step,rule,n_before,n_after,n_removed,rationale
1,keep months <= last_origin-2 (2026-02) with GDP published at last origin,345,338,7,information-availability rule (GDP <= t-2)
2,"form direct h=1 training pairs (need y(s-2), target y(s+1) <= t-2)",338,335,3,lag/lead construction loses rows at sample edges
3,drop pairs whose target or regressor month is in 2020-03..2021-06,335,316,19,COVID outliers would dominate OLS; pre-specified