Finding predictors of corruption from European firm level survey data: a random forest approach

Corruption remains a significant constraint for firms in Europe, despite ongoing institutional reforms. The main goal of this paper is to obtain a list of firm-level variables that can serve as predictors of corruption perception using a machine learning approach. Drawing on agency and institutional theory, we analyse firm-level data from European firms from the World Bank Enterprise Survey (WBES). We employ a Random Forest classifier, which is well-suited for high-dimensional, categorical survey data, capturing non-linear relationships and interactions often missed by traditional models. The model achieves strong predictive performance (ROC AUC = 0.755; Accuracy = 79%). Results show that the most important prediction factors of corruption perception include firm age, size, ownership concentration, legal form, external financial audits, bribery experiences, sector, country group (EU vs. WB), innovation activity, and informal sector competition. The findings support the design of risk-based audits and encourage reforms to reduce informality through streamlined registration processes. The study contributes methodologically by applying machine learning to the field of political economy and expands theoretical insights into firm-level institutional barriers. It is one of the first research to apply Random Forest to firm-level corruption perception in both EU and Western Balkans.