Global Search Regression (gsreg): A new automatic model selection technique for cross-section, time series and panel data regressions

GLUZMAN, PABLO; PANIGO, DEMIAN TUPAC

STATA JOURNAL

STATA PRESS

Lugar: Texas; Año: 2015 vol. 15 p. 325 - 349

1536-867X

This paper presents the main features of Global Search Regression (gsreg), a new automatic model selection technique (AMST) for time series, cross- section and panel data regressions. As other exhaustive search algorithms (e.g. vselect) gsreg avoids characteristic path-dependence traps of standard backward and forward looking approaches (like PCGETS or RETINA). However, gsreg is the first Stata code that: 1) guarantees optimality with out-of-sample selection criteria; 2) allows residual testing for each alternative; 3) retains vselect leaps- and-bound shortcuts for in-sample selection criteria; and 4) provides (depending on user specifications) a full-information dataset with outcome statistics for every alternative model.