2026.08.03(MON)16:20-17:50

Title:

Survey Information Granularity and LLM-Based Economic Behavior Simulation: Evidence from Household Consumption in CHFS Danyang Xie氏(香港科技大学)

Abstract:
This paper uses the Beijing sample of the China Household Finance Survey (CHFS) to evaluate large language models as tools for simulating economic behavior, using household consumption as a real decision setting. We focus on how survey information granularity and historical information affect simulation quality. Using DeepSeek V4 Pro, we construct household-level LLM simulation tasks and compare different input formats, including structured household summaries, fine-grained questionnaire responses, and their combination. We also benchmark the LLM results against traditional methods, including OLS, Random Forest, XGBoost, Ridge, and Elastic Net. We find that fine-grained questionnaire information improves not only household-level consumption fit, but also the fit to the overall distribution, tail behavior, and subgroup heterogeneity in the real data. Historical information mainly helps when current information is limited. Compared with traditional prediction methods, LLMs do not show a consistent advantage in point prediction, but perform better in distributional fit and heterogeneity preservation.

場　所 Venue：

東北大学 経済学研究棟(C16) 4F 大会議室 4F Meeting Room Graduate School of Economics and Management Building (C16)