Reconstructing Large Scale Production Networks

Abstract

Firm-to-firm production networks matter for aggregate propagation, but they are rarely observed. This paper reconstructs national-scale, weighted firm-to-firm networks from two public objects: a sectoral input--output table and the distribution of firm sizes by sector. The algorithm first draws a binary buyer-seller backbone from a sector-aware gravity model and then assigns weights by a minimum-energy program. A Markov closure makes the reconstructed network primitive, so it has a unique stationary distribution. The weighting program keeps one-step firm balances and sectoral flows close to the data; the stationary money vector is then checked ex post and remains close in aggregate. For the United States we reconstruct a network with about 6.5 million firms and 340 million links in roughly four hours on a single workstation. We also reconstruct the networks of Japan, the United Kingdom, Australia, Finland, and Denmark. The Japanese reconstruction, built without any link data, reproduces the heavy-tailed degree regime documented in the country's observed production network. The reconstructed networks exhibit customer tails heavier than supplier tails, though the algorithm treats the two sides symmetrically. We also run computational experiments on the reconstructed networks to assess the systemic risk posed by the failure of individual firms. These experiments show that neither firm size nor degree nor sectoral position is a good proxy for the aggregate losses generated by a firm's failure. For such questions, there is no good substitute for the complete weighted buyer-seller network that we reconstruct. We release the reconstruction code, the generated networks, a Python library, and a graphical