Calibrating Behavioral Parameters with Large Language Models
Researchers from arXiv have developed a novel framework utilizing Large Language Models (LLMs) as calibrated instruments to measure behavioral parameters in economics, such as loss aversion, herding, and extrapolation. These parameters are crucial for asset pricing models but have historically been difficult to quantify reliably. The study analyzed four LLMs across 24,000 agent-scenario pairs, revealing that baseline LLM behavior exhibits a systematic rationality bias, showing attenuated loss aversion and weak herding compared to human benchmarks. However, through profile-based calibration, the researchers induced stable and theoretically coherent shifts, allowing the models to reach or exceed human benchmark magnitudes for biases like anchoring and disposition effects. To validate these findings externally, the calibrated parameters were embedded into an agent-based asset pricing model. The results demonstrated that calibrated extrapolation successfully generated short-horizon momentum and long-horizon reversal patterns consistent with empirical economic evidence. This work establishes explicit measurement ranges and calibration functions for eight canonical behavioral biases, bridging artificial intelligence and behavioral economics.
Wire timeline
Calibrating Behavioral Parameters with Large Language Models
Researchers from arXiv have developed a novel framework utilizing Large Language Models (LLMs) as calibrated instruments to measure behavioral parameters in economics, such as loss aversion, herding, and extrapolation. These parameters are crucial for asset pricing models but have historically been difficult to quantify reliably. The study analyzed four LLMs across 24,000 agent-scenario pairs, revealing that baseline LLM behavior exhibits a systematic rationality bias, showing attenuated loss aversion and weak herding compared to human benchmarks. However, through profile-based calibration, the researchers induced stable and theoretically coherent shifts, allowing the models to reach or exceed human benchmark magnitudes for biases like anchoring and disposition effects. To validate these findings externally, the calibrated parameters were embedded into an agent-based asset pricing model. The results demonstrated that calibrated extrapolation successfully generated short-horizon momentum and long-horizon reversal patterns consistent with empirical economic evidence. This work establishes explicit measurement ranges and calibration functions for eight canonical behavioral biases, bridging artificial intelligence and behavioral economics.
cs.AI updates on arXiv.org