean squared error.
- Stacking: non-negative weights are estimated on validation folds and constrained to sum to one.
- Granger-Ramanathan: an OLS combination regression, with optional intercept behavior set by the method configuration.
When a member has a shorter forecast horizon, the combined forecast for later horizons uses the members that are still present and renormalizes their weights for that horizon.
Density Combination
Density combination accepts normal, quantile, or kernel-grid artifacts. Mixed formats must first be converted to a common representation. The engine evaluates CRPS, log score, negative log score, and PIT values on the aligned holdout window.
The current density path is explicit about approximation. If the source model gives only a point forecast and residual scale, the density is a normal residual approximation, not a model-native predictive distribution.
Fan Charts
Fan charts are built from density-combination output. Central bands are computed from density quantiles. Optional judgment adjustments can shift location, change variance, or tilt skew. Those adjustments are reported in the audit trail.
Tournament Evaluation
The tournament engine ranks forecasts on aligned holdout rows. It reports scorecards, pairwise tests, and model confidence set output. Diebold-Mariano tests are used for unconditional loss comparisons. Clark-West is used only for nested MSPE comparisons where that test is defined.
Giacomini-White conditional predictive ability is not reported as a shipped tournament result because it needs conditioning state variables.
Monitoring Snapshots
Monitoring computes rolling error metrics, deterioration ratios, and threshold alerts from completed forecast and actual paths. It is a snapshot workflow. Scheduled refresh, live feeds, and streaming alert delivery are separate runtime work.
Scenario-Conditional Combination
Scenario-conditional combination applies combination rules under named scenarios. It is useful for comparing how the combined path changes when model forecasts or weights are adjusted. It should not be read as a structural counterfactual unless the source forecasts came from a structural design.
Export Contract
Combination exports include method, source metadata, compatibility signature, validation/test split notes, output payload, and deterministic fingerprint. These fields are present so a reader can replay what was combined and see why incompatible source runs were excluded.