The Metamorphosis of Company Competition into Directed Weighted Graphs

Abstract

This thesis investigates how company competition can be modeled and forecast as a directed weighted graph derived from SEC Form 10 K disclosures. Using annual filings from 2023 and 2024, it develops an end to end pipeline that extracts explicit competitor mentions, derives implicit competition signals from business, product, and market language, resolves company identities, and constructs a directed network whose edges encode competitive relationships and whose weights represent relative competition intensity. On this foundation, the thesis studies a strict one year ahead forecasting problem and evaluates whether graph neural networks can recover future rivalry structure from filing grounded evidence alone. The empirical analysis compares a plain directed GraphSAGE baseline, a weighted GraphSAGE model using graph derived rivalry features, and a weighted GraphSAGE model enhanced with semantic E5 representations. Across repeated temporal experiments, the weighted models consistently outperform the unweighted baseline, with the strongest results achieved by the weighted GraphSAGE model with E5. A complementary fusion analysis further shows that implicit evidence is most valuable when used alongside explicit disclosures, improving the ranking of likely competitors while preserving the central role of directly stated rivalry. Taken together, the results show that filing grounded graph learning offers a promising framework for forecasting future competition structure and for moving beyond binary link prediction toward richer weighted representations of rivalry, while also identifying entity resolution and preprocessing quality as the main remaining obstacles to further improvement.