Identifying the Response to Pressure
The Global Capital Allocation Project has released a new paper “Geoeconomic Pressure”, which leverages recent advances in large language models (LLMs) to identify the areas of the global economy that are particularly vulnerable to geoecoomic pressure and examine how targeted entities respond. We apply the methodology to focus on the 2025 trade war and track firm responses.
A defining characteristic of the modern global system is that major powers increasingly use economic and financial power for geopolitical ends. Using our LLM-based approach, we can distinguish between future versus present geoeconomic pressure and negative versus positive impact on firms, and we also observe firm responses to pressure.
We use these advantages to track and analyze firm responses to the ongoing trade war. In Fig. 1, we decompose the share of American firms reporting effects from current or anticipated tariffs into firms which report a negative impact and firms which report a positive impact.
Figure 1: Self-reported effects of tariffs on American firms. Data from this brief is updated regularly. Visit the Geoeconomic Monitor to explore.
Leveraging Tools from AI
Measuring geoeconomic pressure is inherently challenging because many threats never materialize once the targeted party complies, leaving little visible evidence. Moreover, the nuanced phrasing often obscures the measure, making it more difficult for traditional NLP approaches to capture these subtle signals.
LLMs are more suited to this task as they can provide sophisticated analysis of temporal distinctions and discussions of context. However, that comes with a substantial computational requirement. We take a multi-step approach in measuring whether firms are affected by geoeconomic pressure, the nature of the pressure, and, finally, how they respond to this pressure (Fig. 2).
Figure 2: Multi-step LLM prompting procedure
A first-stage prompt is applied to the entire text corpora to identify the subsample of firms that are affected by geoeconomic pressure at all. A second-stage prompt, which requires significantly longer run-time, is only evaluated on the subsample identified by the first-stage prompt, so as to avoid running the more computationally heavy inference on hundreds of thousands of firm-level documents.
Figure 5: Responses to tariffs of American and non-American firms, 2025
Excerpted from the non-technical brief (PDF). Visit The Global Capital Allocation Project Research Brief site to read the full brief.
