Inputs¶
The model consumes the validated reusable employment products and the derived model relationship periods. It does not read vendor raw data directly.
| Model component | Inputs |
|---|---|
| Worker tower | Employment history before the current relationship, optional job-category and seniority history, and configured static worker features |
| Firm tower | Firm identity, configured firm categoricals, and optional company-description embedding |
| Relationship state | Tenure, seniority, job category, calendar year, and other configured relationship features |
Relationship state is appended to the worker-side representation. This permits match quality to vary by the current period while preserving an employment history that is strictly pre-match.
Observation window¶
The upstream processed-data build applies processed.observation_end_date.
Relationships that remain active at that cutoff are right-censored; only
complete censored annual exposure periods enter the model dataset. Observed
separations at or before the cutoff remain event rows.
The model-only processed.observation_start_date controls left truncation when
model relationship periods are built. Pre-start employment spells remain in the
processed data to preserve relationship start dates, tenure, and prior worker
history, but they do not become model rows. A relationship active on the start
date receives a synthetic boundary period through its next real anniversary, so
separations immediately after the boundary are retained.
Model-period artifact¶
data.build_model_datasets writes sharded relationship periods, embedded
history features, artifact metadata, and per-split firm counts. See the
employment build workflow for its
upstream inputs and the training workflow for the next step.