Determinants of Patent Applications Across European Regions: Evidence from Panel Estimators, K-Means Clustering and Predictive Validation

Europe measures the innovative performance of its regions largely by counting patents, and allocates public resources accordingly. This paper asks what regional patenting reflects, using Regional Innovation Scoreboard data for 245 European regions observed annually from 2016 to 2023, and applies three methods. A correlated random-effects specification decomposes each association into within- and between-region components. K-Means clustering, constructed without reference to the dependent variable, yields four exploratory innovation profiles within which coefficient equality is formally tested. Predictive validation compares seven algorithms under designs withholding random observations, entire regions, entire countries, and later years. Business research effort is the dominant correlate; science–industry collaboration is weaker and does not survive country effects; trademark activity is positive and does. Regional capacity proves remarkably immobile: for the four variables modelled here, differences between regions account for roughly 95 per cent of the variation, so short-run movements carry little information. No difference in the research–patenting association across the four profiles is detected, at a precision the paper reports, while levels of research differ sharply. The ranking of predictive methods reverses as the validation design tightens, flexible algorithms falling behind ordinary least squares once regions and countries are withheld.