imputation_temporal_rule
Back to L2 | Browse all axes | Browse all options
Axis
imputation_temporal_ruleon sub-layerl2_d(layerl2).
Sub-layer
l2_d
Axis metadata
Default:
'expanding_window_per_origin'Sweepable: True
Status: operational
Operational status summary
Operational: 3 option(s)
Future: 0 option(s)
Options
expanding_window_per_origin – operational
Re-fit the imputation model on every expanding window.
Default temporal_rule: at each OOS origin, the imputation model is fit on all data from the sample start through the origin date. Avoids look-ahead while ensuring the model has access to maximum data at each step.
When to use
Default; OOS-safe imputation. Selecting expanding_window_per_origin on l2.imputation_temporal_rule activates this branch of the layer’s runtime.
When NOT to use
When per-origin re-fits are too expensive – consider block_recompute.
References
macroforecast design Part 2, L2: ‘preprocessing is the only layer with a strict A→B→C→D→E execution order; every cell follows the same pipeline.’
Related options: rolling_window_per_origin, block_recompute
Last reviewed 2026-05-04 by macroforecast author.
rolling_window_per_origin – operational
Re-fit the imputation model on a fixed-length rolling window.
Fits the imputation model on the most-recent params.window observations only. Useful when the underlying covariance structure is non-stationary and old data should not influence current imputations.
When to use
Non-stationary panels where covariance drifts.
When NOT to use
When the panel is stationary – expanding window uses more information.
References
macroforecast design Part 2, L2: ‘preprocessing is the only layer with a strict A→B→C→D→E execution order; every cell follows the same pipeline.’
Related options: expanding_window_per_origin, block_recompute
Last reviewed 2026-05-04 by macroforecast author.
block_recompute – operational
Re-fit the imputation model every N origins.
Fits the imputation model once every leaf_config.imputation_recompute_interval origins; intermediate origins reuse the cached fit. Cheap approximation to expanding_window_per_origin.
When to use
Long sweeps where per-origin re-fits are computationally infeasible.
When NOT to use
When precise OOS-safe imputation is critical.
References
macroforecast design Part 2, L2: ‘preprocessing is the only layer with a strict A→B→C→D→E execution order; every cell follows the same pipeline.’
Related options: expanding_window_per_origin, rolling_window_per_origin
Last reviewed 2026-05-04 by macroforecast author.