`target_structure`

Back to L1 | Browse all axes | Browse all options

Axis target_structure on sub-layer l1_b (layer l1).

Sub-layer

l1_b

Axis metadata

Default: 'single_target'
Sweepable: False
Status: operational

Operational status summary

Operational: 2 option(s)
Future: 0 option(s)

Options

`single_target` – operational

Forecast one target series at a time.

The recipe declares one target in L1 leaf_config. All downstream layers (feature DAG, model, evaluation) operate on that single series.

This is the dominant pattern for benchmark studies because most forecasting literature reports per-target metrics; multi-series studies typically compose multiple single-target runs in a sweep.

When to use

Default. Any standard forecasting benchmark.

References

macroforecast design Part 1, L1: ‘data definition is the recipe layer that pins source, target, geography, and horizon – everything downstream branches off these choices.’

Related options: multi_target

Examples

Forecast CPI inflation

1_data:
  leaf_config:
    target: CPIAUCSL

Last reviewed 2026-05-04 by macroforecast author.

`multi_target` – operational

Forecast multiple target series jointly within one cell.

The recipe declares targets: [a, b, c] in leaf_config. The L4 model is fit per-(target, horizon) tuple; the L5 metrics table carries one row per (model, target, horizon, origin).

Useful for vector-target methods (VAR, FAVAR, BVAR) and for studies that compute cross-target metrics (e.g., portfolio MSE).

When to use

VAR-style joint forecasting; cross-target evaluation; replicating papers that report joint metrics.

When NOT to use

Independent per-target studies – those are usually clearer as separate sweep cells.

References

macroforecast design Part 1, L1: ‘data definition is the recipe layer that pins source, target, geography, and horizon – everything downstream branches off these choices.’

Related options: single_target