research paper · plan stage

Gradient-descent survey calibration vs. classical calibration, at scale

Modern policy microsimulation calibrates a candidate frame against thousands of hierarchical administrative targets by gradient descent on a flexible, capped loss. The survey-statistics standard remains classical calibration — raking, GREG, entropy balancing, bounded chi-square. No published comparison runs both families on a target surface at this scale. This paper is that comparison — the calibrate operator's validation dossier, symmetric to the fill operator's (imputation-paper).

status

Plan and method scaffold only. No experiments have run against real data; no results exist yet. Nothing on this page cites a paper finding.

methods compared

Raking (IPF), GREG, chi-square distance, entropy balancing / exponential tilting, bounded chi-square, and populace's gradient-descent calibrator — with and without the hard weight-ratio bound.

blocked on

Frozen target-surface machinery from sparsity-paper and the popdgp harness extraction, both tracked as open issues in their own repos.

target venue

Survey Methodology or Journal of Official Statistics (both diamond open access); arXiv preprint on completion.