Drivers of supply chain resilience in the new energy vehicle industry: An explainable machine learning approach

Under the global dual-carbon goals, supply chain resilience has become a strategic priority for the new energy vehicle (NEV) industry. Using a 2016–2023 panel of 403 Chinese listed NEV firms, we construct an entropy-weighted supply chain resilience index and predict it with Random Forest (RF) and XGBoost, interpreted via SHAP, against a linear regression benchmark. The tree-based ensembles explain 66%–68% of out-of-sample resilience variation, far above the linear benchmark (56%), while models using only supply chain characteristics fail to beat the sample mean. The two ensembles attribute resilience differently: transaction costs rank as the strongest supply chain predictor in RF, whereas chip import dependence dominates in XGBoost, though both agree that supply chain characteristics are stably among the top predictor groups. All five core factors exhibit nonlinear threshold patterns (on the normalized [0, 1] scale): automation first suppresses and then promotes resilience beyond a robot-penetration threshold of 0.19; government subsidies are effective mainly within an intensity window of 0.2–0.5; chip import dependence turns sharply detrimental beyond 0.77; innovation efficiency supports resilience only within an optimal moderate range, its contribution turning negative beyond 0.33; and transaction costs become resilience-enhancing only beyond thresholds of 0.15 (RF) to 0.21 (XGBoost). Short-horizon forecasts are accurate, but rolling forecasts collapse under the 2023 structural break, indicating that resilience prediction is regime-conditional. The study introduces explainable machine learning into NEV supply chain resilience research, clarifies the nonlinear formation mechanism of resilience, and offers threshold-based policy implications for firms, governments, and industry associations, contributing to Sustainable Development Goal 9.