Toward a Self Organizing Literature Review: An Application to the Financial Impacts of Physical Climate Risks

Abstract

Traditional literature reviews rely on selective reading and thematic categorization, a process that is both time-consuming and prone to subjective bias. As scientific fields expand and methodological fragmentation increases, researchers face a growing challenge: how to organize and compare a heterogeneous body of work without losing precision or reproducibility. This paper introduces the Self Organizing Literature Review (SOLR), a data-driven framework that encodes, clusters, and visualizes a corpus of empirical articles according to the models they contain, rather than the topics they discuss. The SOLR maps the structure of the field itself. The framework combines a systematic encoding protocol, in which each article is described by a set of standardized variables capturing its physical proxies, financial indicators, method, and end-user, with two complementary unsupervised learning techniques: self organizing map (SOM) (Kohonen, 1982) and variational autoencoders (VAE) (Kingma and Welling, 2013). Applied to a Burt matrix derived from a complete disjunctive table, the SOM produces a topologically consistent two-dimensional grid of methodological families. The VAE provides a continuous latent-space projection to reveal transitions between research communities that discrete clustering would obscure. We demonstrate the framework on a corpus of 150 peer-reviewed articles addressing the financial impacts of physical climate risks, a field characterized by precisely the kind of disciplinary fragmentation the SOLR is designed to address. The resulting maps reveal coherent clusters corresponding to actuarial catastrophe modeling, banking credit-risk frameworks, market-based valuation approaches, and hybrid models, while exposing geographic biases, methodological blind spots, and underexplored intersections between disciplines. The SOLR offers three properties that conventional reviews lack: reproducibility, since the encoding and clustering steps are algorithmic and can be audited; updatability, since new publications can be integrated into the map without repeating the full review; and comparability, since the position of each article in the latent space can reflect its methodological content. The approach requires low computational resources and is applicable to any interdisciplinary corpus where model diversity is a challenge.