Metraipolis › archive › dc_ea1e423b1476
Estimating Decision Rules from Large Language Model Agents: An Application to Bank Balance-Sheet Management (2026) · DOI 10.21203/rs.3.rs-11429290/v1
- Source
- Metraipolis archive
- Original
- https://doi.org/10.21203/rs.3.rs-11429290/v1
- License (as archived)
- CC-BY-4.0
- Archived
- 2026-10-21T02:51:34.816Z
- SHA-256 of the archived text
205ed5013c472943a0489431650d16285f13ac99afb5565e535ba2caf8541229- Stable URL of this page
https://metraipolis.com/documento/dc_ea1e423b1476
How to cite
Estimating Decision Rules from Large Language Model Agents: An Application to Bank Balance-Sheet Management (2026) · DOI 10.21203/rs.3.rs-11429290/v1 — Metraipolis archive, https://doi.org/10.21203/rs.3.rs-11429290/v1 (CC-BY-4.0); archived copy: https://metraipolis.com/documento/dc_ea1e423b1476 (Metraipolis archive dc_ea1e423b1476, sha256 205ed5013c472943a0489431650d16285f13ac99afb5565e535ba2caf8541229)
BibTeX
@misc{dcea1e423b1476,
title = {Estimating Decision Rules from Large Language Model Agents: An Application to Bank Balance-Sheet Management (2026) · DOI 10.21203/rs.3.rs-11429290/v1},
howpublished = {Metraipolis archive},
url = {https://doi.org/10.21203/rs.3.rs-11429290/v1},
note = {License CC-BY-4.0; archived copy https://metraipolis.com/documento/dc_ea1e423b1476; sha256 205ed5013c472943a0489431650d16285f13ac99afb5565e535ba2caf8541229},
year = {2026}
}Archived text
Reproduced under CC-BY-4.0. Attribution: Metraipolis archive, https://doi.org/10.21203/rs.3.rs-11429290/v1. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Articolo recente con licenza aperta: scheda dai metadati Crossref (posted-content). Titolo: Estimating Decision Rules from Large Language Model Agents: An Application to Bank Balance-Sheet Management Autori: Zlatin Sarastov Pubblicato: 2026-10-20 · editore: Springer Science and Business Media LLC DOI: https://doi.org/10.21203/rs.3.rs-11429290/v1 Licenza dell'articolo: CC-BY-4.0 Abstract (testo degli autori, CC-BY-4.0): Abstract This paper examines whether decisions generated by large language model (LLM) agents can be represented by simple equations in an agent-based model of bank balance-sheet management. In a monthly leader–follower model, a stylised, predominantly deposit-funded euro-area bank chooses deposit rates, lending margins and the composition of bond reinvestment, while depositors and borrowers respond to these decisions and to economic conditions. The initial balance sheet is calibrated to selected indicators from regulatory disclosures. We estimate equation-based approximations of LLM-generated decisions and evaluate them against simulations not used in estimation. Across ten out-of-sample bank-year simulations, the final specification reproduces annual return on equity (ROE) within 0.31 percentage points and meets all fit criteria in seven; none of the deposit-run scenarios meets all criteria. In eighteen simulations with a second LLM family, the ROE difference does not exceed 0.71 percentage points, but only six meet all criteria. Against a market-pricing rule, the original optimisation model, and regression and nearest-neighbour approximations estimated on the same data, the final specification meets all criteria in more years and tracks liquidity most closely in deposit-run years. Instructions without behavioural guidance show that withdrawal magnitudes and fixed-rate borrowers’ price sensitivity depend materially on the instructions supplied; with customer equations estimated under neutral instructions, a two-month withdrawal episode no longer becomes self-sustaining. The results demonstrate a procedure for approximating and evaluating LLM-generated decisions within a defined simulation environment; they do not establish that these responses represent the behaviour of bank managers or customers. JEL classification: C63, C72, G21, E43, G28 Fonte: https://api.crossref.org/works/10.21203%2Frs.3.rs-11429290%2Fv1 — metadati Crossref, consultati il 2026-10-21. Interrogazione: «large language model agents tool use».