Machine Spirits: Speculation and Adaptation of LLM Agents in Asset Markets

Battle Count: 50

The paper is highly relevant to understanding how LLM agents behave in financial markets, which directly impacts quantitative trading strategies. Key findings about speculative bubble formation, strategic adaptation (fundamentalist to trend-following switching), and endogenous market instability are critical for traders deploying or competing against AI agents. The Heuristic Switching Model analogy and the finding that frontier models adapt to exploit less sophisticated agents have direct implications for algorithmic trading. However, the paper focuses more on market dynamics and stability rather than specific trading strategies or alpha generation. The findings about data leakage and informational asymmetry are relevant for understanding competitive advantages in AI-driven trading.

The experimental setup is relatively straightforward: a simple market clearing equation with 6 agents over 50 periods. The main complexity lies in: (1) managing 15 different LLM APIs and local model deployments, (2) designing appropriate prompts that mirror human experiment instructions, (3) implementing the sequential multi-agent interaction protocol, (4) computing various bubble and mispricing metrics, (5) conducting statistical tests (t-tests, Mann-Whitney U, Mincer-Zarnowitz regressions), and (6) running sufficient experimental repeats for robustness. The market model itself is simple (one equation), but the multi-LLM orchestration and analysis pipeline requires moderate engineering effort.

Open-source models use predetermined seeds for reproducibility. OpenAI API provides 'mostly deterministic outputs' with seed setting. However, Google API does not allow reproducibility through seed parameter. Temperature is set to 1 throughout. Robustness checks across temperature (0.3, 0.7, 1.0) and memory (0, 2, 4) parameters are provided. Full prompts and experimental instructions are included in appendices. The base experiment (Hommes et al. 2008) is well-documented. However, stochastic nature of LLM outputs and API changes over time limit full reproducibility.