Destination ranks

This model output records how highly HCT and two firm-level benchmarks rank a worker's observed next employer among sampled alternatives.

flowchart TB
    periods[(Relationship periods and<br/>observed destinations)]
    checkpoint[(Trained HCT checkpoint)]
    benchmarks[(Company-year MQ and<br/>prior-year hiring counts)]
    score[Sample alternatives and rank<br/>the observed destination]
    output[(Destination ranks)]

    periods --> score
    checkpoint --> score
    benchmarks --> score
    score --> output

    classDef source fill:#e8eef7,stroke:#52739e
    classDef process fill:#fff0cc,stroke:#a66b00
    classDef data fill:#eee5f5,stroke:#76528c
    class periods,checkpoint,benchmarks source
    class score process
    class output data

Grain: one sampled worker move
Location: /data/<coverage>/analyses/<mode>/destination_ranks/
Scorer: model/scoring/score_destination_ranks.py

Rows identify the worker, incumbent, observed destination, move date, and model split. They record incumbent and destination MQ, MQ gain, HCT ranks, candidate counts, destination-firm mean-MQ ranks, and prior-year hiring-frequency ranks. The scorer uses processed relationships to recover destination start dates when needed.

python -m scoring.score_destination_ranks \
  --config config/config_one_percent.yaml --mode training
python -m scoring.score_destination_ranks \
  --config config/config_one_percent.yaml --mode full_training

On Slurm, use score-destination-ranks. The job saves progress after source shards; use --resume after interruption and do not combine it with --overwrite.

The next-employer prediction analysis reads both modes without rescoring the model.