Data

Layer 1 answers: what source frame, target y, and predictor x universe does this study start from? Layer 0 fixes runtime policy. Layer 1 defines source data, forecast-time information, target y, and candidate predictor x columns.

Layer 1 contains canonical registry axes plus a small number of hidden compatibility/helper selectors. The primary Navigator tree shows the user-facing decisions; state_selection and sd_variable_selection are hidden lower selectors used by FRED-SD helper resolution.

Group	Axes	Focus
Data Source Mode / Frequency	3	custom_source_policy, dataset, frequency
Forecast-Time Information	3	information_set_type, release_lag_rule, contemporaneous_x_rule
Target (y) Definition	1	target_structure; target IDs, horizons, and dates live in leaf_config
Predictor (x) Definition	1	variable_universe; x column lists and mappings live in leaf_config
FRED-SD Predictor Scope	5	fred_sd_frequency_policy, fred_sd_state_group, fred_sd_variable_group, hidden state_selection, hidden sd_variable_selection
Raw Source Quality	2	raw_missing_policy, raw_outlier_policy before FRED transforms/T-codes
FRED Transform / Frame Availability	3	official_transform_policy, official_transform_scope, missing_availability

The first Layer 1 decision is custom_source_policy:

• official_only: choose a FRED Source Panel.

• custom_panel_only: provide a custom file path and analysis frequency; do not choose a FRED Source Panel.

• official_plus_custom: choose a FRED Source Panel and append custom columns.

Layer 1 does not choose model family, benchmark, researcher preprocessing, feature representation, or evaluation metrics. Those belong to Layer 2 and later.

Relation To Layer Contracts

Layer Contract Design (and architecture/) defines the current L0-L8 map. L1 fixes the data task: source, target, predictor universe, geography, sample window, horizons, and regimes. Layer 0 handles runtime policy plus derived manifest metadata.

Operational Status

Kept in Layer 1:

• data source mode, FRED source panel, analysis frequency;

• data revision / vintage regime, publication lag, same-period x rule;

• target y cardinality and predictor x universe;

• FRED-SD predictor scope;

• raw-source missing/outlier handling before official transforms;

• FRED transform scope and frame availability.

Not owned by Layer 1:

• Layer 2: raw-to-clean preprocessing.

• Layer 3: feature engineering and target construction.

• Layer 4: model fitting, forecasts, benchmarks, ensembles, and tuning.

• Layer 5+: evaluation, statistical tests, interpretation, and output.

Data sources

Built-in FRED sources have dedicated recipe-author notes:

• FRED-MD — monthly U.S. macro panel.

• FRED-QD — quarterly U.S. macro panel.

• FRED-SD — state-level mixed-frequency panel.

For column dictionaries see the FRED datasets section under for_researchers/.

For the consolidated Layer-1 data policies axis-by-axis, see ../data_policies.md.