Modeling and Analyzing Time-Dependent Network of Firms

Since 1888 the economic and legal data related to firms in the Swiss Confederation has been collected in the registry of commerce, a public archive administered by the government, which contains documentation about juridical entities conducting business. The aim is to record and publish legally relevant administrative events (e.g., the creation or the fusion of two companies, the names of partners, board members, and company directors having signatory authority) and ensure the protection of third parties. In particular, it promotes the security of trades by conferring certain legal effects on registered facts. Historically, this data has been recorded as natural language text, written in French, Italian, and German, in a printed format. Nowadays this data is recorded in digital format and enriched with metadata about the involved companies. However, the complete semantics still needs to be extracted from natural language. For example, when a firm has a new owner or partner, such information is available only in the raw text, and similarly, when a firm merges with another one, the buyer firm needs to be extracted from natural language. This information is essential to support many tasks in economic studies, in particular, the ones requiring networks (i.e., graphs) of direct firm-firm connections, such as subsidiary and branch relationships, as well as indirect firm-owner-firm links.

Furthermore, the challenging aspect of these networks is the time dimension: the existence of links is subject to firm lifecycle events, such as change of ownership, birth, or exit - events that are documented in the registry of commerce as well. In the context of NRP project n. 187353, CodeLounge at Software Institute at USI, Lugano, in collaboration with the Institute of Economics, developed an approach that is able to retrieve the single, time-variant relationships between firms, and between firms and people, that constitute the aforementioned networks.

In this thesis, we explored approaches to store performance-critical data and keep track of the graph-like nature of the networks of firms’ transformations made over time. This model supports both (1) longitudinal analysis, i.e., tracking one or multiple networks over time and space to see how they evolve and their sensitivity to external relevant events (i.e., change in tax rate), and (2) snapshots over time of all firms in Switzerland, aggregated geographically or on other dimensions (i.e., state, business purpose). To hold the relationships between different companies as a priority we store, model, and query the network of firms as edges connecting nodes. We define a graph structure typical of the No-SQL Graph databases as an alternative to the classical tabular SQL database. We analyzed the technologies currently available as Neo4j and RDF for graph data storage and GraphX and GraphFrames for large-scale graph data processing.

In the end, to validate the model we collaborated with experts in the domain of economics to obtain relevant questions on the data, that will serve as requirements. Using this data model we demonstrated that it is possible to efficiently answer questions from the experts and provide understandable results. For example, we were able to analyze how companies evolved over fifteen years by examining their survival through direct or indirect takeovers. We also described the distribution of firms across different addresses through the investigation of companies’ geographical locations.