Homo Silicus is Hyper-Rational: Why LLM Agents Fail to Replicate Attention-Driven Trading

Abstract

Can large language model (LLM) agents serve as proxies for human investors in behavioral finance experiments? I deploy 96 GPT-4-family agents in a staggered difference-in-differences design, exposing treated agents to exogenous attention shocks and viral social media signals about meme stocks while holding the fundamentals constant. Whereas human retail investors substantially increase net purchases of attention-grabbing stocks, LLM agents reduce buying propensity by 11.67 percentage points (SE = 4.92, p = .018), equivalent to preventing one purchase for every 8.6 treated agent-period observations. The direction and magnitude of this effect are robust across all specifications, although conservative small-sample variance corrections widen the confidence interval. Agents also display a reversed disposition effect, selling losers 3:1 relative to winners, directly contradicting the human pattern driven by loss aversion. To explain these divergences, I develop the Normative-Descriptive Divergence (NDD) model, formalizing how prescriptive training corpora ('avoid hype,' 'cut losses') generate systematic departures from the emotion-driven behavior observed in human investors. Even when explicitly prompted to exhibit FOMO, agents rationally retreat when attention shocks impose calculable costs, revealing a strict constraint hierarchy in which quantitative optimization overrides persona-level instructions. These findings establish boundary conditions for “Homo Silicus” research: LLMs approximate normative economic benchmarks but fail to replicate biases rooted in emotional processing.

Keywords: Investor attention, retail trading, behavioral finance, Generative AI, difference-indifferences, LLM Agents