Supply chain readiness and export performance in Africa: evidence from interpretable machine learning

Purpose This study examines whether a multidimensional supply-chain readiness system predicts export performance across African economies and whether the relationships persist when export performance is measured through aggregate value, extensive product breadth and intensive product depth rather than exports relative to GDP. Design/methodology/approach The analysis begins with a 54-country, 1985–2024 country-year grid while retaining original missing observations. Export intensity is retained as a benchmark. Additional outcomes comprise log merchandise export value, the number of exported HS-6 products and log export value per HS-6 product. Ridge, Random Forest, Extra Trees and Gradient Boosting are assessed with temporal, random, country-group and rolling validation. Findings Extra Trees remains the strongest benchmark model (R-squared = 0.651). Predictive performance is also high for log merchandise exports (R-squared = 0.926), exported HS-6 product breadth (R-squared = 0.891) and log exports per HS-6 product (R-squared = 0.824). The complete-case benchmark, estimated without numerical imputation, records R-squared = 0.737. Country demeaning lowers R-squared to 0.242 for export intensity and 0.283 for log merchandise exports, confirming that pooled machine learning captures substantial between-country structure. Within-country importance nevertheless retains industry employment for export intensity and manufacturing value added for merchandise export value as leading readiness signals. Originality/value The study combines interpretable machine learning with alternative export-performance margins, explicit missing-data sensitivity and a pooled-versus-within-country bridge. The results distinguish predictive cross-country structure from within-country change and provide a more cautious basis for connecting supply-chain readiness to export breadth and depth.