Decoding SME innovation: a machine-learning profile of practice bundles and the place of public funding

This paper profiles the firm-level characteristics associated with innovation in small and medium-sized enterprises (SMEs) using interpretable machine learning. Drawing on 26,270 Spanish SMEs from the 2018 Community Innovation Survey, we benchmark several classifiers with out-of-sample evaluation; the selected model reaches a test-set accuracy of 89.87% (95% CI: 89.03%–90.66%). The outcome is a single indicator of any product and/or business-process innovation, and all results refer to that combined outcome. Non-R&D innovation expenditure, in-house R&D and the purchase of new technology contribute most to predictive performance, and cooperation opens a further route in the decision rules. Public funding contributes less and is associated with innovation mainly when it coexists with these activities rather than on its own. A surrogate decision tree identifies several alternative profiles that distinguish innovators from noninnovators. The study contributes an out-of-sample validated predictive profile of the innovating SME, evidence that the discriminating signal lies in bundles of practices rather than in isolated inputs, and interpretable decision rules that yield testable hypotheses for confirmatory research. Because the design is cross-sectional, exploratory and predictive, the reported relationships are associations identified by the model, not causal effects.