Mafia Connections: Infiltration in Corporate Ownership

Learned representations have transformed the measurement of unstructured data in economics. We extend it to relational data, showing that Temporal Graph Networks encode economically meaningful behavior in dynamic corporate ownership networks. Using a high-resolution Italian ownership graph anchored on 5,700 judicially confiscated firms, we train a TGN that is never supervised on confiscation to produce time-varying firm embeddings, and we summarize their geometry with an Infiltration Proximity Index (IPI): a real-time measure of how densely a firm’s latent neighborhood is populated by firms whose confiscation is already legally known. We validate the index along three dimensions. It forecasts confiscation out of sample up to four years ahead, with a higher area under the ROC curve than the full set of firm-level financial variables at every horizon and for every classifier, and with far fewer missed confiscations at the cost of flagging more firms that are never confiscated; the geometry it summarizes places firms confiscated only later closer to firms already confiscated at the time of measurement; and it responds coherently to local changes in ownership. We then use the index to date firms’ transitions into a high-risk regime in a staggered difference-in-differences design. Around the dated transition, firms display sharp scale expansion, rising liabilities and receivables, cost reallocation, and persistent illiquidity, with only temporary profit gains, patterns consistent with firms operating as conduits for financial flows rather than as profit maximizers.