every paper, one per strategy
Papers.
microcosm publishes one paper per strategy — the operator's or
referee's validation dossier. Each paper lives at
/{strategy}/paper, the stable citation
URL; this page is the index for CVs, grant reports, and academic
visitors.
Weight-aware imputation for policy microsimulation
Regime-gated, chained, weighted-bootstrap quantile-regression forests, benchmarked against unweighted/weighted QRF, OLS, linear quantile regression, and hot-deck matching. 21 pages.
read the paper →Gradient-descent survey calibration vs. classical calibration, at scale
Comparing modern gradient-descent calibration against raking, GREG, entropy balancing, and bounded chi-square. Plan and method scaffold only — no results yet.
read the plan →L0 regularization for subnational microsimulation calibration
Target-informed sparse sampling: keeps 57,240 of 337,704 candidate records with a 4.74% microcosm loss.
read the paper →The population-view harness, as a standalone referee
Energy distance, calibration-blind coverage, classifier AUC, and an uncapped tail block — proven inside imputation-paper; the standalone popdgp methods paper has not been started.
read the status →What running imputation and calibration together does
Realized landmines, repairability, the target-vs-survey Pareto frontier, effective-support collapse — four questions, zero experiments run yet.
read the plan →Microcosm Dynamics: an open, scored dynamic microsimulation model
The model's earnings, family and disability components each pass a test on held-out or external data. Four blind tests compare it with DYNASIM3, and none of them uses those tested components. We plan to connect the forward earnings law to the projections, apply the retirement earnings test, add behavioral claiming and run the full benefit rules in the Axiom rules engine.
read the paper →New populations.
Country model papers (tax and benefit rules by jurisdiction) live on policyengine.org research, not here — this index covers the microcosm data-construction stack and its applications.