Corporate distress models built on accounting ratios alone treat every firm-year as if it were drawn from the same economic environment, yet default clusters sharply in downturns and tightening cycles. This paper develops a stacked ensemble that combines a regularized logistic regression, a random forest, XGBoost and LightGBM through an out-of-fold logistic meta-learner, and feeds it twelve firm-level financial ratios together with eight macroeconomic indicators (GDP growth, inflation, the policy rate and its change, the credit spread, unemployment and its change, and exchange-rate depreciation). The model is estimated on a firm-year panel of 54,750 observations on 5,398 firms over 2003 to 2022 generated by a calibrated data-generating process that embeds the leverage thresholds, liquidity cliffs and rate-sensitivity interactions reported in the empirical literature, and it is evaluated on a strictly out-of-time test window (2017 to 2022) that contains an unseen recession year.
On the test window the ensemble with macroeconomic inputs reaches an AUC of 0.852 (95% bootstrap interval 0.836 to 0.867) against 0.842 for a ratios-only logistic benchmark, and raises the share of distressed firms captured in the riskiest decile from 55.0% to 57.5%. Adding macroeconomic indicators improves the ensemble's AUC by 0.007 (interval 0.0003 to 0.013) and its average precision from 0.318 to 0.342. Tree-based SHAP decomposition attributes about one third of predictive signal to macroeconomic variables, led by GDP growth and the change in unemployment. The gains are concentrated where linear models are misspecified: under a purely linear benchmark design the ensemble offers no advantage over logistic regression, although macroeconomic inputs still help. A notable cost is calibration in the recession year, where the macro-augmented ensemble overstates the realized distress rate (24.4% predicted against 13.8% observed). We argue that macro-augmented ensembles are best used for ranking and early warning, with explicit recalibration or scenario overlays when the macroeconomic state lies near the edge of the training history.