Working paper, October 15, 2026

From Individual Decisions to Collective Outcomes: A Scoping Review of LLM Agents in Financial Markets

John Garcia, School of Management, California Lutheran University, Thousand Oaks, CA, USA

Personal websiteORCID 0000-0002-7062-0032

Here you can search the 599 papers, inspect the 30 price-setting designs and check a claim against Table 6.

Read the review Explore the evidence Download the paper from SSRN

Where the market-level evidence sits

Each dot is one of the 271 primary studies in the synthesis set. Gold dots are the studies in which agents’ decisions set prices.

Of these, 1 ran concurrent human markets: Henning et al. (2025). Counts from Online Resource 1; see Sections 3.2 and 6.1. See the 30 price-setting designs

Four conclusions

From the Introduction of the review. Each conclusion opens to the evidence the review gives for it.

Individual behavior belongs to a configuration (a model, version, reasoning mode, prompt, and elicitation instrument) rather than to “LLMs” as a class of models.

The evidence

GPT-3.5 made budget-set choices more consistently than human subjects given identical tasks in a parallel experiment, although its consistency declined under an alternative-price presentation (Chen et al., 2023). Loss aversion varies across releases of one model family, from above the human level in GPT-3.5, through human-like values in GPT-4 and GPT-4o, to below neutrality in GPT-4.1 (Wang et al., 2025d). Responses to prompt-embedded reference points differ three-fold across models (Garcia, 2026a).

In the few designs that vary them, aggregate outcomes depend on the trading institution and agent population composition, not solely on the decision rule.

The evidence

In one study that varied both the trading institution and the model family while holding payoffs fixed, both factors materially changed the outcomes (Chupilkin, 2026; Section 6.2). Otherwise identical forecasting markets bubble or price at the fundamental depending on which model populates them; a six-model market bubbles in about half of its runs even though its bubble-forming models are a minority; and frontier models that forecast fundamentals among their own kind switch to trend following among bubble-forming agents (Saxena et al., 2026). Measuring a rule in a homogeneous population therefore excludes how it responds to other agents (Section 6.4).

Dependence is a plausible aggregation channel, but its importance relative to individual bias remains unmeasured.

The evidence

In one prediction-market study, pooled honest-agent errors correlate at 0.696, which a variance proxy equates to approximately 1.4 independent forecasters among ten agents (Begin et al., 2026). In contrast, synchronized purchases in a live fleet demonstrate common actions, which can be correct responses to shared information rather than correlated errors (Barton et al., 2026b). Human bubble-forming markets show a similar common error component (Saxena et al., 2026), and none of the 28 price-setting designs read in full compares the market consequences of correlated errors with those of individual bias.

Direct aggregation evidence is scarce.

The evidence

Of the 271 primary studies in the synthesis set, 215 elicit decisions independently, and 30 allow agents’ decisions to affect prices (Online Resource 1). Among these 30, one runs human markets under the same design, and three compare agent dispersion with human markets of the same design, with two finding less heterogeneity among LLM agents. These comparison counts hold under the alternative treatments of uncertain codes examined in Section 8. The first and fourth conclusions rest on many studies and on the audit counts. The second and third rest on a few designs, and 23 of the 30 price-setting designs are working papers; they are better read as hypotheses that these designs support rather than as established regularities (Online Resource 2 summarizes the evidence behind each conclusion).

Use the companion

Read the reviewThe full text with linked citations, tables and figures. Explore the evidenceSearch and filter the 599 papers in the corpus and see how each one was coded. Price-setting designsThe 30 designs in which agents’ decisions set prices, audited on 12 components. Check a claimTurn Table 6 into a checklist for your own study or a referee report. Research agendaThe eight priorities proposed in Section 9. Data and methodsHow the corpus was built, the use of LLM agents, and the Online Resources.

Abstract

Large language models (LLMs) increasingly make, recommend, and simulate financial decisions, yet most evidence concerns agents deciding alone. This scoping review asks when that evidence generalizes to interacting agents and financial markets. From a corpus of 599 papers available through September 2026, it combines a 276-paper narrative synthesis with an audit of 295 decision-system papers, organized by three levels (individual decisions, coordination within decision systems, and interaction among participants) and a cross-cutting validation layer. Evidence does not transfer automatically across levels. Individual behavior depends on model configuration and elicitation, whereas market outcomes depend on the trading institution and population composition. In one replication, frontier models that forecast fundamentals in homogeneous populations switch to trend following among bubble-forming agents. Direct aggregation evidence remains scarce: of the 271 primary studies in the synthesis, 30 let agents’ decisions set prices. Dependence among agents sharing models, prompts, and information is a plausible aggregation channel, but its market consequences relative to individual bias remain unresolved. The review derives claim-specific validation criteria for human fidelity, explanatory value, artificial-agent participation, tool performance, and measurement. For regulators, financial institutions, and market researchers, these criteria distinguish evidence about isolated decisions from evidence about price discovery, common exposures, and market stability.

Keywords: Large language models · LLM agents · Agent-based modeling · Experimental asset markets · Behavioral finance · Algorithmic monoculture

JEL Classification: C45 · C63 · C92 · D84 · G14 · G41

Three decision levels and a validation layer

Levels 1–3 are linked: individual decision rules enter decision systems, and the outputs of these systems are the actions faced by other participants. The validation layer, labeled Level 4, asks whether a measurement taken at one level is valid for use at the next level.

From the caption of Fig. 2.

Open the interactive framework

OUTCOMESAggregate outcomesPrices, efficiency, bubbles, runs, and the dispersion of advice LEVEL 3Interaction among participantsSeparately motivated agents affect one another through prices, payoffs, networks, or shared advice, including mixed markets LEVEL 2Coordination within a decision systemSpecialized LLM roles (analysts, risk managers, traders, critics) exchange information inside one pipeline serving one objective LEVEL 1Individual decisionWhat a model chooses, estimates, forecasts, or advises when it faces a task alone, under a stated prompt and elicitation instrument VALIDATION LAYERValidation across levelsIs a measurement taken at one level valid for its use at the next?

Papers in the corpus by year

The corpus contains 25 papers from 2023, 75 papers from 2024, 221 papers from 2025, and 278 from January 1 to September 27, 2026.

Table 1 and Online Resource 1. The 2026 bar ends on September 27.

Papers per year

Corpus papers by the year of the version recorded; 2026 runs to September 27.

252023

752024

2212025

2782026