Continuous ADAM: Double Auction Markets

• Continuous Artificial Double Auction Market is a simulated market using a limit order book where orders are continuously matched using price-time priority to generate transaction prices.
• It features diverse agent ecologies, from zero-intelligence noise traders to strategic market makers, highlighting practical insights into liquidity and market volatility.
• Analytical and computational models, including queueing theory, game theory, and formal verification, are used to optimize matching efficiency and validate market mechanisms.

Searching arXiv for recent and foundational papers on continuous double auctions, artificial markets, and verified CDA implementations. A Continuous Artificial Double Auction Market (ADAM) is an artificial or simulated market whose central exchange institution is a continuous double auction, typically implemented as a limit order book in which buy and sell orders arrive sequentially, interact with standing opposite-side liquidity, and generate transaction prices through immediate matching rather than batch clearing. In the research literature, ADAMs appear as formally specified matching engines, stochastic queueing models, latent-liquidity theories, and heterogeneous-agent simulations for finance, regulation, and electronic markets; across these strands, the common object is a continuously cleared double-sided market in which price formation is endogenous to order flow and matching rules (Garg et al., 2024, Donier et al., 2015, Mahfouz et al., 2021).

1. Institutional core and matching semantics

The canonical ADAM institution is a single-product continuous double auction that processes an online stream of instructions. In one formalization, the primitive commands are Buy w, Sell w, and Delete id, with each non-delete order represented as $w=(i,t,p,q)$, where $i$ is a unique order identifier, $t$ a unique timestamp, $p$ a limit price, and $q$ a maximum quantity. The exchange maintains resident bid and ask books $(B,A)$ and processes each command by an online transition

$$(B,A,\text{Command }w)\xmapsto{P}(\hat B,\hat A,M).$$

A buy is matched against earlier resident asks with ask price $\le p$; a sell is matched symmetrically against resident bids with bid price $\ge p$. The priority rule is price-time priority: competitiveness of price first, timestamp second. The incoming order is matched for the largest possible quantity, partial fills are allowed, unmatched residual quantity remains resident, and the post-state must satisfy positive bid-ask spread, so no executable pair remains in the book (Garg et al., 2024).

Artificial-market simulators usually enrich this core with exchange-operational detail rather than replacing it. A NASDAQ-like simulator built in ABIDES uses a price-then-FIFO matching engine, supports limit orders, market orders, and cancellation of outstanding limit orders, and runs as a discrete-event simulation with variable electronic network latency and agent computation delays. In that setting, the observable limit-order-book state used for learning is the z-score normalized ask and bid prices and volumes over the first five price levels (Mahfouz et al., 2021).

Application-specific CDA variants may impose stricter quoting rules. In a cloud-computing CDA, for example, a new bid had to satisfy $b'>Obid$, a new ask had to satisfy $i$0, and a match occurred whenever $i$1, with transaction price

$i$2

That market also reset the outstanding bid and ask at the beginning of each trading round, illustrating that ADAM implementations can differ materially in queue persistence, admissibility rules, and within-round state semantics even when they retain continuous double-auction logic (Shi et al., 2013).

2. Agent ecologies and application domains

ADAM research is as much about participant populations as about the auction mechanism. One prominent agent-based model combines zero-intelligence and limited-intelligence trader classes inside a CDA/LOB exchange. Its trader families are background agents, market makers, market takers, and directional traders; the concrete population includes 5000 noise agents, 100 value agents, 3 market makers, 3 TWAP agents, 3 VWAP agents, 5 momentum traders, and 5 mean-reversion traders. Noise agents submit random one-shot market orders, value agents compare current midpoint $i$3 with an estimated fundamental $i$4, market makers cancel old quotes and repost symmetric buy/sell orders around $i$5, and directional traders act on moving-average signals (Mahfouz et al., 2021).

Other ADAMs emphasize institutional traders. A leveraged-ETF market model contains $i$6 normal agents plus one leveraged ETF agent. Normal agents form expected returns from a weighted combination of fundamental, technical, and noise components,

$i$7

while the ETF tracks target leverage $i$8 through a rebalancing quantity $i$9 and submits market orders whenever $t$0. In that environment, increasing the minimum rebalancing threshold reduces the number of ETF trades, lowers total rebalancing volume, and decreases market volatility (Yagi et al., 2020).

Information heterogeneity is another recurrent design dimension. In one artificial stock market with a CDA mechanism, agents are informed, uninformed, switchers, or zero-intelligence traders. Informed traders set $t$1, whereas uninformed traders use

$t$2

Switchers decide whether to pay an information cost $t$3 to observe current fundamental value $t$4 or remain uninformed. The reported result is structurally two-sided: market volatility rises with the percentage of switchers in the population, but for a fixed switcher percentage volatility falls when a larger proportion of switchers actually purchases information at a given time (Liu et al., 2013).

Minimal ecologies remain important as benchmarks. Comparative simulations of ZI, ZIP, and GD traders in CDA-like markets found that the zero-intelligence model best reproduced several empirical-style statistical properties, including persistent large price fluctuations and heavy-tailed return distributions. This result is often read as evidence that continuous market interaction and decentralized matching can generate realistic aggregate structure even when agents are strategically primitive (Tseng et al., 2010).

3. Analytical theories of liquidity, queues, and equilibrium

A major analytical strand models ADAMs through latent supply and demand rather than explicit agent code. In a dynamic theory of supply and demand, the market is described by time-varying cumulative curves $t$5 and $t$6, with marginal supply and demand densities

$t$7

After an auction at price $t$8, executable liquidity is removed by truncation around the clearing price. In a moving frame $t$9, the large-population dynamics obey

$p$0

$p$1

As the inter-auction time $p$2, marginal liquidity vanishes linearly near the transaction price, cumulative supply and demand vanish quadratically, local liquidity collapses at the price, and impact crosses from linear to square-root, with

$p$3

This makes continuous clearing a source of endogenous hypersensitivity rather than a neutral approximation to Walrasian exchange (Donier et al., 2015).

A second analytical strand treats the CDA as a queueing system. In a simplified market with integer prices $p$4, unit-size orders, limit-order arrivals at rates $p$5, and market-order arrivals at rates $p$6, the total ask and bid queues are two independent $p$7 queues. Under symmetry,

$p$8

The model has three regimes: an ergodic regime $p$9 with mobile, diffusive prices; a non-ergodic but not fully trapped regime $q$0; and a strongly non-ergodic regime $q$1 in which the best quotes remain permanently populated and transaction prices collapse to a two-value random telegraph process (Radivojević et al., 2013).

In the low-traffic limit $q$2, the same stylized CDA admits an exact embedded price Markov chain because standing books are almost always empty and newly arriving limit orders are immediately transformed into market orders. The invariant price distribution can then be solved exactly on the finite price grid, and the approximation remains useful for stationary price distributions when $q$3 (Scalas et al., 2016).

A complementary game-theoretic line studies one-shot strategic shading. In a single-item CDA game with infinite buyers and sellers, traders precommit to bids $q$4 or asks $q$5 before the auction starts, and trade occurs when the standing maximum bid exceeds the standing minimum ask. For linear supply and demand, the model yields at most one Bayesian Nash equilibrium in the class of differentiable strictly increasing strategies. Competitive one-price behavior is not an equilibrium, but in the linear case total expected profit under the Bayesian Nash equilibrium equals the total expected profit under the competitive benchmark (Ruijgrok, 2012).

4. Learning, inference, and automated design

ADAMs are also used as laboratories for inverse problems. In one simulated CDA/LOB, opponent modelling is cast as supervised multiclass classification from market context and action: the input is the top-five-level book state together with order direction, price, and size, and the target is the trader archetype. The same environment supports behavioral cloning as supervised regression of order price and order size from order-book observations. Market makers and market takers are comparatively easier to identify than directional traders, which reflects the fact that some policies leave stronger state-conditioned signatures in order flow than others (Mahfouz et al., 2021).

Not all learning papers in this area use literal continuous clearing, but several are directly informative for ADAM. In a repeated double auction with two-sided bandit feedback and average pricing, buyers bid upper confidence bounds, sellers bid lower confidence bounds, and the resulting decentralized learning rule achieves social regret $q$6 while buyers and sellers that actually exchange goods incur $q$7 individual regret. The market there is synchronized and batch-cleared rather than continuous, so it functions more as a price-discovery benchmark than as a strict CDA model (Basu et al., 2022).

Within a genuine CDA simulator, empirical game-theoretic analysis has been combined with reinforcement learning to test whether candidate equilibria have profitable unilateral deviations in a larger policy space. In a market with 25 background traders and 1 market maker, reinforcement-learning agents interacting with the CDA as a POMDP found profitable deviations against arbitrary non-equilibrium profiles and in an expanded “NoFlip” action regime, but not against the reported EGTA equilibria in the more comparable “FlipKnown” regime. The interpretation offered is that the EGTA equilibria likely have only negligible regret with respect to the enlarged policy space examined (Wright, 2016).

Automated mechanism design treats continuity itself as a design variable. A grey-box framework decomposes a double auction into matching policy $q$8, quote policy $q$9, shout-accepting policy $(B,A)$0, clearing condition $(B,A)$1, pricing policy $(B,A)$2, and charging policy $(B,A)$3. In that grammar, continuous clearing is $(B,A)$4, batch clearing is $(B,A)$5, and probabilistic clearing $(B,A)$6 interpolates between them. The standard CDA benchmark is

$(B,A)$7

but the search procedure often discovers hybrid mechanisms with probabilistic clearing and alternative matching rules, showing that continuous clearing is one point in a broader design space rather than a universal optimum (Niu et al., 2010).

5. Formal verification, complexity, and executable correctness

The most rigorous implementation work in the area treats the CDA matching engine itself as a formally verified artifact. A verified continuous double-auction implementation specifies correctness through three properties: positive bid-ask spread after each instruction, price-time priority, and conservation of quantities under matching and deletion. An earlier result proves that any two online algorithms satisfying these properties produce identical matchings on every input sequence, so the specification is complete. The efficient verified implementation replaces an $(B,A)$8 list-based matcher with a red-black-tree-based routine eProcess_instruction running in $(B,A)$9 time on $$(B,A,\text{Command }w)\xmapsto{P}(\hat B,\hat A,M).$$0 instructions, and a matching lower bound shows that any continuous-double-auction implementation requires $$(B,A,\text{Command }w)\xmapsto{P}(\hat B,\hat A,M).$$1 time. The design uses four synchronized trees—$$(B,A,\text{Command }w)\xmapsto{P}(\hat B,\hat A,M).$$2, $$(B,A,\text{Command }w)\xmapsto{P}(\hat B,\hat A,M).$$3, $$(B,A,\text{Command }w)\xmapsto{P}(\hat B,\hat A,M).$$4, and $$(B,A,\text{Command }w)\xmapsto{P}(\hat B,\hat A,M).$$5—to support both best-price extraction and deletion by identifier in $$(B,A,\text{Command }w)\xmapsto{P}(\hat B,\hat A,M).$$6 time. The development is machine-checked in Coq, extracted to OCaml, and intended not only as an executable auction engine but also as an automatic checker for historical trade logs: replay the order stream, compute the unique legal match sequence, and compare it with the exchange’s reported trades (Garg et al., 2024).

6. Scope, misconceptions, and adjacent double-auction models

This body of work suggests that ADAM is best understood as a layered construct rather than a single model: a matching kernel, an agent ecology, and an evaluation criterion. A common misconception is to treat any “double auction” result as automatically applicable to continuous artificial markets. That is not the case. Some important adjacent models are explicitly batch-cleared or sealed-bid. A discrete-time centralized double auction for divisible assets proves that, under nonstrategic bidding, zero trade occurs precisely at Pareto-efficient allocations and repeated auctions converge toward individually rational Pareto allocations, but that mechanism is a repeated call auction rather than a CDA (Pennanen, 2020). Likewise, concurrent supply-chain auctions use protocol-generated synthetic bids and asks to coordinate linked markets, but their operation is deadline-based and batch-cleared rather than continuously matched (Babaioff et al., 2011).

A second misconception is that ADAM realism monotonically increases with trader sophistication. Comparative evidence does not support that as a general law: in one classic comparison, ZI traders matched several empirical-style heavy-tail and fluctuation properties better than ZIP or GD traders (Tseng et al., 2010). Conversely, mechanism-only work such as verified CDA matching does not by itself supply strategic interaction, market impact, multi-asset structure, or agent adaptation (Garg et al., 2024). Dynamic liquidity theory, in turn, explains square-root impact and vanishing local liquidity near price without providing a full exchange simulator (Donier et al., 2015). The encyclopedia consequence is that “Continuous Artificial Double Auction Market” names a research program organized around a common institution, not a single canonical architecture.