Nathan Delacrétaz

Climate disclosure mandates are expanding, while ESG teams remain too small to cover all holdings, monitor compliance, and conduct deep analysis simultaneously. Survey-based frameworks reach only willing participants, leaving non-participating companies, recent acquisitions, and competitors unassessed. We show that AI can extract ESG information directly from published annual reports, achieving 80% accuracy on a 36-criterion framework without input from the reporting entity. Tested on 30 Swiss real estate annual reports in three languages, compact AI models running on standard office computers perform within one percentage point of costly cloud services while keeping portfolio data on-premise. The key finding is not that one system dominates, but that different configurations fit different workflows: precision-oriented models reduce false alarms in portfolio screening, whereas recalloriented models better support compliance tasks where missed disclosures carry legal risk. Off-the-shelf models already deliver production-grade accuracy without customization or training investment. Since report quality explains more performance variation than model choice, we propose a tiered review process: automatic acceptance for straightforward reports, targeted verification for flagged cases, and full manual review for high-stakes or ambiguous documents. More broadly, AI-enabled extraction can turn unstructured disclosures into structured inputs for automated workflows, reducing repetitive analyst workload and bringing qualitative analysis closer to the rigor and scalability of quantitative methods.

Rural populations are particularly exposed to increasing weather variability, notably through agriculture. In this paper, we exploit longitudinal data for Turkish provinces from 2008 to 2018 together with precipitation records over more than 30 years to quantify how variability in a standardized precipitation index (SPI) affects out-migration as an adaptation mechanism. Doing so, we document the role of three potential causal channels: per capita income, agricultural output, and local conflicts. Our results show that negative SPI shocks (droughts) are associated with higher out-migration in rural provinces. A mediated-moderator approach further suggests that changes in per capita income account for more than one quarter of the direct effect of droughts on out-migration, whereas agricultural output is only relevant for provinces in the upper quartile of crop production. Finally, we find evidence that local conflict fatalities increase with drought and trigger out-migration, although this channel is distinct from the direct effect of SPI shocks on out-migration.