LLMs as Calibrated Instruments for Behavioral Finance Parameters
A recent study published on arXiv (2602.01022) presents a novel framework that views large language models as calibrated tools for measuring behavioral parameters in asset pricing. Analyzing four models alongside 24,000 agent-scenario pairs, researchers identified a consistent rationality bias in the baseline behavior of LLMs, including diminished loss aversion, minimal herding, and nearly nonexistent disposition effects when compared to human standards. By employing profile-based calibration, significant, stable, and theoretically sound adjustments were made to parameters like loss aversion and herding, often matching or surpassing benchmark levels. To ensure external validity, these calibrated parameters were integrated into an agent-based pricing model, revealing short-term momentum and long-term reversal patterns that align with empirical finance.
Key facts
- arXiv paper 2602.01022
- Four LLMs tested
- 24,000 agent-scenario pairs
- Baseline LLMs show attenuated loss aversion
- Baseline LLMs show weak herding
- Baseline LLMs show near-zero disposition effects
- Profile-based calibration shifts loss aversion, herding, extrapolation, anchoring
- Calibrated extrapolation produces short-horizon momentum and long-horizon reversal
Entities
Institutions
- arXiv