Prediction and Learning of Exporting Firms: A Study of Colombia

This paper applies machine learning techniques to predict which manufacturing firms in Colombia are likely to become exporters, using data from the Encuesta Anual Manufacturera (EAM) and Encuesta de Desarrollo e Innovación Tecnológica (EDIT) for the period 2015–2019. The objective is to estimate each firm’s “distance to export” through a probability score learned from the characteristics of existing exporters. Among the different algorithms tested, Logit with LASSO regularization delivers the best predictive performance, correctly identifying nearly three out of four actual exporters. Building on these predictions, the study introduces an exporting score, a probability measure that ranks firms by their proximity to the export margin. This score captures heterogeneity among non-exporters, anticipates entry and exit dynamics, and highlights sectoral and geographic clusters of latent export potential. In addition, the analysis shows that a set of firm level characteristics consistently emerge as the most relevant predictors across models: importer status, firm size, and combined spillovers, complemented by operational variables such as value added, inventories, and quality certification. The findings offer valuable insights for export promotion policies, enabling more targeted support for firms likely to enter international markets.