Semantic analysis in actuarial risk assessment: A framework for portfolio construction and concentration risk measurement

Traditional actuarial risk assessment relies primarily on historical return correlations to evaluate portfolio concentration and tail risk. However, corporate textual disclosures contain forward-looking information about business models, risk exposures, and strategic similarities that may not be fully captured in historical price relationships. This paper develops a systematic framework for incorporating semantic analysis of corporate disclosures into actuarial risk measurement, demonstrating applications to portfolio construction, concentration risk assessment, and stress testing. Using JSE All Share constituents, we apply natural language processing techniques to extract semantic similarity matrices from corporate narratives across business model, geopolitical, and ESG dimensions. Through shrinkage estimation, we combine these semantic correlations with traditional methods to produce enriched risk measures. Our analysis reveals that semantic correlations identify concentration risks not visible in historical data, with Effective Number of Bets reductions of 75%. The framework provides practical tools for incorporating qualitative corporate information into quantitative risk assessment processes.