we-solve (beta)

Econ-283 · economics · Economics

AI trading and market efficiency

LLM status Open unverified

Statement

Original subject group: Asset pricing and financial markets

Audit classification: Published research agenda. The agenda supports studying algorithmic trading and stability. A specific learning game and collusion statistic are editorial; strategic algorithmic collusion is outside a strictly nongame catalogue.

Definitions and precise research target

Fix a repeated asset market with a fundamental payoff $V$, price $P$, private signals, an order-processing rule and strategic traders. A learning system is specified by its training observations, objective, action space, update rule and computational restrictions. Let $s$ denote the market share using a specified family of artificial-intelligence trading systems. Define an intervention that varies $s$, with competitive and informational responses of other traders allowed. At share $s$, measure price inefficiency $E[(P-V^{\mathrm{fund}})^2]$, where $V^{\mathrm{fund}}$ is the declared date-specific fundamental present value, effective spreads, trading volume and the expected resource loss of market disruption. Define a collusion statistic as the persistent excess of joint trading profits or spreads over a separately computed competitive benchmark under identical information and technology. The task is to identify how these outcomes change with $s$ and characterize conditions under which learning supports tacit collusion or correlated destabilizing behavior. Vary signal precision, feedback, algorithm diversity and entry while keeping the comparison explicit. Historical backtests cannot reveal all equilibrium responses to widespread adoption; live or experimentally controlled validation must address adaptation and data leakage. Particular learning environments can exhibit inefficient outcomes without proving that all AI trading has that effect.

Additional formal definitions and scope (editorial). Choose a finite evaluation horizon $T$, discount weights and a terminal payoff $V\in L^2$, and fix the market's auction/clearing rule and admissible orders. Fundamental value at date $t$ is a prespecified conditional present value, rather than the undiscounted terminal payoff at all dates. The algorithm specification includes initialization, exploration randomness, training/evaluation split and the optimization target. The share $s$ is explicitly measured by capital, trader count or order flow. Compare price error and profits with an equilibrium benchmark using identical signals, risk tolerance, action space and transaction costs. Excess profits alone do not identify collusion: distinguish market power, risk-bearing rents and information rents, and define the tested conduct restriction. Strategic collusion variants use game-theoretic concepts; a catalogue restricted to nongame economic theory should keep an adoption/identification formulation and label the strategic extension as outside that strict boundary.

For the identification part, declare an admissible class $\Theta$ of structural models, including preferences, technologies, shock laws, measurement/sampling rules and any equilibrium selector. If $O$ denotes the complete observable history and $T(\theta)$ the target defined above, the identified set is $\mathcal I_T=\{T(\theta):\theta\in\Theta,\ \mathcal L_\theta(O)=\mathcal L_{\rm obs}(O)\}$; point identification means this set is a singleton. A $do(a)$ comparison replaces stated policy/technology equations while fixing the declared exogenous law and re-solving the remaining equations. These definitions do not assert that an identification theorem holds without further restrictions.

Variants and assumption changes

  1. Competitive learning adoption (editorial). Treat traders as price takers and specify information acquisition; estimate adoption effects on price information and spreads.
  2. Strategic algorithms (editorial; game-theoretic). Fix a repeated trading game and learning rule; test explicitly defined departures from a competitive conduct benchmark, permitting correlated algorithm behavior.

References and source audit

  1. Official January 9, 2026 web version; locator: The Financial System / Technological innovation; Prudential Architecture / Resilience. Broad agenda; exact model editorial. URL: https://www.bankofengland.co.uk/research/bank-of-england-agenda-for-research.
  1. Bank of England. 2026. Bank of England Agenda for Research, 2025–2028. The Financial System / Technological innovation; Prudential Architecture / Resilience. https://www.bankofengland.co.uk/research/bank-of-england-agenda-for-research.

Common conventions: Source mathematical conventions

These conventions apply to each entry unless explicitly replaced.

Models and identification. A primitive model $m$ specifies all probability laws, feasible choices, information, equilibrium and measurement rules used by its target. The admissible class $\mathcal M$ is fixed independently of outcomes. If $P_O(m)$ is its observable probability law and $\tau(m)$ the target, its identified set at observable law $P$ is \[ \mathcal I_\tau(P)=\{\tau(m):m\in\mathcal M,\ P_O(m)=P\}. \] Point identification means that this set is a singleton. An empty set signals incompatibility of data and assumptions. Coordinate bounds need not describe the joint set. Bounds are sharp only if every retained target is attainable or arbitrarily approachable by compatible models. Finite bounds require explicit restrictions on unobserved quantities; unrestricted confounding, hidden wealth, extrapolation or arbitrary equilibrium selection can yield uninformative sets.

Causal interventions. A potential outcome $Y_i(p)$ is the outcome under a complete policy regime $p$, including timing, financing, information and market spillovers. The target population and units are fixed before treatment. With interference, use the complete assignment vector or a justified exposure mapping; individual no-interference notation alone cannot identify an economy-wide intervention. A difference of mean potential outcomes does not require the cross-world joint distribution. Conditioning on post-treatment migration, survival, enrollment or employment changes the estimand and needs separate justification.

Statistical statements. An estimator, confidence set, forecast or test specifies a sampling law, model class and loss/coverage criterion. Uniform $(1-\alpha)$-coverage means $\inf_{m\in\mathcal M}\Pr_m\{\tau(m)\in C_n\}\geq1-\alpha$, or an explicitly asymptotic version. The whole parameter space trivially has coverage, so informativeness requires a stated width, risk or power objective. A forecasting improvement is relative to a named benchmark, information set and evaluation population.

Equilibrium and optimization. An equilibrium comprises feasible plans, optimality, beliefs where needed, budget constraints and all market/resource clearing. The primitives or a stated benchmark define each correspondence $\mathcal E(p;m)$. Nonemptiness is assumed or proved, never inferred just from compact policy sets. A selected-equilibrium target uses a declared selection rule; a robust target quantifies over all permitted equilibria. Use suprema or infima unless attainment is established. An $\varepsilon$-optimal policy is feasible and within $\varepsilon$ of the finite optimum value. Report an empty feasible set separately.

Welfare and accounting. Normative utility, interpersonal normalization, social weights, numeraire, discounting and the use of public receipts are primitives. Behavioral utility need not coincide with the welfare criterion. Household welfare includes ownership income and rents once; adding profits again to compensating variation generally double-counts them. A monetary equivalent is defined by an explicit transfer and price convention, with monotonicity and range conditions. Average effects, mechanism contributions, transfers and welfare losses are different targets. Infinite sums require convergence and dynamic feasible plans require terminal or no-Ponzi conditions.

Source versions. A working-paper date, revision date and journal publication year are distinguished. A DOI for a journal version is not automatically the DOI of an earlier manuscript. Locators refer to the stated version and printed pages where specified. A metadata-only check does not verify a claim about a full-text conclusion. Every entry's source subsection narrows the provenance claim accordingly.

Statement licence: CC-BY-4.0 · Source file

Contributions 1

Econ-283-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-283

AI trading and market efficiency Correlated information technologies and a collusion benchmark

Question and scope

The target is the effect of AI trading on price accuracy, market quality and strategic conduct under a specified adoption intervention. This attempt separates two tractable mechanisms: replacing independent information with a correlated technology, and sustaining dealer markups by a repeated-game rule. It does not predict that actual learning algorithms collude or identify the causal effect of observed AI adoption.

A competitive information benchmark

A one-period asset pays $V\sim N(0,v)$ at maturity, where $v>0$. There is no further payoff innovation between the trading date and maturity, so $V$ is also the latent contemporaneous full-information fundamental. It is not observed by market makers. Discounting and risk premia are zero; the error studied below is the gap from this full-information benchmark, not from the market makers' own conditional expectation. A share $s\in[0,1]$ of fixed information-processing capacity uses AI; this intervention replaces, rather than adds to, human capacity.

For $0<s<1$, the public order-processing summaries are \[ H_s=V+\xi_H,\qquad A_s=V+Z+\xi_A, \] where $\xi_H,Z,\xi_A,V$ are independent centered Gaussians, with variances $1/[h(1-s)]$, $\rho$, $1/(as)$ and $v$, respectively. Here $a,h>0$ and $\rho\ge0$. The common component $Z$ represents correlated model error shared by the AI technology. At $s=0$ omit $A_s$; at $s=1$ omit $H_s$. Market makers observe both summaries and know their true law. Their zero-expected-profit quote is $P_s={\mathbb{E}}[V\mid H_s,A_s]$. This is an explicit nonstrategic benchmark; signals are observations, not equilibrium truthful reports by strategic traders.

Theorem (Exact pricing error and optimal capacity share): The price mean squared error is \[ M(s)={\mathbb{E}}(P_s-V)^2= \left[v^{-1}+h(1-s)+\frac{as}{1+\rho as}\right]^{-1}. \] For $\rho>0$, the minimizing share is \[ s^*={\operatorname{clip}}_{[0,1]}\left(\frac{\sqrt{a/h}-1}{\rho a}\right). \] For $\rho=0$, it is $s^*=1$ when $a>h$, $s^*=0$ when $a<h$, and every share when $a=h$. If $a>h$ but $a/(1+\rho a)^2<h$, the unique optimum is interior: increasing AI capacity improves price accuracy below $s^*$ and worsens it above $s^*$.
Proof: Conditional on $V$, the two summaries have independent errors with variances $[h(1-s)]^{-1}$ and $\rho+(as)^{-1}$. Multiplying their Gaussian densities by the prior density and completing the square in $V$ gives posterior precision $v^{-1}+h(1-s)+[\rho+(as)^{-1}]^{-1}$. The posterior mean is $P_s$ and the posterior variance is the reciprocal of this precision; it does not depend on the realized signals. Taking expectations gives the MSE formula. Its continuous extension gives the endpoints.

Minimizing $M$ is equivalent to maximizing $I(s)=h(1-s)+as/(1+\rho as)$. For $\rho>0$, $I'(s)=-h+a/(1+\rho as)^2$ and $I''(s)=-2\rho a^2/(1+\rho as)^3<0$. Solving $I'=0$ and clipping proves the optimum and the strict comparative statics. For $\rho=0$, $I$ is affine and the remaining assertions follow. ∎

The deterioration at high AI shares does not contradict the value of information: increasing $s$ removes a source of independent human information. If human capacity were instead held fixed, precision would be $v^{-1}+h+as/(1+\rho as)$, which is nondecreasing in $s$. This distinction is part of the adoption intervention and cannot be settled by a regression coefficient labeled "AI share."

Proposition (Misspecifying the common error): Suppose market makers incorrectly set $\rho=0$ when forming their quote. Let $D_0=v^{-1}+h(1-s)+as$ and $\widetilde P_s=[h(1-s)H_s+as A_s]/D_0$. Then \[ {\mathbb{E}}(\widetilde P_s-V)^2=D_0^{-1}+\rho\left(\frac{as}{D_0}\right)^2. \] In particular, as nominal AI precision $a\to\infty$ at a fixed $s>0$, this error tends to $\rho$, whereas the correctly specified posterior error tends to $[v^{-1}+h(1-s)+\rho^{-1}]^{-1}$ for $\rho>0$.
Proof: Expand $\widetilde P_s-V$ as a linear combination of the independent variables $V,\xi_H,\xi_A,Z$. With $\rho=0$ the sum of its variance terms is $D_0^{-1}$, either by direct multiplication or the posterior calculation above. The additional independent $Z$ term has coefficient $as/D_0$, giving the formula. Taking limits proves the last claim. No excess trading profit or collusive behavior is required for this error. ∎

A fully specified repeated dealer benchmark

Consider a different market with a public asset value and identical risk-neutral dealers. One unit of a trading service is demanded each period. Its marginal resource cost is $c\ge0$ and customers will pay up to $\bar p>c$. There are $n\ge2$ dealers. Simultaneous quotes lie in $[c,\bar p]$; the lowest quote wins, and ties divide demand equally. Quotes and allocations are publicly observed, payoffs are price less cost times allocated demand, and dealers discount future periods by $\delta\in[0,1)$. Fundamentals and inventory risk are constant. The one-period competitive benchmark has every dealer quote $c$, earning zero.

Fix $p^H\in(c,\bar p]$. The following deterministic automaton is completely specified: start in state C; in C quote $p^H$; remain in C only if every observed quote equals $p^H$; after any other quote enter state P forever; in P quote $c$. It has no training stage or exploration. It is a candidate algorithmic policy whose incentives can be checked, not a claim about reinforcement-learning convergence.

Theorem (An exact excess-markup equilibrium): The stated automaton profile is a subgame-perfect equilibrium if and only if $\delta\ge1-1/n$. On its cooperative path the spread over resource cost is $p^H-c$ and joint per-period profit is $p^H-c$, strictly above the competitive benchmark despite identical information, costs and risk.
Proof: In punishment state P, quoting above $c$ yields no sales while quoting $c$ yields zero profit, so no deviation is profitable. In state C, the continuation value from cooperation is $(p^H-c)/[n(1-\delta)]$. A first deviation above $p^H$ earns zero now and thereafter. A first deviation below $p^H$ earns at most $p^H-c$ now and zero thereafter. Thus $1/[n(1-\delta)]\ge1$ makes every first deviation weakly unprofitable. Any deviating strategy either never first deviates, giving the same payoff, or has such a first date, so this check proves sequential optimality after every public history.

Conversely if $1/[n(1-\delta)]<1$, choose a sufficiently small $\eta>0$ with $p^H-\eta\ge c$ and $(p^H-c-\eta)>(p^H-c)/[n(1-\delta)]$. Undercutting by $\eta$ gives a profitable deviation. The profit and spread expressions on the cooperative path follow from equal division of the unit demand. ∎

This theorem makes the benchmark behind "collusion" explicit. It provides no inference that high observed profits imply this automaton: unobserved costs, differentiated services or risk-bearing could alter the competitive counterfactual. Nor does a low pricing MSE rule out the markup equilibrium, since the asset fundamental in this example is public and correctly known.

Identification and the unresolved adoption problem

With observed fundamental realizations, the first model's MSE is estimable at a given share, but the causal curve additionally requires the specified replacement intervention and stable $a,h,\rho,v$. Correlation may change with training-data overlap or vendor concentration. Strategic orders would require a new signal-extraction equilibrium. Private adoption incentives, capital allocation, learning stability and out-of-sample adaptation have not been derived here.

The Bank of England's official research agenda was checked, especially its financial-system discussion of algorithmic trading, market dynamics and technological innovation. It motivates the subject; it does not supply the Gaussian model, the automaton or an empirical collusion conclusion. All mathematical claims used here are proved above.

Completion Estimate

24%

This is a post hoc, heuristic estimate for the full catalogue target. The attempt derives exact pricing errors under correlated AI information and a precise repeated-dealer markup equilibrium. It does not derive learning dynamics or identify the adoption intervention with endogenous strategic responses, spreads, volume, and disruption losses in the full market model.

References

Go to contribution →