White Paper:
Why Companies Die

In this whitepaper we demonstrate how to enhance the predictability of corporate failure using web search and large language models (LLMs).

Using Cosimo, an agentic search system powered by NOSIBLE, we curated one of the largest cross-sectional datasets linking corporate failures and media coverage. This dataset contains more than 10,000 company failures worldwide. It includes bankruptcies, insolvencies, shutdowns, dissolutions, and other material events. For each failure we retrieve and analyze all media coverage in the period leading up to and immediately following the failure.

We then fine-tune large language models (LLMs) to read this coverage, reason about it, and predict the likelihood of failure 1 to 4 quarters out. LLMs operate on language, not financial features which are often subject to lags. We demonstrate that LLMs are able learn recurring patterns in how risk emerges and evolves over time. Our analysis shows that these language-derived signals can anticipate collapse prior to conventional methods.

Because the input is language, the same models can be pointed at any entity provided that they have media-like coverage. It does not matter whether that media-like coverage is in the public or private domain. In essence, LLMs are general-purpose risk sensors that can pay attention to everything, all the time. We argue that this novel approach offers a practical means of putting an end to the blind spots that financial institutions suffer from.