% paper/sections/literature.tex — UQO WP 27, "The Invisible Hand of Machines"
% Section: Related literature. Every citation key is in paper/references.bib and every claim is
% backed by the corresponding row of docs/literature_ledger.md.
\section{Related literature}
\label{sec:literature}

The paper sits at the intersection of six literatures: agent-based and heterogeneous-agent models of
financial markets; the microstructure of double auctions and the measurement of price discovery;
information aggregation with correlated signals, herding and feedback trading; liquidity spirals,
fire sales and the measurement of contagion; algorithmic trading, flash crashes, algorithmic
monoculture and algorithmic collusion; and, closest to our question, the rapidly growing work that
places large language model (LLM) agents in economic experiments and market simulations. We review
each in turn and close with a statement of what the present paper adds and where it overlaps with
existing work.

\subsection{Agent-based and heterogeneous-agent models of financial markets}
\label{sec:literature:abm}

Our baseline population of rule-based traders descends from agent-based computational finance, which
showed that interacting heterogeneous rules reproduce the fat tails, volatility clustering and bubbles
of real markets \citep{lebaron2006}. In the Santa Fe artificial stock market the speed of belief
updating governs whether prices track fundamentals or display excess volatility and technical-trading
regimes \citep{lebaron1999}; in adaptive belief systems, discrete-choice switching between predictors
produces increasingly complex price dynamics as the intensity of choice rises \citep{brock1997}, and
the share of trend followers determines whether the fundamental price is stable \citep{brock1998}.
Interacting noise traders who switch between optimistic, pessimistic and fundamentalist states generate
scaling laws and clustered volatility endogenously \citep{lux1999,lux2000}; herd behaviour among
speculators yields bubbles and crashes in a mean-field model of contagious optimism \citep{lux1995}; and
recruitment-type opinion dynamics generate persistent herding states without any informational reason
\citep{kirman1993}. These mechanisms matter for us because they describe what a \emph{classical}
population does on its own, before any LLM is introduced, and because trend following and herding are
precisely the channels through which a correlated LLM order flow can be amplified.

Closer to our exchange design, \citet{chiarella2002} simulate a double auction populated by
fundamentalist, chartist and noise traders and reproduce order-book regularities, and
\citet{chiarella2009} show that heterogeneous rules in a limit-order book generate fat tails,
volatility clustering and order-flow autocorrelation, with the chartist share driving excess
volatility. Leverage in an agent-based model produces fat tails and clustered volatility through
synchronized margin-driven selling \citep{thurner2012}, which is the closest antecedent of our leverage
treatment. Multi-agent reinforcement-learning markets reproduce the same stylized facts
\citep{lussange2021} and are a precursor of LLM-populated markets. The surveys of \citet{hommes2006}
and \citet{hommes2021} summarize the laboratory and agent-based evidence that expectation heterogeneity
drives aggregate instability; we follow their lead in treating an LLM policy as one more expectation
type. \citet{farmer2009} argue that agent-based models are the appropriate tool for out-of-equilibrium
dynamics and for policy counterfactuals that cannot be solved analytically; \citet{axtell2025}
identify validation, calibration and scale as the open frontiers, which our matched-seed design and
power analysis address directly. For calibration and validation we use the moment set of structural
stochastic volatility models \citep{franke2012} and the catalogue of stylized facts: heavy tails
\citep{mandelbrot1963,cont2001}, conditional heteroskedasticity \citep{engle1982,bollerslev1986}, the
long memory of absolute returns \citep{ding1993}, and the long memory of order-flow signs coexisting
with nearly efficient prices \citep{lillo2004}.

\subsection{Double auctions, zero intelligence and the measurement of price discovery}
\label{sec:literature:lob}

The continuous double auction is the institution we study, and its experimental pedigree is long:
\citet{smith1962} showed that double auctions with induced values converge to competitive equilibrium
with few traders, and \citet{smith1988} showed that experimental asset markets with known dividend
streams nevertheless produce bubbles and crashes; our dividend-paying asset and bubble metrics follow
that design. The zero-intelligence tradition established that market structure, not trader
rationality, drives allocative efficiency \citep{gode1993}, a point reviewed by \citet{ladley2012}. A
zero-intelligence model of the continuous double auction predicts spread and price impact from
order-flow rates alone \citep{farmer2005}, and the statistical theory of \citet{smith2003} ties
spread, depth and price diffusion to the same parameters. These results fix the benchmark against
which any effect of ``intelligence'' must be measured: a change in the composition of a population can
move liquidity merely by changing order-flow rates, which is why we report realized order and volume
shares alongside the assigned treatment. On the empirical side we rely on the order-book facts of
\citet{bouchaud2002}, the Markovian book model of \citet{cont2010}, the linear relation between
order-flow imbalance and short-horizon price changes \citep{cont2014}, the survey of
\citet{gould2013}, and the treatment of impact and the slow digestion of order flow in
\citet{bouchaud2018}. Our event-driven exchange, with per-agent latency and a price--time matching
engine, follows the design of ABIDES \citep{byrd2020}, and our realism checklist is that of
\citet{vyetrenko2020}.

The outcome variables are standard measures of price discovery and market quality. The price-impact
coefficient $\lambda=\Cov(v,Y)/\Var(Y)$ of \citet{kyle1985} and the adverse-selection spread of
\citet{glosten1985} tie liquidity to the information content of order flow. \citet{hasbrouck1991}
measures that content by the permanent price impact in a VAR of quote revisions and trades;
\citet{hasbrouck1993} decomposes the transaction price into an efficient random-walk component and a
pricing error whose standard deviation is a market-quality metric; \citet{hasbrouck1995} attributes
price discovery across venues through information shares, which we adapt to LLM versus rule-based
order flow. Variance ratios reject the random walk under autocorrelation \citep{lo1988}, the effective
spread can be inferred from serial covariance \citep{roll1984}, and illiquidity can be measured as
absolute return per unit of volume \citep{amihud2002}. \citet{bai2016} define informativeness by how
well prices forecast future fundamentals, and \citet{bond2012} argue that informativeness is the
welfare-relevant object because real decisions feed on prices. \citet{hendershott2014} separate
transitory price pressure from permanent information effects, which justifies our separation of the
pricing error from the efficient-price variance. Because the simulator observes the latent value, we
can compute these statistics against the truth rather than against a proxy, which is the main
advantage of a simulated economy for this question \citep{fama1970}.

\subsection{Information aggregation with correlated signals, herding and feedback}
\label{sec:literature:info}

The theoretical core of the paper is the aggregation of \emph{correlated} private signals.
\citet{hellwig1980} shows that competitive prices aggregate dispersed private signals imperfectly
because of noise supply, and \citet{grossman1980} that prices cannot be fully informationally
efficient when information is costly. With strategic informed traders, \citet{foster1996} show that
the correlation structure of signals, not only their precision, governs trading intensity and price
informativeness, and \citet{back2000} that more correlated information leads to faster revelation.
\citet{goldstein2015} show that information diversity across traders is complementary, so that a loss
of diversity reduces informativeness about some components of value; \citet{dugast2018} show that
abundant cheap data can lower informativeness by crowding out costly deep information, a close parallel
to shared-model crowding; and \citet{kondor2012} warn that public information can raise disagreement
through higher-order beliefs, so that shared inputs do not trivially homogenize behaviour.
\citet{pasquariello2007} ties information heterogeneity among strategic speculators to contagion and
shows that homogeneity reduces it. Our stylized model (Section~\ref{sec:theory}) borrows the
correlated-error structure of these papers and reinterprets the common component as the signature of a
shared foundation model or prompt family.

Herding supplies the behavioural counterpart. Sequential decision-makers who follow their predecessors
destroy information aggregation \citep{banerjee1992}, informational cascades become more likely as
signals become more correlated \citep{bikhchandani1992}, reputational concerns make managers mimic one
another \citep{scharfstein1990}, and short-horizon speculators herd on the same information source
even when it is unrelated to fundamentals \citep{froot1992}. The buy-side imbalance statistic of
\citet{lakonishok1992} is the standard empirical measure of institutional herding, and we compute it
for LLM cohorts. Feedback trading closes the loop: the share of trend followers determines the
stability of the fundamental price in heterogeneous-belief models \citep{brock1998}, and
contagious optimism among speculators produces bubbles and crashes \citep{lux1995}. Our feedback
extension (Proposition~\ref{prop:feedback}) is a mechanical version of this amplification.

\subsection{Liquidity spirals, fire sales, leverage and contagion measurement}
\label{sec:literature:spirals}

The leverage treatment and the H3 and H6 hypotheses draw on the literature on funding constraints.
\citet{brunnermeier2009} show that market liquidity and funding liquidity reinforce each other through
margins, producing liquidity spirals; our margin-call mechanism follows this logic. Liquidation values
fall when potential buyers are themselves constrained \citep{shleifer1992}, constrained arbitrageurs can
amplify rather than correct mispricing \citep{shleifer1997}, forced sales push prices below
fundamentals and propagate losses \citep{shleifer2011}, and financially constrained arbitrageurs may
under-provide liquidity \citep{gromb2002}. Endogenous margins amplify booms and crashes in the leverage
cycle, with leverage caps as a policy lever \citep{geanakoplos2010}, and procyclical intermediary
leverage amplifies price moves \citep{adrian2010}. In the agent-based setting, \citet{thurner2012}
obtain fat tails and clustered volatility from synchronized margin-driven selling. Across assets,
losses of convergence traders propagate through reduced risk-bearing capacity \citep{kyle2001},
distressed selling induces endogenous correlation \citep{cont2013}, fire-sale spillovers scale with
common exposures and leverage \citep{greenwood2015}, cross-market rebalancing transmits shocks between
assets with unrelated fundamentals \citep{kodres2002}, and dense networks are robust to small shocks
but fragile to large ones \citep{acemoglu2015}. For measurement we use the spillover index of
\citet{diebold2009}, its generalized directional form \citep{diebold2012} and its network
interpretation \citep{diebold2014}, and we heed the warning of \citet{forbes2002} that correlation-based
contagion tests must adjust for heteroskedasticity.

\subsection{Algorithmic trading, flash crashes, monoculture and algorithmic collusion}
\label{sec:literature:algo}

The evidence on algorithmic and high-frequency trading is, on average, favourable to liquidity and
price discovery: algorithmic trading improves liquidity \citep{hendershott2011}, high-frequency traders
trade in the direction of permanent price changes and against transitory pricing errors
\citep{brogaard2014}, and algorithmic strategies in foreign exchange improve price efficiency even
though they are more correlated with each other than human trades \citep{chaboud2014}. That last
finding is the one we take most seriously: correlation among algorithms is a documented fact, and its
consequences depend on what the algorithms are correlated \emph{on}. High-frequency market makers
behave as new market makers with tight inventory control \citep{menkveld2013}, a feature we build into
the market-maker policy; \citet{menkveld2016} reviews the speed, adverse-selection and liquidity
consequences, and \citet{foucault2016} show how fast traders reacting to common news signals affect
informativeness and liquidity. The darker evidence concerns stress. During the Flash Crash,
high-frequency traders did not trigger the crash but their synchronized inventory management amplified
it \citep{kirilenko2017}; order-flow toxicity rose before the crash and liquidity provision withdrew
when flow became toxic \citep{easley2011}; the August 2007 quant unwind revealed that common strategies
produce simultaneous deleveraging \citep{khandani2011}; and ultrafast extreme events proliferate in a
machine ecology operating beyond human response time \citep{johnson2013}. \citet{budish2015} show that
continuous-time trading creates a latency arms race with correlated sniping and propose frequent batch
auctions, which motivates our latency-matching treatment and call-auction robustness arm.

The algorithmic-monoculture literature gives the homogenization mechanism a name. \citet{kleinberg2021}
show that when many decision-makers rely on the same algorithm, outcomes can be worse even if that
algorithm is individually more accurate; \citet{bommasani2021} identify homogenization as a
structural risk of foundation models, and \citet{bommasani2022} document outcome homogenization
across deployers of shared models. On the output side, generative AI raises individual creativity but
reduces collective diversity \citep{doshi2024}, standard LLM outputs show diminished diversity of
thought relative to human samples \citep{park2024brm}, and LLM assistance homogenizes human ideation
\citep{anderson2024}; conversely, ensembles of twelve LLMs match human crowd forecasting accuracy
\citep{schoenegger2024}, which illustrates the diversity benefit our H4 tests in a market. The
algorithmic-collusion literature shows that learning algorithms can coordinate without being told to:
Q-learning pricing algorithms learn tacit collusion \citep{calvano2020,klein2021}, algorithmic pricing
raised margins in German retail gasoline \citep{assad2024}, asymmetric adoption softens competition
\citep{brown2023}, LLM pricing agents reach supracompetitive prices with a sensitivity to prompt
wording \citep{fish2024}, and LLM sellers in continuous double auctions collude more when they can
communicate \citep{agrawal2025}. In securities markets, \citet{colliard2026} show that Q-learning market
makers learn adverse-selection pricing but quote markups that shrink with the number of algorithms,
and \citet{dou2025} show that reinforcement-learning informed speculators in a Kyle-type market sustain
supracompetitive profits through price-trigger strategies or homogenized learning biases, lowering
price informativeness and liquidity. \citet{dou2025} is the nearest theoretical neighbour of our
mechanism; the difference is that their channel is strategic, whereas ours is a shared-error structure
in signals, and that they have no language models.

The policy literature has already drawn the systemic conclusion. \citet{gensler2020} argue that deep
learning increases uniformity and interconnectedness because models built on the same data generate
correlated predictions; \citet{danielsson2022} show that AI in risk management can create endogenous
systemic risk through common models and procyclical responses; the \citet{fsb2024} warns that common AI
models and data sources could raise correlations in trading, lending and pricing;
\citet{danielsson2025} model AI as amplifying existing vulnerabilities and making crises faster;
\citet{aldasoro2025} list monoculture and correlation among the financial-stability concerns of an
intelligent financial system; and \citet{meng2026} propose a unified model in which the AI adoption
share and signal correlation raise systemic coupling convexly, with 13F evidence of tail-loss
amplification. None of these provides a controlled experiment; the present paper is an attempt to
supply one.

\subsection{LLM agents in economics and markets}
\label{sec:literature:llm}

\paragraph{LLMs as simulated subjects.}
\citet{horton2023} proposed using LLMs as simulated economic agents to replicate classic behavioural
experiments at low cost; \citet{aher2023} showed that LLMs can simulate multiple humans and replicate
human-subject studies with fidelity that varies by model and task; and \citet{argyle2023} showed that
conditioning on demographic backstories produces ``silicon samples'' matching sub-population response
distributions. GPT choices in budgetary tasks score higher on rationality than humans and are largely
stable across demographic prompts \citep{chen2023pnas}; chatbot choices in behavioural games fall within
the human distribution and are more cooperative than the median human \citep{mei2024}; in repeated
games LLMs cooperate well but coordinate poorly \citep{akata2025}; and expectations elicited from an
LLM reading newspaper articles track professional forecasters and are extrapolative
\citep{bybee2023}. \citet{manning2024} automate hypothesis generation and testing with LLM subjects, and
\citet{immorlica2024} propose a framework in which generative agents are economic agents with new
primitives such as prompts and context. Generative agents with memory, reflection and planning produce
believable emergent social behaviour \citep{park2023}, social-network simulations reproduce
information propagation \citep{gao2023s3}, and EconAgent produces realistic macroeconomic dynamics
\citep{li2024econagent}. \citet{anthis2025} conclude that LLM social simulations are a promising
method provided that validity concerns---homogeneity, sycophancy, drift---are addressed; homogeneity is
the concern this paper turns into a treatment.

\paragraph{Single-agent trading and backtests.}
ChatGPT-derived sentiment on headlines predicts next-day returns \citep{lopezlira2023chatgpt}.
FinMem \citep{yu2023finmem}, TradingGPT \citep{li2023tradinggpt} and FinAgent \citep{zhang2024finagent}
evaluate one LLM agent, or a small team with layered memory and distinct characters, by backtesting on
historical prices. In these designs the market does not react to the agent, so nothing can be learned
about equilibrium prices, and \citet{lopezlira2025memorization} show that such backtests are
contaminated by memorization of the historical period. Both points motivate a synthetic economy with
an endogenous price and no history to memorize.

\paragraph{LLM-populated market simulators.}
A second generation places many LLM agents on a matching engine. ASFM couples LLM traders to a real
order-matching system and finds reactions to rate and inflation shocks consistent with economic priors
\citep{gao2024asfm}; StockAgent adds bulletin-board communication and reports that GPT- and
Gemini-driven agents differ systematically in trading behaviour \citep{zhang2024stockagent};
TwinMarket scales to a social-media-connected population, reproduces bubbles and crashes, and reports
that agents of one model family are more homogeneous \citep{yang2025twinmarket}; FCLAgent combines
LLM-decided direction with rule-based sizing and reproduces path-dependent patterns absent from
conventional agent-based models \citep{hashimoto2025}; StockSim provides an open order-level simulator
with latency, slippage and price--time matching \citep{papadakis2025}; and \citet{li2026switch} find
that LLM agents' trading-style switching in a year-long simulation is only partially consistent with
behavioural-finance theory. \citet{lopezlira2025trade} is the closest in spirit: LLM agents in a
persistent order-book simulation adhere to assigned strategies, generate price discovery and bubbles,
and are analysed with partial-dependence-style perturbations. These papers establish that LLM markets
\emph{can} produce the patterns we study. They are descriptive in the sense that the population
composition is fixed, the information set of LLM agents is not held equal to that of the agents they
replace, and the fundamental value, the efficient price and the transaction price are not separated.

\paragraph{Experimental-finance replications.}
A third strand replicates laboratory markets with LLM subjects. In Smith--Suchanek--Williams-type
markets, LLM agents price near fundamentals and form only muted bubbles, with lower variance in their
strategy text than humans \citep{henning2025}. In a simulated call auction, LLM agents exhibit
disposition effects and extrapolative beliefs, and prompt interventions causally change bubble size
\citep{ouyang2026}. Across fifteen models, \citet{saxena2026} find behaviour ranging from fundamental
pricing to human-like bubbles, with mixed populations unstable and advanced models adapting to exploit
weaker ones. Replications of \citet{smith1962} find slower or absent convergence with model-family
differences \citep{struski2026}, and the trading institution alone moves the efficiency of identical
LLM agents from 56\,\% to 89\,\% \citep{chupilkin2026}. These studies answer the question of
behavioural fidelity---whether LLMs behave like humans---and their answers conflict. Our results do not
require fidelity: they require only that errors be correlated within a model.

\paragraph{Correlated behaviour from shared inputs.}
The single paper that measures correlated trading from shared inputs is \citet{song2026}: overlap in
retrieved documents among retrieval-augmented trading agents raises mean pairwise action correlation
from 0.04 to 0.41 and increases position volatility without raising profits. That design is a backtest
without market feedback, so it cannot speak to informativeness or liquidity. \citet{cho2025} document
peer conformity in LLM multi-agent systems, with the confidence gap and format of peer information
controlling its strength. In LLM re-creations of herding experiments, \citet{hansen2025} find that AI
agents rely more on private signals than humans but can be prompted to herd when maximizing profit.
Finally, LLM service behaviour drifts over months \citep{chen2024hdsr}, which is why we cache every
response and replay from a frozen decision library.

\subsection{What this paper adds}
\label{sec:literature:adds}

We should be candid about overlap. The hypothesis that shared models correlate trading is not ours: it
is stated by \citet{gensler2020}, the \citet{fsb2024} and \citet{danielsson2025}, modelled as
homogenized learning biases by \citet{dou2025}, and measured in a backtest by \citet{song2026}.
Model-family homogeneity inside LLM markets has been observed by \citet{yang2025twinmarket},
\citet{zhang2024stockagent} and \citet{saxena2026}. LLM-populated order-book simulators exist and are
open source \citep{lopezlira2025trade,papadakis2025,gao2024asfm}. Our simulator is therefore not the
contribution; the design is. First, we treat the LLM penetration share as a randomized treatment
assigned to trader slots by an independent seed stream, so that matched pairs across the share share
every other stream---fundamentals, news, endowments, latencies, policy assignment, activation,
tie-breaks---and differences in outcomes are causal effects on the same market history
(Section~\ref{sec:identification}). Second, we separate three prices: the latent value known to the
simulator only, the public value implied by announced dividends and news, and the transaction price, so
that an information-efficiency loss can be distinguished from transitory pressure
\citep{hasbrouck1993,hendershott2014}. Third, we derive the shared-error mechanism in a Kyle--Hellwig
setting with a model-specific common error and an effective number of models, and then test it by
intervention---prompt diversity at equal share, a two-model mix, inherited information sets, memory,
communication cohorts---so that the effective number of models is manipulated rather than observed.
Fourth, we report the realized dose (the share of orders and volume actually originated by LLM agents)
alongside the assigned share, and we run fairness designs with equal decision opportunities, equal
dollar budgets and matched latency, because naive comparisons confound ``being an LLM'' with being slow
and infrequent. Fifth, the leverage interaction and the safeguard counterfactuals connect the LLM
literature to \citet{brunnermeier2009}, \citet{thurner2012} and \citet{kirilenko2017}, with thresholds,
estimators and seed counts fixed before the first confirmatory call.