Default Profiles
Defaults are part of the public contract. They must be named, versioned, and written to the manifest.
Draft profile:
macroforecast-default-v1
Current MVP profile:
Choice |
Value |
|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
preprocessing |
dataset t-code transforms only, no extra preprocessing |
sample period |
explicit |
data vintage |
current vintage when omitted |
|
|
|
|
|
|
|
|
This default still writes legacy-compatible
feature_builder=target_lag_features; the compiler also records canonical
Layer 2 block/runtime provenance. New full recipes may use explicit feature
blocks directly.
The profile also keeps the current Layer 3 compatibility names:
model_family=ar means candidate forecast generator family ar, and
benchmark_family=zero_change means the zero_change generator is assigned
the benchmark/baseline role.
Frequency resolution:
fred_mdmeans monthly.fred_qdmeans quarterly.fred_sdalone requiresfrequency.fred_md+fred_sdmeans monthly.fred_qd+fred_sdmeans quarterly.monthly data converted to quarterly uses a 3-month average with a warning.
quarterly data converted to monthly uses linear interpolation with a warning.
The MVP profile is intentionally narrow. It should expand only after the Experiment API and layer audit are stable.
The final profile should also specify:
default information set
default sample split
default benchmark
default model
default preprocessing
default metric
default statistical test behavior
default importance behavior
default artifact policy
default reproducibility mode
Design requirements:
no hidden defaults in artifacts
old default profiles remain reproducible
changing a default profile is a versioned change
users can inspect and override defaults