Dynamic platforms can be tamed by looking at the long run: stationary-distribution methods turn asynchronous, stochastic marketplaces into tractable static programs, guiding priority rules and information policies when money is unavailable and queue-and-transfer designs when payments are permitted.

Main Finding

Dynamic platform-mediated markets with asynchronous, stochastic arrivals can be transformed into tractable static design problems by characterizing the long-run stationary distribution of the system induced by admissible mechanisms. Using queueing theory and a Positive-Recurrent Regenerative Mechanism (PRRM) framework, complex dynamic incentive and allocation problems (both NTU and TU) reduce to linear programs over stationary distributions subject to implementability (flow/balance) and participation constraints. Key qualitative implications: optimal queue/admission rules typically have simple cutoff/support structures; waiting can act as a non-monetary screening device but imposes welfare losses relative to monetary mechanisms; when transfers are available the designer balances competition across time via dynamic auction/inventory policies rather than wasteful queues.

Key Points

• Framework and primitives

  • Continuous-time M/M/1 baseline: buyers Poisson(λ), items/service completions Poisson(µ). Buyers value v, incur flow waiting cost c; items incur holding cost d to platform; firm profit per served item π.
  • Mechanisms: regenerative, non-anticipatory mappings from arrival histories to outcomes that induce a queue-state process θt ∈ Z (positive = buyers waiting, negative = items in inventory).
  • Focus on PRRMs: regenerative + positive recurrent → unique stationary distribution p over queue states; designer optimizes long-run averages.

• Methodological contribution

  • Transform the dynamic mechanism design into a static LP: choose stationary distribution p ∈ Δ(Z) maximizing an objective (weighted buyer welfare and platform/firm profit) subject to
    • Stationarity/flow (balance) constraints (Bk and B−ℓ) that relate arrivals and departures,
    • Participation/break-even constraints for buyers and firm,
    • Incentive compatibility (buyers must prefer recommended entry/exit behavior), and implementability by some PRRM.
  • Border-style reduced-form characterization adapted to dynamic queues allows feasible p to be described succinctly.

• Main structural results (NTU)

  • First-best (ignoring incentive constraints except participation) yields a finite interval support {−L, …, K} for p and balance constraints bind throughout interior states → implementable by cutoff/admission probabilities.
  • Cutoff/admission policies: optimal mechanisms often implement simple thresholds on when arrivals are admitted or items stored.
  • Queueing as screening: waiting can substitute for prices to screen types, but it creates social waste (waiting costs). Consequently, welfare-maximizing platforms may prefer pooling/random allocation (i.e., avoid screening via queues) to reduce wasted waiting.
  • Priority rules and information policy matter: admissions control, priority discipline (FCFS, LCFS, SIRO), and the information revealed to arrivals change incentives and performance. Under a full toolkit, FCFS tends to re-emerge as optimal in many settings (Che & Tercieux).

• Transferable Utility (TU) setting

  • With monetary transfers available, the central problem shifts to how to allocate competitive pressure intertemporally: when to sell now vs. keep buyers/items in queue/inventory for future matches.
  • The optimal dynamic auction (Che & Choi) characterizes how to dynamically control entry, allocation, and pricing to manage an inventory of buyers and goods across time. Transfers remove the need for wasteful waiting-based screening but introduce dynamic allocation and pricing tradeoffs.

• Incentives and information

  • Buyers form beliefs about the queue state from the induced stationary distribution; information disclosure (partial, full, or none) affects entry/abandonment behavior and thereby steady state.
  • Optimal mechanisms can be non-Markovian (depend on arrival order/history since last null time), because arrival order can be used instrumentally even under memoryless arrivals.

Data & Methods

• Type of work: theoretical survey synthesizing recent papers and methods (not empirical). Core sources include Che & Tercieux (forth.), Baccara/Lee/Yariv (2020), Madsen & Shmaya (2025), Che & Choi (2025), and classical queueing literature.
• Baseline model: analytical M/M/1 continuous-time queueing environment with Poisson arrivals and exponential service times. Buyers and items are homogeneous in base exposition; extensions permit heterogenous types for screening problems.
• Key mathematical tools
  • Queueing theory (regenerative processes, positive recurrence, steady-state distributions).
  • Linear programming (LP) over stationary distributions (relaxed/first-best and incentive-constrained versions).
  • Border-style reduced-form characterization for implementable stationary distributions.
  • Lagrangian/duality arguments to derive cutoff support structure and binding of balance constraints.
• Assumptions/limitations
  • Poisson/exponential/memoryless assumptions simplify analysis and make arrival-order instruments tractable; many results rely on steady-state ergodicity.
  • Survey focuses on single-server-type or aggregated single-queue representations; multi-dimensional or networked matching, non-Poisson arrival processes, and richer dynamic preferences require extensions.
  • Incentive analysis generally framed as Bayes–Nash equilibrium using stationary beliefs; stronger sequential-rationality refinements may be needed for detailed dynamic strategic behavior.

Implications for AI Economics

• Platform algorithm design
  • When designing AI-mediated allocation systems (ride-hailing dispatch, gig platforms, cloud marketplaces, blockchains), treat the long-run stationary distribution as the primary object: algorithmic policies should be evaluated by their induced steady-state tradeoffs between welfare, profit, and waiting/holding costs.
  • Simple, implementable rules (admission cutoffs, inventory thresholds, FCFS under full toolkit) can be optimal; thus, AI systems can aim for interpretable threshold policies learned or calibrated from estimated λ, µ, c, d, π.
• Screening vs. pricing
  • In environments where monetary transfers are constrained or undesirable, platforms (and their ML policies) will use waiting/priority as screening tools. AI designers must account for the social/welfare cost of induced waiting and may prefer randomized pooling or reduced screening to avoid waste.
  • If transfers are available, ML/AI should optimize dynamic pricing and allocation policies that balance present versus future competition rather than rely on queue-induced screening.
• Incentives, information disclosure, and belief modeling
  • Information released to agents (e.g., estimated wait times, queue length, priority signals) is a design lever with incentive consequences. AI systems that predict and disclose system state must incorporate strategic responses into learning and evaluation.
  • Stationary-belief reasoning is a useful approximation: agents who join a platform repeatedly will form steady-state beliefs—algorithms should be robust to such steady-state reasoning.
• Algorithmic/estimation implications
  • Estimation needs: λ and µ (and their heterogeneity and nonstationarity) are key inputs. Practical deployment requires methods to estimate arrival rates and waiting costs in real time and to adapt thresholds.
  • Reinforcement learning: the LP-over-stationary-distributions insight suggests value in RL formulations that target steady-state objectives and enforce flow/balance constraints (e.g., constrained MDPs, ergodic objectives) rather than only episodic reward maximization.
  • Non-Markovian policies can be optimal: ML architectures that can condition on order/history since last empty state (or use latent state summarization) may outperform purely Markov policies.
• Empirical predictions and testing
  • Predict cutoff/admission behavior and support bounds (finite queue-support) as signatures of optimal policies.
  • Expect tradeoffs: platforms that cannot use transfers will exhibit more pooling/random allocation and shorter optimal queues than transfer-enabled platforms.
  • FCFS priority should be observed more often when platforms have rich admission and information toolkits and welfare objectives; deviations may signal constraints or different objective weights.
• Directions for further AI-economics research
  • Extend to non-Poisson arrivals, heterogenous multi-attribute types, networked matching, multi-server systems.
  • Empirically estimate the welfare cost of queueing-based screening in real platforms and compare to pricing mechanisms.
  • Design and test RL algorithms that directly optimize the LP/steady-state objective under implementability constraints and incentive-aware agent models.
  • Study robustness and safety: how strategic manipulations of arrival behavior and misreporting interact with learned dynamic policies.

References (select) - Che, Y.-K., & Tercieux (forth.), Dynamic steady-state market design via stationary distributions. - Che, Y.-K., & Choi (2025), Dynamic auctions balancing inventory and entry over time. - Baccara, Lee, & Yariv (2020), etc.