horizon_set#

Back to L1 | Browse all axes | Browse all options

Axis horizon_set on sub-layer l1_f (layer l1).

Sub-layer#

l1_f

Axis metadata#

  • Default: 'derived'

  • Sweepable: False

  • Status: operational

Operational status summary#

  • Operational: 5 option(s)

  • Future: 0 option(s)

Options#

standard_md – operational#

Standard FRED-MD horizons: {1, 3, 6, 9, 12, 18, 24} months.

The canonical multi-horizon set used in the McCracken-Ng / Stock-Watson tradition for monthly forecasting. Models are fit per-horizon (when forecast_strategy = direct) and metrics report per-(model, horizon) rows.

When to use

Default for monthly studies. Comparable to published monthly benchmarks.

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.’

  • McCracken & Ng (2016) ‘FRED-MD: A Monthly Database for Macroeconomic Research’, Journal of Business & Economic Statistics 34(4). (doi:10.1080/07350015.2015.1086655)

Related options: standard_qd, single, custom_list, range_up_to_h

Last reviewed 2026-05-04 by macroforecast author.

standard_qd – operational#

Standard FRED-QD horizons: {1, 2, 4, 8} quarters.

Quarterly counterpart of standard_md.

Configures the horizon_set axis on l1_f (layer l1); the standard_qd value is materialised in the recipe’s fixed_axes block under that sub-layer.

When to use

Default for quarterly (FRED-QD) studies.

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.’

  • McCracken & Ng (2020) ‘FRED-QD: A Quarterly Database for Macroeconomic Research’, Federal Reserve Bank of St. Louis Review.

Related options: standard_md, single, custom_list

Last reviewed 2026-05-04 by macroforecast author.

single – operational#

A single horizon (defaults to h=1).

Forecasts only one horizon per cell. Sets leaf_config.target_horizons = [N] to override the default of 1. Faster than multi-horizon studies and clearer metrics tables when the study’s question is single-horizon.

When to use

One-shot studies (h=1 nowcasting, h=12 long-horizon ablation).

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: standard_md, standard_qd, custom_list

Parameters

name

type

default

constraint

description

target_horizons

list[int]

None

Optional when horizon_set=single. If provided, must be a list of exactly one positive integer. If omitted, runtime defaults to [1].

Single-element horizon list. The one value sets the forecasting horizon h for all models in the cell loop.

Last reviewed 2026-05-04 by macroforecast author.

custom_list – operational#

User-supplied horizon list (any non-empty integer set).

Requires leaf_config.target_horizons: [int...]. Useful for non-standard horizon comparisons (e.g., {1, 2, 3, 6, 12} or {6, 12, 24, 36}).

When to use

Replication of papers with non-standard horizon sets; ablation studies.

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: standard_md, standard_qd, range_up_to_h

Parameters

name

type

default

constraint

description

target_horizons

list[int]

—

Required when horizon_set=custom_list. Non-empty list of positive integers. Duplicate values are silently de-duplicated; order preserved.

Explicit list of forecasting horizons h. One model fit per horizon per cell when forecast_strategy=direct.

Last reviewed 2026-05-04 by macroforecast author.

range_up_to_h – operational#

Every horizon from 1 to leaf_config.max_horizon (inclusive).

Equivalent to custom_list with [1, 2, ..., max_horizon]. Useful for direct-h forecasting where the study wants dense horizon coverage (e.g., 1-12 months).

When to use

Dense horizon studies with direct-h forecasting.

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: custom_list, standard_md

Parameters

name

type

default

constraint

description

max_horizon

int

—

Required when horizon_set=range_up_to_h. Must be a positive integer >= 1. Produces horizons [1, 2, …, max_horizon].

Upper bound of the dense horizon range. The runtime expands this into tuple(range(1, max_horizon+1)) before building the cell loop.

Last reviewed 2026-05-04 by macroforecast author.