# Novelty positioning — WP 27 "The Invisible Hand of Machines"

Prepared 2026-10-27 by the literature-review member. All citation keys refer to paper/references.bib; every item was verified (see docs/literature_ledger.md).

# 1. What the existing LLM-and-markets literature does

(i) Single-agent trading benchmarks and backtests. FinMem (yu2023finmem), TradingGPT (li2023tradinggpt), FinAgent (zhang2024finagent) and the InvestorBench family evaluate one LLM agent (sometimes a small team) against a historical price path. The market does not react to the agent; the object is the agent's P&L or Sharpe ratio. Lopez-Lira and Tang (lopezlira2023chatgpt) is the return-predictability analogue (ChatGPT headline scores predict next-day returns). These papers say nothing about equilibrium prices because the price series is exogenous, and Lopez-Lira, Tang and Zhu (lopezlira2025memorization) show that such backtests are contaminated by memorization of the historical period.

(ii) LLM-populated market simulators without causal identification. ASFM (gao2024asfm), StockAgent (zhang2024stockagent), TwinMarket (yang2025twinmarket, NeurIPS 2025), FCLAgent (hashimoto2025), StockSim (papadakis2025) and Lopez-Lira's open-source simulator (lopezlira2025trade) all place many LLM agents on a matching engine and report emergent bubbles, crashes, or reactions to macro shocks. They are descriptive: they show that LLM markets can produce certain patterns, usually for one population composition and a handful of seeds. Two of them already contain observations close to our mechanism — TwinMarket notes that agents driven by one model family are more homogeneous and links this to herd behaviour, and StockAgent reports systematic differences between GPT- and Gemini-driven agents — but neither randomizes the LLM share, controls the information set, or separates the fundamental value from the efficient price and the transaction price. Lopez-Lira (lopezlira2025trade) is the closest in spirit (persistent order book, heterogeneous roles, perturbation analysis "like partial-dependence plots"), yet the perturbations are over market conditions, not over the penetration of LLM agents, and there is no matched-seed design.

(iii) Experimental-finance replications with LLM subjects. Henning et al. with Camerer (henning2025) run Smith–Suchanek–Williams markets with LLM traders and find near-fundamental pricing and muted bubbles; Ouyang and Sui (ouyang2026) find disposition effects and extrapolation in an open call auction and show prompt interventions move bubble size; Saxena, Pangallo, Hommes, Caccioli and del Rio-Chanona (saxena2026) test 15 models and find everything from fundamental pricing to human-like bubbles, with mixed populations unstable; Struski et al. (struski2026) and Chupilkin (chupilkin2026) replicate Smith (1962) double auctions and show slower convergence and strong dependence on the trading institution. These are the right benchmark for behavioural fidelity (do LLMs behave like humans?) but they hold the population 100 % LLM (or 100 % human) and ask a different question from ours.

(iv) Homogeneity, herding and monoculture. The algorithmic-monoculture literature (kleinberg2021, bommasani2022, bommasani2021) and the output-diversity evidence (doshi2024, park2024brm, anderson2024) establish that shared models homogenize decisions. In LLM multi-agent systems, Cho et al. (cho2025) document peer conformity. The single paper that measures correlated trading from shared inputs is Song (song2026, ICLR 2026 workshop): overlap in retrieved documents raises pairwise action correlation among RAG trading agents from 0.04 to 0.41. That is a backtest on crypto data with no market feedback — the price does not respond to the agents — so it cannot speak to price informativeness or liquidity spirals. Hansen and Lee (hansen2025, Fed FEDS) re-create herding experiments with LLMs and find less herding than humans unless agents are told to maximize profit.

(v) AI collusion and RL agents in markets. calvano2020, klein2021, fish2024 (LLM pricing agents reach supracompetitive prices; prompt wording matters), agrawal2025 (LLM sellers in CDAs collude more with communication), colliard2026 (Q-learning market makers quote markups) and dou2025 (RL informed speculators in a Kyle market lower price informativeness through "homogenized learning biases"). Dou, Goldstein and Ji is the nearest theoretical neighbour: it also ties a common learning process to reduced price efficiency. Its mechanism is strategic (price-trigger punishment, over-pruning bias in Q-learning), not a shared-error structure in signals, and it has no LLMs.

(vi) Policy literature. Gensler and Bailey (gensler2020), the FSB (fsb2024), Danielsson and Uthemann (danielsson2025), Aldasoro et al. (aldasoro2025) and Meng and Chen (meng2026) all argue that common models and data sources will raise trading correlations and amplify stress. None provides a controlled experiment; meng2026 offers a reduced-form 13F estimate.

# 2. The gap this paper fills

1. Exogenous, randomized LLM penetration with matched seeds. We treat llm_share as a treatment assigned to trader slots by an independent random stream, holding every other stream (fundamentals, news, policy assignment, activation, tie-breaks) fixed across the pair. Differences in outcomes between φ = 0 and φ > 0 are therefore causal effects of LLM penetration on the same market history, estimated with ≥ 30 seeds and a power calculation for the confirmatory design. No existing LLM-market paper does this; the simulators in (ii) report single compositions and the experiments in (iii) compare 100 %-LLM to 100 %-human markets.

2. Three prices, not one. We separate the latent fundamental V_t (known to the simulator only), the public value V^pub_t (Kalman/HMM posterior on announced dividends and news) and the transaction price, and we compute Hasbrouck's efficient-price / pricing-error decomposition (hasbrouck1993), variance ratios (lo1988), Kyle lambda (kyle1985), forecasting-regression informativeness (bai2016) and information shares (hasbrouck1995) on simulated data where the truth is observed. Existing LLM simulators report returns, volatility and "bubbles" relative to a dividend-discount value but do not distinguish information-efficiency loss from transitory pressure.

3. A shared-model correlated-error mechanism, derived and then tested by intervention. The theory section nests a static Kyle/Hellwig market in which each trader's signal error has a model-specific common component η_g; the share φ_g using model g and the effective number of models M_eff = 1/Σφ_g² pin down price informativeness, order-flow correlation and pricing error in closed form, with feedback (trend demand) and leverage (margin) extensions giving amplification 1/(1−λγ) and 1/(1−λX/m). The correlated-signal microstructure literature (foster1996, back2000, goldstein2015, pasquariello2007) has the ingredients but never applies them to shared foundation models, and the monoculture literature (kleinberg2021, bommasani2022) has the homogenization idea without a market-clearing price. The simulator then intervenes on the mechanism directly: same vs. diverse prompts (H4), provider mix, communication cohorts (H5), memory, and frozen decision-library replay vs. online calls — so M_eff is manipulated, not merely observed. Song (song2026) manipulates information overlap but without price feedback; TwinMarket and StockAgent observe model-family homogeneity without manipulating it.

4. Fairness designs by token, USD and decision-opportunity budgets, with matched latency. Because LLM agents decide every 20 ticks and incur latency, naive comparisons confound "being an LLM" with "being slow and infrequent". We report realized LLM shares of agents, orders and volume, and run budget-matched treatments (equal tokens, equal dollars, equal decision opportunities) and matched-latency treatments. No prior LLM-market study controls for these; most do not even report the realized share of volume.

5. Leverage × homogeneity interaction and policy counterfactuals. H3 (bubble/crash frequency and liquidity spirals rise in φ × leverage) connects the LLM literature to brunnermeier2009, thurner2012 and kirilenko2017; H6 evaluates circuit breakers and position limits by crash depth and time-to-absorb, in the spirit of budish2015. The policy papers in (vi) call for exactly this evidence.

# 3. Where we must be honest

• The idea that shared models correlate trading is not ours. It is in gensler2020, fsb2024, danielsson2025, dou2025 (as "homogenized learning biases") and empirically in song2026. Our contribution is identification and mechanism, not the hypothesis. The introduction should say so.
• Model-family homogeneity inside LLM markets has already been observed (yang2025twinmarket, zhang2024stockagent, saxena2026). We turn an observation into a treatment.
• LLM-populated order-book simulators exist and are open source (lopezlira2025trade, papadakis2025, gao2024asfm). Our simulator is not the novelty; the design is. We should cite and, where feasible, reuse their validation metrics (vyetrenko2020).
• Behavioural fidelity is contested. henning2025 finds LLMs more fundamental-driven than humans; ouyang2026 and saxena2026 find bubbles; hansen2025 finds less herding. Our results on informativeness do not require LLMs to be human-like — they require only that errors be correlated within a model — but we must not claim external validity for human-plus-LLM markets without caveats.
• Prompt sensitivity and drift (fish2024, chen2024hdsr) threaten replicability. The frozen decision-library treatment and the mock provider are our answer; the limitation remains for online runs.
• Scale. TwinMarket runs thousands of agents; our N = 10,000-slot runs replace only a fraction with online LLM calls and use the decision library for the rest. We should report this clearly rather than imply fully online populations.
• The stylized model is a one-period Kyle variant with a fixed response intensity β. The CARA extension relaxes it, but the strategic effects in dou2025 and colliard2026 (learning to collude, markups) are outside our theory and we should not claim to nest them.