This paper investigates the firm-level predictors of artificial intelligence (AI) innovation in Europe using a machine learning (ML) approach. Drawing on a matched dataset that combines patent information from OECD REGPAT with firm-level financial and structural indicators from ORBIS Intellectual Property, we model the probability that a firm filed at least one AI-related patent at the European Patent Office (EPO) in 2020. AI patenting is identified using the World Intellectual Property Organization (WIPO) taxonomy. At the same time, the predictor set includes sectoral dummies, innovation and technological capabilities, balance sheet indicators, and interaction terms, resulting in a high-dimensional feature space. We compare traditional econometric models (Probit) with ML classifiers, including LASSO, Elastic Net, and Random Forest. Results reveal that ML algorithms significantly outperform Probit in identifying AI innovators, with Random Forest achieving the highest sensitivity (0.815) and balanced accuracy (0.854). These improvements are particularly valuable given the small share of firms performing AI patenting in the sample (3.7%). Variable importance analysis highlights ICT specialisation, patent stock, firm size, and market concentration as key drivers of AI innovation. Moreover, several interaction terms—such as those involving size, sector, and innovation intensity—emerge as critical in improving classification performance. These findings suggest that heterogeneities and nonlinear factors should be considered when predicting AI patenting. From a policy perspective, our results highlight that ML could help improve the tailoring of innovation policies by identifying firms with transformative potential. Supporting digital infrastructure, facilitating the diffusion of innovation beyond ICT-intensive hubs and tailoring strategies to firm size could make Europe's AI transformation more inclusive. Overall, this study demonstrates the analytical power of ML in uncovering complex innovation dynamics and providing actionable insights for scholars and policymakers.

Read the document