What Makes an Economic Agent? Behavioral Fidelity in LLM Simulations

Simulated users are increasingly represented by detailed personas. But a detailed description does not necessarily contain the information needed to simulate a particular decision. We study this problem in household finance, where liquid savings, debt, and available credit are known to shape behavior. We construct 800 synthetic U.S. households calibrated to Federal Reserve surveys and ask two frontier LLMs to answer financial questions under different representations of the same household, from the full balance sheet to demographics alone and detailed public personas. We benchmark the simulations against New York Fed survey data, where households carrying credit card debt put about 25 percentage points more of a windfall toward debt repayment than households without such debt. When given the household's actual financial information, the models reproduce large differences across households. With demographics or a detailed persona alone, the gap falls below 3 percentage points. Adding just four financial variables (income, liquid assets, credit card debt, and available credit) restores the differences. By contrast, finance descriptions generated from the persona do not, because they often fail to reflect the household's actual financial situation. For user simulation, what matters is not simply how much a persona says, but whether it captures the circumstances that shape the behavior being simulated.