Predictor de Ingresos e Informalidad

Problem

Public conversation about informal labor in Peru argues from anecdote. ENAHO microdata already exists, but almost nobody outside academia turns it into a number a policymaker or a curious visitor can actually use, or shows the tradeoff between a model that explains and one that only predicts.

Approach

A model tournament (OLS, Mincer, Lasso, RF, GB) plus an informality classifier — built to show the whole path, not just the model that won.

Nine income specifications run under the same 80/20 split and 5-fold cross-validation, from a literal reading of a classroom Mincer equation up to gradient boosting on log income, selected by cross-validated MAE rather than by peeking at the test set. The deployed model (E9, gradient boosting) beats the best interpretable specification by S/79 of MAE — a gap read as the contribution of nonlinearities and interactions a linear form can’t capture, not as proof the simple model is wrong. No specification’s R² clears 0.5 in soles, which the README frames against Mincer’s own textbook R² of 0.25-0.35 in log income — a different scale, cited so the number isn’t mistaken for a stronger fit than it is.

A second, separately trained classifier estimates informal-employment risk from the same microdata: gradient boosting reaches 0.96 PR-AUC against a 0.68 prevalence baseline, with an operating point chosen on out-of-fold training probabilities (never test) so that, of every 1,000 workers flagged, roughly 900 are genuinely informal. A structural ablation is reported alongside it — removing firm size alone drops PR-AUC to 0.957 — precisely so the score isn’t read as more than what it is: a targeting tool for programs, not a forecast of anyone’s future employment status.