This paper shows that the notion of relevant markets can be readily operationalized using machine learning tools, called ``embeddings''. Exploiting data on 14 million firm-to-firm transactions for Belgium for the year 2014, we identify which suppliers tend to co-supply the same business customers. Embeddings then allow to identify which firms are similar to each other in a latent space, leveraging only information on their co-occurrence in the production network. Crucially, we find that firms that are similar in the latent space defined by embeddings, also turn out to be similar along observable characteristics that are often central to market definitions, such as geography or product space. Within narrowly defined NACE 4-digit industries, we find that suppliers that are geographically distant from each other are less similar in the embeddings space, and they tend to co-occur less often supplying the same customers. Further, using measures of product similarity proposed in the literature we show a positive correlation between firms' embedding similarity and measures of product similarity. Finally, we provide a case study of the ready-mix concrete sector, and show how the embeddings algorithm can delineate relevant markets for individual firms within this sector.

About the speaker

Glenn Magerman is Associate Professor of Economics at ECARES, Université Libre de Bruxelles, a Research Affiliate at CEPR London/Paris, and a Research Fellow at VIVES, KU Leuven. He is also a Fulbright alumnus at Stanford University. He is Visiting Fellow at the Oxford Martin School, and academic consultant for the multiyear research network “Challenged to Monetary Policy” of the European Central Bank.

His research agenda focuses on production networks and global value chains. In particular, he studies which factors determine for productivity and growth, how economic shocks are transmitted from one agent to another, how networks contribute to aggregate outcomes such as growth, welfare, inequality and inflation, and how to develop policies in such networks.