we-solve (beta)

Econ-111 · economics · Economics

Algorithmic pricing and collusion

LLM status Open unverified

Statement

Original identifier: I4 (atlas item 74). Classification: Algorithmic competition robustness research program. Original tractability label: Medium (editorial, not a complexity result).

Definitions and assumptions

Consider a specified repeated differentiated-product price game with firms $i=1,\ldots,n$, demand $q_i(p,s)$, marginal costs $c_i$, and common state $s$. Each independently installed algorithm maps its own information history and internal learning state into a price distribution. Define a benchmark $p^N(s)$ as a selected one-shot Nash equilibrium of the same demand and cost game; competition need not mean zero markups. Define average excess profit or price relative to this benchmark, weighted by prespecified quantities. Specify information sharing, vendor overlap, exploration, replacement, entry, and discounting. Independence of installation does not imply statistical independence of observed prices.

Formal research question

Characterize algorithm and market classes for which supracompetitive behavior is robust. For example, with $G_t$ the declared profit gap above the benchmark, determine conditions yielding \[ \Pr\!\left(\liminf_{T\to\infty}\frac1T\sum_{t=1}^TG_t\geq\varepsilon\right)\geq1-\alpha \] for fixed $\varepsilon>0$ and $\alpha\in(0,1)$, or establish finite-horizon analogues when learning does not converge. Evaluate robustness to heterogeneous vendors and profitable unilateral deviations. Separately identify causal effects of adoption on market outcomes using randomized adoption or credible exogenous variation, with explicit assumptions about simultaneous adoption and spillovers.

Interpretation and scope

The bare existence question has already received affirmative constructive evidence in particular simulations. The substantive extension concerns when that behavior survives realistic costs, demand shocks, algorithm turnover, and entry. Parallel prices can reflect common costs or competitive response and therefore do not identify tacit cooperation. Nor does an above-Nash average price prove legal collusion: the economic benchmark and legal agreement standard are distinct. A learning algorithm may fail to converge, so a limiting-price claim requires a convergence proof or a justified stochastic stationary distribution. Finite-run simulation estimates should report sensitivity to random seeds, initialization, and exploration schedules and cannot prove the displayed probability bound for all environments. Deviation tests need to allow firms to modify algorithms, not merely change one price while retaining a fixed learning rule. Empirical adoption studies additionally confront endogenous timing and local market exposure. A model-based robustness result and a causal market estimate are complementary deliverables with different evidential requirements.

Source audit and status

[S1] provides Q-learning simulation evidence; [S2], published in 2024, provides German gasoline-market adoption evidence; [S3] analyzes legal-economic issues involving autonomous agents. All citations are verified. Consequently this item is corrected from an unanswered existence claim to a robustness and identification frontier. The displayed asymptotic criterion is an editorial target and is not asserted as a theorem or universal empirical fact established by these papers.

References

    },leftmargin=*]

  1. Emilio Calvano, Giacomo Calzolari, Vincenzo Denicolò, and Sergio Pastorello (2020). Artificial Intelligence, Algorithmic Pricing, and Collusion. American Economic Review 110(10), 3267--3297. https://www.aeaweb.org/articles?id=10.1257/aer.20190623

    Source role: Constructive simulation evidence for algorithmic supracompetitive prices. Locator: Abstract and article metadata.

  2. Stephanie Assad, Robert Clark, Daniel Ershov, and Lei Xu (2024). Algorithmic Pricing and Competition: Empirical Evidence from the German Retail Gasoline Market. Journal of Political Economy 132(3), 723--771. https://doi.org/10.1086/726906

    Source role: Market-specific evidence on adoption and competition. Locator: Abstract and article metadata.

  3. Joseph E. Harrington, Jr. (2018). Developing Competition Law for Collusion by Autonomous Artificial Agents. Journal of Competition Law & Economics 14(3), 331--363. https://doi.org/10.1093/joclec/nhy016

    Source role: Legal-economic conceptual analysis. Locator: Abstract and article metadata.

Common conventions: Source mathematical conventions

Unless an entry states otherwise, agent and firm sets are finite. State and action spaces are measurable, strategies and outcome maps are measurable, probability laws are specified, and expectations exist for the targets under discussion. Infinite-horizon discount factors lie in $(0,1)$ and discounted objectives are finite. Norms, tolerances and complexity input sizes must be specified for computational comparisons. Constants or primitives that appear locally are inputs, not empirical facts asserted by this catalogue.

Identification. A statistical model $m\in\mathcal M$ implies an observable law $P_m$ and target $\theta(m)$. At law $P$, its identified set is \[ \Theta(P;\mathcal M)=\{\theta(m):m\in\mathcal M,\ P_m=P\}. \] Point identification holds when this set is a singleton, or equivalently when $P_{m_1}=P_{m_2}$ implies $\theta(m_1)=\theta(m_2)$ for all compatible models. Sharp bounds mean bounds attained, or approached when the boundary is not attained, by compatible models. Writing this set is a definition, not a solution: the substantive task is to establish informative restrictions and derive or estimate the set. An unrestricted-confounding model may give only trivial bounds.

Inference. Identification, estimability and valid uncertainty quantification are separate questions. Fix $\alpha\in(0,1)$. A confidence procedure $C_n$ over a declared law class $\mathcal P$ has asymptotic uniform coverage at level $1-\alpha$ when \[ \liminf_{n\to\infty}\inf_{P\in\mathcal P}\Pr_P\{\theta(P)\in C_n\}\geq1-\alpha. \] For set-identified targets the coverage event must be defined separately, for example coverage of the full identified set. If an entry uses identification-indexed models instead of uniquely indexed laws, the same coverage convention is applied over those models. Pointwise limits do not establish uniform validity, and asymptotic validity does not guarantee finite-sample precision. Sample sizes, dependence structures and weak-identification sequences are part of each application.

Counterfactuals. $Y_i(a)$ is a potential outcome under a specified intervention, including its timing, implementation and versions. It is not $Y_i$ conditional on voluntarily choosing $a$. A full intervention vector permits interference; shorthand scalar treatment notation does not silently impose no interference. Assignment, selection, attrition, anticipation and measurement must be specified. A claim about an instrument's local effect is not automatically a population policy effect.

Economic equilibrium. A counterfactual model includes best responses, constraints, feasibility, market clearing, state transitions and expectations consistent with the stated information. When several equilibria exist, either a declared selector is an input or the target is set-valued. Comparisons across policy regimes must use the stated selection convention. Reduced-form relationships are not presumed invariant after an equilibrium-changing intervention.

Optimization and welfare. Optimization is over the stated feasible set. If attainment is not established, $\sup$ or $\inf$ and $\varepsilon$-optimal choices replace an asserted maximizer or minimizer. For example, with value $V=\sup_{a\in\mathcal A}W(a)<\infty$, an $\varepsilon$-optimal set is $\{a\in\mathcal A:W(a)\geq V-\varepsilon\}$ for $\varepsilon>0$. Benefits, costs, risk and health or environmental outcomes can be aggregated only using declared units and conversion weights. Interpersonal utility weights, the treatment of fiscal transfers and foreign profits, discounting, and the behavioral welfare criterion are explicit inputs. A comparative-static derivative additionally requires existence and appropriate differentiability of the selected counterfactual.

Sources. Local labels [S1], [S2], and so forth restart in every question. Locators identify the relevant abstract, model, section or result at the level actually checked; a bibliographic or abstract-level verification is not represented as inspection of an entire proof. Working papers are distinguished from later publications and changing versions. Source summaries are paraphrased. All URL checks and editorial formulations refer to the audit date stated above.

Statement licence: CC-BY-4.0 · Source file

Contributions 1

Econ-111-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-111

Recurrent algorithmic pricing with turnover and credible deviations

Problem ID: L41-109. Status: Unresolved. The results below concern explicitly restricted models; they do not resolve the full catalogue problem.

Question and declared economic model

The catalogue asks when algorithmic prices stay above a same-market Nash benchmark, survive replacement and unilateral algorithm changes, and can be distinguished empirically from common shocks. We analyze an explicitly installed public-state pricing rule. Its invariant distribution and incentive condition can be calculated exactly. This is a robustness result for a specified rule, not a convergence theorem for independently trained learning systems.

There are two firms, zero marginal cost, discount factor $\delta\in(0,1)$, and stage demand \[ q_i(p_i,p_j)=\max\{0,a-bp_i+dp_j\},\quad a>0,\quad b>d\ge0. \] Prices lie in $[0,a/(b-d)]$, a compact action set. Profits are $p_iq_i$. The selected symmetric one-shot equilibrium price and the joint-profit-maximizing price among symmetric profiles are \[ p^N=\frac{a}{2b-d},\qquad p^C=\frac{a}{2(b-d)}. \] At these profiles define \[ \pi_N=\frac{ba^2}{(2b-d)^2},\quad \pi_C=\frac{a^2}{4(b-d)},\quad p^D=\frac{a+dp^C}{2b},\quad \pi_D=\frac{(a+dp^C)^2}{4b}. \] For $d>0$, $\pi_D>\pi_C>\pi_N$. The formulas also reveal the degenerate case $d=0$, when competitive and joint prices coincide and no strictly positive gap is possible.

Lemma (Global stage best responses): Against either $p^N$ or $p^C$, the global best-response price in the specified action interval is $(a+dp_j)/(2b)$. Consequently the above profits are the actual best-deviation, cooperation, and punishment profits.
Proof: Where demand is positive, profit is $p_i(a+dp_j-bp_i)$, a strictly concave quadratic with that maximizer and positive maximum. Where demand is zero, profit is zero. The maximizer lies within the action interval for each displayed opponent price, so truncation does not change it. Substitution gives the formulas. At $p^N$ the best response equals $p^N$. Maximizing $2p(a-(b-d)p)$ along symmetric profiles gives $p^C$. Direct substitution, or the strict quadratic improvement from $p^C\ne p^D$ for $d>0$, gives the strict inequalities. ∎

Turnover and recovery as an explicit public process

The public state is $C$ or $N$. In state $C$ both rules quote $p^C$; in state $N$ they quote $p^N$. After prices are observed, an intended next state is $C$ if current state is $C$ and both complied, and is $N$ otherwise. An independent public transition then changes intended $C$ to actual $N$ with probability $\lambda$, and intended $N$ to actual $C$ with probability $r$. Assume \[ 0<\lambda,r<1,\qquad \lambda+r<1. \] Thus $\lambda$ models public disruption of a cooperative regime, while $r$ models its public restoration. These transitions are primitives, observable to both firms. A deviation in state $N$ cannot change the transition rule. Future random transition draws are unavailable at the time of pricing.
Theorem (Exact credibility and long-run gap): For $d>0$, the public-state rule is sequentially optimal against every unilateral history-dependent pricing algorithm if and only if \[ \delta(1-\lambda-r)\ge \frac{\pi_D-\pi_C}{\pi_D-\pi_N}. \tag{1} \] When both firms use the rule, from either initial state, \[ \lim_{H\to\infty}\frac1H\sum_{h=1}^H \left(\frac{\pi_{1h}+\pi_{2h}}2-\pi_N\right) =\frac{r}{\lambda+r}(\pi_C-\pi_N) \quad\text{almost surely}. \tag{2} \] Thus for every $\varepsilon$ strictly below the right-hand side of (2), the catalogue's limiting-gap event has probability one. Strict (1) and a strict gap inequality persist under sufficiently small perturbations of these model parameters.
Proof: Let $V_C,V_N$ be discounted payoffs following the rule, and write $\gamma=\delta(1-\lambda-r)$. Their Bellman equations imply \[ V_C-V_N=(\pi_C-\pi_N)+\gamma(V_C-V_N) =\frac{\pi_C-\pi_N}{1-\gamma}. \] A deviation in state $C$ changes the probability of next state $C$ from $1-\lambda$ to $r$. Its largest possible current gain is $\pi_D-\pi_C$. Hence obedience is optimal exactly when \[ \pi_D-\pi_C\le\gamma(V_C-V_N), \] which rearranges to (1). In state $N$, the current prescribed price is a global stage best response and the next-state distribution does not depend on the deviator's action.

These one-period inequalities imply optimality against an arbitrary replacement algorithm: for every chosen action, current profit plus discounted expected prescribed continuation value is bounded above by the current state's $V$. Iterating conditional expectations through $H$ periods gives the same inequality for any history-dependent deviation with remainder at most $\delta^H\max|V|$, which tends to zero. Necessity follows from using the single best current deviation when (1) fails.

Under compliance the state transition matrix is \[ \begin{pmatrix}1-\lambda&\lambda\\r&1-r\end{pmatrix}. \] For completeness, its long-run occupation fraction can be obtained by regeneration at returns to $C$. A cycle consists of one $C$ observation and, with probability $\lambda$, a geometric number of $N$ observations of mean $1/r$. Successive cycles are independent with finite mean lengths. The strong law for their lengths gives fraction $1/(1+\lambda/r)=r/(\lambda+r)$ in $C$; an incomplete final cycle has asymptotically negligible length, as follows from finite mean and the strong law. Only $C$ periods have excess profit. This proves (2). All displayed functions are continuous on the strict parameter domain, proving the perturbation claim. ∎

A feasible parameter check and a turnover warning

For $a=b=1$ and $d=1/2$, \[ p^N=2/3,\quad p^C=1,\quad p^D=3/4, \qquad\pi_N=4/9,\quad\pi_C=1/2,\quad\pi_D=9/16. \] The credibility threshold is $9/17$. Taking $\delta=0.95$, $\lambda=0.05$, and $r=0.10$ gives $\gamma=0.8075>9/17$. The asymptotic average profit gap is $(2/3)(1/18)=1/27$. These are exact model calculations, not simulated estimates or observed market effects.
Proposition (Absorbing turnover destroys the asymptotic criterion): If $r=0$ and $\lambda>0$, then under compliance state $N$ is reached in finite time almost surely and the long-run average excess profit is zero, regardless of whether initial cooperation satisfies its incentive constraint.
Proof: The probability of surviving $h$ consecutive $C$ periods is $(1-\lambda)^h$, tending to zero. After the first disruption the process stays in $N$. A bounded finite prefix contributes zero to the limiting average. ∎
This separates persistent collusion from a long but ultimately transient episode. A finite-horizon simulation could easily confuse the two when $\lambda$ is small.

Remaining identification and learning questions

The public state and restoration mechanism coordinate behavior explicitly. The result does not show that separate vendors learn this rule, or that private noisy monitoring admits the same punishments. Independent execution errors can be represented by a public disruption rate only when their detection and resulting transitions actually follow the stipulated process. Unobserved common costs require an expanded benchmark.

An empirical adoption effect would also need an assignment design and exposure definition. Equations (1)--(2) establish no causal effect of installing an algorithm in an actual market and make no legal finding. The source papers [Calvano,Assad] motivate the simulation and empirical sides of that remaining task.

References

  • [Calvano] E. Calvano, G. Calzolari, V. Denicolò and S. Pastorello (2020), Artificial Intelligence, Algorithmic Pricing, and Collusion, American Economic Review 110(10), 3267--3297. Primary abstract and metadata: https://www.aeaweb.org/articles?id=10.1257/aer.20190623. The article studies learning simulations; our public-state theorem is a separate calculation.
  • [Assad] S. Assad, R. Clark, D. Ershov and L. Xu (2024), Algorithmic Pricing and Competition: Empirical Evidence from the German Retail Gasoline Market, Journal of Political Economy 132(3), 723--771. https://doi.org/10.1086/726906. This is market-specific adoption evidence, not a proof of (1) or (2).

Completion Estimate

25%

Go to contribution →