UQO Working Papers · WP27 · v1.0
The Invisible Hand of Machines: Price Discovery in Markets Populated by LLM Agents
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
LLM traders make a market deep, calm and slow to learn
Raising the LLM share from 0 to 95 percent of slots tightens spreads from 18.6 to 6.1 basis points, multiplies depth by seven and leaves pricing error unchanged, while the five-tick variance ratio climbs from 1.1 to 2.1.
Abstract
What happens to price discovery when a growing share of the traders in a market are autonomous language-model agents? The paper builds an event-sourced continuous double-auction exchange with exact accounting, a latent dividend process that separates fundamental value from public value and transaction price, seven calibrated classical trader types, and an adapter that lets a language model trade from a canonical snapshot through a strictly validated JSON action. Randomized, matched-seed experiments assign 0 to 100 percent of trader slots to a small reasoning model (gpt-5.4-nano), holding fundamentals, news, endowments, latencies and information sets fixed, under a pre-registered analysis plan with seven matched seeds per live cell and thirty per classical cell.
A stylized model with correlated signal errors predicts that a shared model degrades informativeness, raises order-flow correlation and amplifies leverage-driven spirals, and that model diversity restores informativeness. The simulator confirms the predictions on coordination and diversity and rejects them on prices: raising the LLM share tightens quoted spreads from 19 to 6 basis points, multiplies depth by seven, cuts price impact by 60 percent and leaves the pricing error unchanged, while the market absorbs latent information more slowly and returns become predictable (variance ratio 2.1, lag-one autocorrelation 0.32 at full penetration).
No bubble or crash appears in 112 live runs, and with leveraged traders present the LLM agents halve the pricing error and remove the bubble episodes that leverage creates. Prompt diversity and a two-model mix lower coordination and shorten absorption at a liquidity cost; a communication channel and memory raise coordination and pricing error slightly. The unifying mechanism is patient, public-value-anchored liquidity provision by agents who act infrequently. All results are causal statements about the simulator, reproducible from a one-command replication package; the Anthropic cells of the design are implemented but could not be run.
Key findings
18.6 → 6.1 bp
quoted spread as LLM share rises to 95%
×7
order-book depth at full LLM penetration
−60%
price impact per 100 lots
0
bubbles, crashes or halts in 112 live runs
0.02 → 0.32
lag-one return autocorrelation, 0% to 95% LLM
2.16% → 0.96%
tracking error with leverage, 0% vs 25% LLM
Scientific contributions
- A counterfactual market with randomized LLM share
- One seed fixes dividends, news, endowments, latencies and slot assignments; the treatment replaces a stratified subset of slots by LLM agents inheriting their information set, so matched pairs differ only in who decides.
- Four prices kept apart
- Latent fundamental value, public value from a Kalman filter on announced dividends and news, transaction price and quote midpoint are distinct, so price discovery is measured against the truth rather than a known target.
- Theory confirmed on coordination, rejected on prices
- LLM agents deciding at the same tick are on the same side nine times out of ten and a communication channel raises coordination, yet pricing error is unchanged (−0.18 percentage points, interval [−0.67, +0.30]) and spreads fall by 12.5 basis points (d = −2.2).
- LLM agents dampen leverage spirals
- With 19 percent leveraged traders, replacing 25 percent of slots by LLM agents lowers tracking error from 2.16 to 0.96 percent (p = 0.03) and removes the bubble episodes (0.7 per run to 0) that leverage creates in the classical economy.
- The cost is slower, predictable price paths
- Latent information is absorbed more slowly as the share rises; the five-tick variance ratio climbs from 1.14 to 2.12 and return autocorrelation from 0.02 to 0.32 (d = 3.5), a drift that trend followers can exploit.
- Diversity trades coordination for liquidity
- Diverse prompts or a two-model mix at 25 percent lower pairwise flow correlation from 0.40 to 0.32–0.30 and absorb shocks four ticks faster, at a spread cost of 0.5–4.6 basis points.
Methodology
Methodology (4)
Event-sourced limit-order-book simulator
Continuous double auction with price–time priority, partial fills, fees, per-agent latency, order TTL, price bands, circuit breaker, position limits, call auctions and margin-driven liquidation; integer-cent accounting checked after every fill with deterministic replay (24 unit, property and integration tests).
Stylized correlated-error model
Six propositions on informativeness versus shared-model share, order-flow correlation and synchrony, effective number of models, feedback amplification, leverage spirals and safeguards; each primitive maps to a simulator parameter and the sign of every prediction was pre-registered.
Common random numbers and paired inference
Named random streams per seed; outcomes at the run level (pricing error, absorption time, variance ratios, spreads, depth, Kyle lambda, bubble and crash thresholds, synchrony); paired t with 95 percent intervals, Cohen's d, Wilcoxon, seed-fixed-effects OLS, Holm within hypothesis families and local projections.
Strictly validated LLM adapter
OpenAI Responses API with JSON schema, invalid actions mapped to hold and logged, bounded memory, optional cohort communication channel and prompt personas, USD and token caps, SQLite decision cache and a frozen k-NN decision library for zero-cost replay at 1,000 traders.
Research data
- Simulated exchange economy
- Entirely synthetic: 100-trader markets of 600 ticks with dividends every 20 ticks, a latent AR(1) innovation, private signals and news; seven classical trader types calibrated to stylized facts (near-zero return autocorrelation, fat tails, volatility clustering, 10–30 bp spreads); every run stores config, manifest, ticks, trades, agents and LLM decisions as parquet.
- Live LLM decision records
- 80,031 decisions by gpt-5.4-nano (and gpt-5.4-mini in two-model and equal-budget arms) in the confirmatory design plus 35,652 in the dose arm, seven matched seeds per live cell across six shares, three cognition arms, a leverage interaction, two safeguards and two fairness arms; total API spend 56 USD.
- Classical validation design
- Sixteen cells × thirty seeds without any language model, plus a 360-run calibration sweep, reproducing bubbles under leverage and trend-heavy populations, misleading-news effects and two-asset contagion through common funding.
Reproducibility
- One-command replication: make replicate regenerates classical designs, theory checks, figures and tables without an API key in about five minutes; make confirmatory reruns the live design for about 60 USD
- Pre-analysis plan in docs/identification.md and docs/protocol.md; every post-plan change logged in docs/research_log.md; literature ledger of 143 verified references
- Every run directory carries a manifest with config hash, git revision, seeds, model ids, prompt version and realized LLM shares; 24 tests including hypothesis-based property tests
- Full LaTeX source of the 101-page paper with proofs, prompts, algorithms and hyperparameters; SSRN 7679801
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.0current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 101 p. |