Preamble: In the spirit of open inquiry and intellectual candour, what follows is Simudyne’s considered perspective on the application of agent-based simulation in capital markets—a perspective shaped by years of toil, reflection and practical engagement, but one that neither claims omniscience nor presumes consensus.
Legal Disclaimer: This document reflects Simudyne’s perspective on agent-based simulation (ABM) in capital markets as of April 2025. It represents our best current understanding and interpretation of the subject, informed by our experience and expertise, but does not claim to offer a comprehensive, definitive, or peer-reviewed account of the field. References to patent rights herein include granted U.S. patents, allowed claims pending formal grant, and pending continuation claims currently under examination. Specifically, Simudyne’s Synchronisation patent (WO2020188373A1) is protected under allowed claims 1–9, limited to trading system simulation. Broader claims relating to non-financial applications remain pending and, at present, are neither granted nor allowed. Any discussion of potential applications of the Synchronisation technology outside trading systems refers solely to the subject matter of pending claims and does not imply existing patent protection in those domains. This document is provided for informational purposes only and does not constitute legal advice. Patent status may evolve based on the outcome of ongoing applications and administrative proceedings. Simudyne makes no representation or warranty as to the enforceability, scope, or eventual outcome of any pending claims. Readers are encouraged to seek independent advice and conduct their own due diligence before relying on the information contained herein for commercial, academic, regulatory, or investment purposes.
Caveat emptor: the views expressed herein reflect our perspective at a point in time and should not be construed as definitive nor exhaustive.
Introduction
Over the past fifteen years, we at Simudyne have witnessed and participated in the rising prominence of agent-based models (ABMs) as a foundational tool for simulating complex dynamics in capital markets. In the last five years especially, peer-reviewed studies, institutional reports, and regulatory initiatives have increasingly advocated for agent-based simulations to complement—and in some cases challenge—traditional financial models. This shift reflects an emerging consensus that markets are complex adaptive systems populated by heterogeneous participants, whose interactions under stress or regime change can defy static or purely empirical modelling [1]. The failures of equilibrium-based models to anticipate systemic events such as the 2008 global financial crisis and the 2010 ‘Flash Crash’ underscored the need for new modelling paradigms. In response, academics, regulators, and practitioners—including ourselves—have turned to agent-based approaches to capture emergent phenomena and market malfunctions, test policy interventions, and train AI agents in silico [2, 3].
At Simudyne, we have invested more than a decade operationalising these principles. Our simulation platforms—Pulse and Horizon—have been deployed in production settings at global market infrastructures, including Hong Kong Exchanges and Clearing (HKEX) and London Stock Exchange Group (LSEG). Through this work, we have transformed ABM from a research curiosity to an institutional-grade, real-time decision infrastructure [4].
Recent scholarship has reinforced this paradigm shift. Axtell and Farmer (2025) argue that agent-based modelling should be recognised not merely as a technique, but as a computationally enabled branch of economics and finance, capable of addressing complexity and emergent phenomena that traditional models cannot [5]. Their analysis situates ABMs alongside simulation-based methods used across the natural sciences, highlighting that financial markets—like ecosystems and weather systems—are best understood through dynamic, agent-level simulation rather than closed-form equations.
Advantages of Agent-Based Models vs. Traditional Approaches
We have seen first-hand that traditional financial models—whether econometric, equilibrium-based, or data-driven—often struggle to account for unprecedented scenarios and adaptive behaviours. These models typically assume representative agents and rely on historical correlations or closed-form equilibria. Such simplifications fail when confronted with out-of-sample events, structural breaks, or novel technologies [6]. By contrast, agent-based simulations enable us to construct dynamic, heterogeneous, and granular models of financial markets that are explicitly designed to test “what-if” scenarios outside the historical record [7].
In our own technology, we have augmented this flexibility with patented innovations. For example, our Synchronisation interface [8] allows real trading systems and AI algorithms to plug into live simulations without modification, facilitating the evaluation of decision engines under stress conditions that could never be tested safely in production environments [4, 8]. We have also pioneered Deterministic Execution [9], ensuring that large-scale, distributed simulations of markets can be replayed exactly, providing an auditable basis for regulatory stress testing and AI algorithm certification [4, 9].
Agent-based simulations uniquely capture the dynamic interaction and feedback loops inherent in financial systems. As Bookstaber et al. (2018) observe, “a static representation may be adequate for some tasks, but understanding the propagation of shocks through the financial system requires a dynamic approach” [10]. Simudyne’s Deterministic Execution patent ensures that such dynamic simulations are fully reproducible and auditable even when executed at scale across distributed systems [4, 9].
As Axtell and Farmer emphasise, one of the defining features of ABMs is their capacity to model how micro-level interactions among heterogeneous agents generate macro-level market phenomena—something that static, representative-agent models structurally exclude [5]. This micro-to-macro emergence is not an incidental feature but the very rationale for agent-based simulation in economics and finance.
ABMs incorporate heterogeneity of agents, which is crucial to reproduce real-world phenomena like crashes or contagion cascades [11, 12]. The Bank of England has observed that ABMs provide a “flexible framework” to evaluate both aggregate and distributional effects of policies [12]. Our research on liquidity risk modelling [13] and stylised fact reproduction [14] directly demonstrates this capacity.
This approach to empirical validation aligns with Axtell and Farmer’s assessment that ABMs gained legitimacy in finance precisely because of their ability to reproduce well-documented stylised facts—such as fat-tailed returns and volatility clustering—without imposing them by assumption [5]. Our own validation efforts build upon this empirical tradition to reproduce these stylised facts, while incorporating recent refinements, such as those by Ratliff-Crain et al. (2023), that assess which stylised facts remain robust in contemporary markets [15]. This finding underscores the importance of focusing our model calibration and validation efforts on the stylised facts that remain empirically robust and relevant for modern capital markets. It also affirms our commitment to maintaining the empirical relevance of our simulation platforms as markets evolve.
Critically, ABMs transcend the rationality assumption of classical financial theory. By embedding heuristics, behavioural biases, and learning rules, we can simulate the messy reality of market decision-making. Although this flexibility introduces calibration challenges, our research and practical deployments have shown that the explanatory and predictive power gained far outweighs these difficulties [15, 16]. Through advanced calibration techniques, including data-driven agent profiling, we have addressed many of these concerns [17].
Market Microstructure Modelling and Liquidity Dynamics
One of the most significant applications of our simulation technology is in modelling market microstructure—the granular processes by which trades are executed, liquidity is provided or withdrawn, and prices are formed. Over the past decade, researchers and practitioners have increasingly turned to ABMs to model trading at a micro level to reproduce the statistical “stylised facts” of market behaviour: fat-tailed return distributions, clustered volatility, autocorrelation of returns, and long memory effects [18]. Our own working paper on stylised facts on both the Turquoise and HKEX venues confirms that our simulation stack replicates these emergent patterns with high fidelity [13, 14].
We have also demonstrated that agent-based simulation is uniquely suited to analysing flash crashes and liquidity crises, which defy prediction by traditional models. Our detailed simulations of the 2010 Flash Crash and similar high-frequency events reveal how liquidity can evaporate endogenously through interactions of algorithmic agents—particularly when feedback loops, adverse selection, and defensive herding behaviours amplify initial exogenous shocks [19].
Moreover, our platform has been used to evaluate the impact of specific market regulations—such as price variation limits, short-selling bans, and tick size adjustments—in a safe, controlled environment [20]. By leveraging our Synchronisation interface [8], we have enabled exchanges and regulators to test these interventions in accelerated time and under extreme market conditions without capital at risk or operational disruption [4].
We have further operationalised Interactive Agent-Based Simulation (IABS), allowing trading strategies to be evaluated in dynamic environments where other agents adapt in response to the strategy under test. Our work on “Delivering Algo Performance” exemplifies how institutions can use this approach to benchmark and stress test their trading algorithms against emergent systemic effects, rather than relying on static backtests [21].
Systemic Risk, Liquidity Crises, and Contagion Modelling
Beyond the microstructure level, we have long recognised that agent-based simulation provides an essential framework for systemic risk analysis. The financial crises of the past two decades have underscored the need for more advanced modelling approaches to address contagion, liquidity spirals, and network effects. As Bookstaber et al. (2014) observe, “Risk management methods, most notably Value-at-Risk (VaR), are based on historical data and are simply not designed to work in a financial crisis.” They further explain that while stress tests represent an improvement by decoupling future crisis assessments from historical data, they still fall short in a critical regard. Stress testing fails to account for the dynamics, feedback, and complexities inherent in a financial crisis. The authors contend that understanding systemic risk requires a fundamental re-evaluation of models to encompass the internal workings of the financial system, such as crowded trades, asymmetric information, liquidity shortages, and interconnectedness. They argue that “Agent-based models are well suited to deal with the issues of crisis dynamics and feedback. These models track the actions of agents, assessing their reactions to events on a period-by-period basis, and updating system variables accordingly. Agent-based models incorporate heterogeneity and allow agents to operate according to idiosyncratic, and potentially less-than-optimal, rules governing financial institutions.” As such, agent-based models allow us to simulate the interconnected behaviour of banks, asset managers, funds, and infrastructures at the granularity needed to trace shock propagation paths.
This view is echoed by Axtell and Farmer, who write that systemic financial crises involve an “ecology of purposive interacting agents” and that “systemic risk is an emergent phenomenon that comes about due to the nonlinear interaction of individual agents,” making them inherently unsuited to analysis by traditional risk models. They explicitly advocate to think of systemic risk “as a complex system and [to] study it using agent-based modeling” [5].
Our work with HKEX and other institutions has validated how liquidity stresses can emerge endogenously within simulated limit order books and propagate through market networks, producing solvency and funding contagion effects [13]. These simulations consistently reveal how liquidity and solvency risks interact non-linearly and discontinuously, leading to systemic fragility [24, 39]. In many cases, agent-based simulation results have directly informed regulatory policy evaluations, including liquidity coverage ratio assessments and central bank intervention strategies [25, 26].
Theoretical Rationale for Multi-Agent Environments
Our advocacy for agent-based simulation is grounded not only in empirical success but also in theoretical rigour. Capital markets are quintessential complex adaptive systems. They cannot be reduced to static equilibria or representative agent models [27]. Instead, markets are emergent phenomena arising from the interaction of heterogeneous, boundedly rational, and adaptive agents operating under uncertainty [28]. Our entire technology stack, including our patented scalability and determinism infrastructure, is designed to model precisely these dynamics at scale [4, 8, 9, 40].
ABMs provide structural and causal insights unavailable in traditional data-driven models. Because we encode explicit behavioural rules and interactions, we can run controlled experiments—altering one rule or agent type and observing the systemic consequences [29]. Moreover, agent-based simulations offer out-of-sample robustness: the structural assumptions about behaviour and interaction hold even under novel conditions, unlike machine learning models that rely on past data correlations [30]. Our deterministic execution technology ensures that such structural experiments are reproducible and explainable to regulators and institutional stakeholders [9].
Emerging Trend: Cognitive Agents and AI in Simulated Markets
A particularly significant development we are pioneering is the integration of advanced AI techniques—most notably large language models (LLMs), reinforcement learning (RL), and hybrid cognitive architectures—into agent-based simulations of financial markets. This evolution reflects not only the growing computational feasibility of such integrations but also our recognition that the behavioural complexity of real markets demands cognitively richer agents.
Historically, agent-based financial simulations employed relatively simple behavioural heuristics—momentum trading, fundamental valuation, noise strategies—encoded in rule-based agents. While such abstractions succeeded in reproducing certain emergent phenomena, they fell short of capturing the cognitive, social, and adaptive dimensions of real-world market participants. The inclusion of AI-driven agents addresses this limitation by allowing agents to learn, adapt, and reason in a manner closer to human traders, institutional investors, or algorithmic trading systems operating under uncertainty.
Recent research has demonstrated that large language models, trained on vast corpora of financial news, trading dialogues, and general knowledge, can approximate human-like decision-making in market contexts. For instance, Koa et al. (2024) introduced a Massively Multi-Agent Role-Playing simulation in which each market participant was instantiated as an LLM-powered agent, conditioned on specific trading roles and risk preferences. Intriguingly, the study observed that market equilibria emerging from the interaction of these cognitive agents exhibited stylised facts previously achieved only by carefully hand-coded ABM rules [31].
This shift from predictive to reasoning-based approaches aligns with Jagdish’s (2024) framework distinguishing between “System 1” and “System 2” in AI systems. As Jagdish argues, while traditional AI models excel at pattern recognition and statistical forecasting (System 1), they often struggle with the kind of counterfactual reasoning and scenario planning that human market participants routinely perform (System 2). By integrating LLMs with simulation capabilities, we create agents that don’t merely forecast prices based on historical patterns but actively ‘think through’ complex market scenarios, considering multiple potential futures and adapting their strategies accordingly. This distinction between prediction and simulation-powered thinking is particularly crucial in financial markets, where non-stationarity and strategic interactions create environments where historical patterns frequently break down [38].
In our own work, we have advanced this frontier through research on neuro-symbolic traders—agents whose behaviour combines structured symbolic reasoning with neural learning mechanisms. Our study, “Neuro-Symbolic Traders: Assessing the Wisdom of AI Crowds in Markets,” evaluated how these agents, when embedded in interactive simulations, impacted market behaviour. We observed that these agents tended to suppress price fluctuations and reduce volatility, highlighting potential risks to market stability due to collective overconfidence and diminished diversity in trading behaviour [32].
We have also incorporated reinforcement learning agents into our simulations. Unlike static rule-based agents, RL agents continuously update their strategies through interaction with the simulated market environment, effectively learning to exploit, stabilise, or destabilise market conditions based on reward structures. This dynamic adaptation opens the door to studying emergent phenomena such as strategic manipulation, coordination failure, or adversarial dynamics between competing algorithms—risk scenarios regulators are increasingly concerned about in the context of AI-driven markets. Furthermore, agent-based simulators provide a powerful tool for optimising policies and protocols, enabling the identification of strategies that enhance market stability and robustness against manipulation, offering a forward-looking approach to safeguarding market integrity in the age of AI.
Our research on deep hedging provides a concrete example of this hybrid AI-ABM approach. In our paper “Deeper Hedging: An Agent-Based Model for Effective Deep Hedging”, we integrate a deep reinforcement learning agent into an agent-based market model, showing how AI agents can autonomously learn novel and robust hedging strategies in environments characterised by stochastic volatility and transaction costs [33]. Notably, these RL agents co-evolve with heterogeneous market participants in the simulation, exposing not only the performance but also the systemic effects of AI-driven strategies under different market conditions.
The theoretical justification for incorporating cognitive agents lies in complexity economics and behavioural finance. Markets are not merely aggregations of optimising agents operating under rational expectations; they are social, informational, and feedback-driven ecosystems, in which heterogeneous actors interpret, misinterpret, and adapt to one another’s behaviour in real time. LLM-powered and RL-based agents allow us to encode bounded rationality, learning dynamics, and social feedback loops within the simulation environment itself, moving beyond the static, equilibrium-oriented models of classical finance.
In the evolving field of financial market simulations, it is becoming increasingly evident that authentic market dynamics do not arise merely from agents engaged in statistical guesswork about future prices. Rather, they emerge from agents endowed with computational processes that mirror human reasoning—agents who do not simply predict, but who construe. These agents construct internal models of other participants, entertain counterfactual scenarios, and revise their mental models as new information collides with old beliefs. By embedding these cognitive faculties within our simulated agents—and situating them within what Jagdish describes as ‘world models’—we enable them to rehearse actions, anticipate consequences, and engage in reflective decision-making before executing trades. This approach builds upon the intellectual foundation laid by Yann LeCun, who contends, rightly, that the attainment of human-level artificial intelligence will demand architectures capable of modelling the world and learning through interaction, transcending the narrow confines of the text-prediction paradigm. Jagdish echoes this insight, recognising that such cognitively enriched agents are not merely statistical calculators but participants in a living, evolving simulation of market behaviour [38].
Yet this approach is not without challenges. First, cognitive agents require careful calibration to ensure behavioural plausibility without introducing artefacts. Second, the computational burden of large-scale cognitive simulation is non-trivial, demanding efficient synchronisation and parallel execution architectures—precisely the technical barriers we have addressed through our patented scalability and determinism technologies [4, 8, 9]. Third, and perhaps most important, cognitive agents introduce new dimensions of emergent systemic risk: the interaction of adaptive AI agents can lead to opaque, nonlinear behaviours that neither their designers nor regulators fully anticipate. For this reason, leading institutions and regulatory bodies have begun advocating for controlled ABM environments as safe arenas in which to rehearse and analyse AI-agent dynamics before their deployment in live markets [34].
Drawing on Jagdish’s framework, we can conceptualize our simulation environments as computational thinking laboratories where AI-driven trading strategies undergo rigorous evaluation not merely against historical data, but within dynamic, adaptive ecosystems that challenge their fundamental assumptions. Simulation encourages exploration of unknown unknowns, making it particularly valuable for stress-testing strategies against scenarios that lie outside the historical record—including the black swan events that pose the greatest threat to financial stability [38].
These efforts are converging on what we and others term cognitive realism: the design of agent-based market models that do not merely mimic transactional behaviour, but embed human-like cognition, social learning, and bounded rationality into the very fabric of simulation. Our ongoing research, including synthetic data generation using cognitive agents [17] and experimental evaluation of market dynamics under LLM-based trading populations [31], stands at the forefront of this emerging paradigm.
Institutional and Regulatory Perspectives
We are not alone in recognising the strategic importance of agent-based simulation. Central banks, financial regulators, and leading market infrastructures have increasingly adopted ABMs to complement their existing modelling toolkits. Institutions such as the Bank of England, European Central Bank, and U.S. Office of Financial Research have integrated ABMs into their systemic risk oversight and macroprudential policy analysis [35, 36].
Regulatory bodies are also exploring the use of simulation sandboxes to evaluate AI trading systems, test rule changes, and rehearse systemic crises before they manifest in live markets [34]. Our own collaborations with HKEX, LSEG, and others exemplify how agent-based simulation is being operationalised as part of regulatory infrastructure. Simudyne is working directly with these institutions to provide scalable, explainable, and reproducible simulation environments based on our patented technology stack [4].
Further, regulatory discussions now explicitly acknowledge the systemic risks posed by AI-driven trading. Reports from the Financial Stability Board and Basel Committee reference the potential for emergent, nonlinear dynamics arising from algorithmic trading systems—a risk uniquely suited to analysis in controlled, agent-based environments [36]. Our simulation work has already been used in regulatory sandboxes to evaluate algorithmic trading behaviour, systemic risk contagion, and policy interventions.
The MITRE Corporation’s recent research further supports this trend, emphasising the necessity of simulation environments for the safe and rigorous evaluation of AI agents’ emergent behaviours and potential failure modes—an approach that, in complex socio-technical systems, often necessitates multi-agent simulation frameworks and scalable, auditable infrastructure [37]. Importantly, Axtell and Farmer also note that the historical barriers to institutional adoption of ABM—chiefly computational cost and data availability—have largely fallen away. The conditions that previously confined ABMs to academic research now favour their operational deployment by regulators, exchanges, and market participants [5]. We believe this convergence of regulatory, academic, and technological perspectives marks the beginning of a new standard: that no advanced trading algorithm or policy change should be deployed in live markets without first being evaluated in a high-fidelity agent-based simulation that reacts to the impact of said algorithm or policy change.
Conclusion
Our experience over the past decade has convinced us that agent-based modelling is no longer an experimental technique confined to academic circles. It has become essential infrastructure for capital markets research, risk oversight, and AI strategy development. The literature reviewed here, combined with our own intellectual property, operational experience, and regulatory collaborations, demonstrates that ABMs uniquely capture emergent phenomena, systemic feedbacks, and dynamic interactions that elude traditional models.
The integration of cognitive agents and advanced AI techniques into agent-based environments marks the next frontier. We believe the convergence of AI and ABM will define how financial markets are analysed, regulated, and understood in the coming decade. Our work at Simudyne reflects a singular conviction: that in a world of accelerating technological change and systemic uncertainty, simulation is no longer optional. It is the proving ground where tomorrow’s financial systems must first be built, tested, and secured.
As we look ahead, our commitment is clear. We will continue to invest in the advancement of agent-based simulation technology, the integration of cognitive and adaptive agents, and the creation of safe, ethical, and scientifically sound environments in which the future of finance can be rehearsed before it is lived.
This trajectory reflects the broader academic consensus. Axtell and Farmer (2025) conclude that agent-based modelling should become a standard component of the financial modelling toolkit, alongside econometrics and DSGE models [5]. Our work operationalises this recommendation, bringing ABM from academic proof-of-concept to production-grade simulation infrastructure.
References
[1] Mizuta, K. (2021). “Agent-Based Artificial Markets and Financial Market Regulation: A Review.” SSRN.
[2] Bookstaber, R., Paddrik, M., & Tivnan, B. (2014). “An Agent-Based Model for Financial Vulnerability.” Office of Financial Research Working Paper No. 14-03.
[3] Bookstaber, R. (2014). “Stress Tests to Promote Financial Stability: Assessing Progress and Looking to the Future.” Office of Financial Research Working Paper.
[4] Simudyne Ltd. (2020). “Simulation Platforms for Capital Markets: Technical Overview.” Simudyne White Paper.
[5] Axtell, R., & Farmer, J.D., “Agent-Based Modeling in Economics and Finance: Past, Present, and Future.” Journal of Economic Literature 2025, 63(1), 197–287.
[6] Farmer, J.D. (2012). “A Complex Systems Approach to Constructing Better Models for Managing Financial Markets and the Economy.” The European Physical Journal Special Topics.
[7] Tesfatsion, L. (2006). “Agent-Based Computational Economics: A Constructive Approach to Economic Theory.” Handbook of Computational Economics, Volume 2.
[8] Simudyne Ltd. (2020). “WO2020188373A1 Synchronisation Interface for Live Decision Systems in Simulation.” Patent application.
[9] Simudyne Ltd. (2023). “US Patent 11,733,971 B2: Deterministic Execution in Distributed Simulation Environments.” Patent.
[10] Bookstaber, R., Paddrik, M., & Tivnan, B. (2018). “An Agent-Based Model for Crisis Liquidity Dynamics.” Office of Financial Research Working Paper No. 18-04.
[11] LeBaron, B. (2006). “Agent-Based Computational Finance.” Handbook of Computational Economics, Volume 2.
[12] Bank of England. (2025). Agent-based modelling at central banks: Recent developments and new challenges (Working Paper No. 1050).
[13] Simudyne Ltd. (2024). “Agent-Based Liquidity Risk Modelling for Financial Markets.” Working Paper (in preparation for peer-review publication).
[14] Simudyne Ltd. (2023). “Reproducing Market Stylized Facts Using Agent-Based Simulation.” Working Paper (in preparation for peer-review publication).
[15] Ratliff-Crain, E., Koehler, M.T.K., Tivnan, B.F., & Bagrow, J.P. (2023). “Revisiting Stylized Facts for Modern Stock Markets.” IEEE International Conference on Big Data.
[16] Delli Gatti, D., Desiderio, S., Gaffeo, E., Cirillo, P., & Gallegati, M. (2011). “Macroeconomics from the Bottom-up.” Springer.
[17] Simudyne Ltd. & Refinitiv. (2019). “Synthetic Reality: Synthetic Market Data Generation at Scale Using Agent-Based Modelling.” Refinitiv White Paper.
[18] Cont, R. (2001). “Empirical Properties of Asset Returns: Stylized Facts and Statistical Issues.” Quantitative Finance, 1(2).
[19] Gao, K., Vytelingum, P., Weston, S., Luk, W., & Guo, C. (2022). High-Frequency Financial Market Simulation and Flash Crash Scenarios Analysis: An Agent-Based Modelling Approach. arXiv:2208.13654v1 [q-fin.TR].
[20] Mizuta, T. (2019). “An Analysis of the Effect on Market Price of the Parameters in Artificial Markets: An Agent-Based Approach.” Computational Economics.
[21] Balch, T., et al. (2019). “Using Interactive Agent-Based Simulation to Evaluate Trading Strategies.” Proceedings of the ACM Conference on AI in Finance (ICAIF ‘19).
[22] Farmer, J.D., & Foley, D. (2009). “The Economy Needs Agent-Based Modelling.” Nature, 460.
[23] Bookstaber, R., Paddrik, M., & Tivnan, B. (2018). “An Agent-Based Model for Financial Vulnerability.” Journal of Economic Interaction and Coordination, 13(2).
[24] Bookstaber, R., & Kenett, D.Y. (2016). “Looking Deeper, Seeing More: A Multilayer Map of the Financial System.” Office of Financial Research (OFR) Brief Series, 16-06.
[25] Bank of England. (2019). “Simulating Liquidity Stress in the Derivatives Market.” Staff Working Paper No. 838.
[26] European Central Bank. (2018). “Assessing the Procyclicality of CCP Margins with a Simulation-Based Approach.” Working Paper Series No. 2112.
[27] Arthur, W.B. (1999). “Complexity and the Economy.” Science, 284(5411).
[28] Farmer, J.D., & Foley, D. (2009). “The Economy Needs Agent-Based Modelling.” Nature, 460.
[29] Lux, T., & Marchesi, M. (1999). “Scaling and Criticality in a Stochastic Multi-Agent Model of a Financial Market.” Nature, 397.
[30] Gualdi, S., Tarzia, M., Zamponi, F., & Bouchaud, J.-P. (2015). “Tipping Points in Macroeconomic Agent-Based Models.” Journal of Economic Dynamics and Control, 50.
[31] Koa, J., Wang, X., Kumar, V., & Singh, S. (2024). “Massively Multi-Agent Role-Playing Market Simulation with LLM-based Agents.” OpenReview Preprint.
[32] Stillman, N.R., Baggott, R., Lyon, J., Vytelingum, P., Zhang, J., Zhu, D., & Chen, T. (2024). “Neuro-Symbolic Traders: Assessing the Wisdom of AI Crowds in Markets.” Working Paper (in preparation for peer-review publication).
[33] Gao, K., Weston, S., Vytelingum, P., Stillman, N., Luk, W., & Guo, C. (2023). “Deeper Hedging: An Agent-Based Model for Effective Deep Hedging.” Proceedings of the Fourth ACM International Conference on AI in Finance, 270-278.
[34] Sidley Austin LLP. (2024). “Emergent Behaviour in AI-Driven Trading: Legal and Compliance Implications.” Law Review Article.
[35] Bank of England. (2025). “Agent-Based Modeling at Central Banks: Recent Developments and New Challenges.” Bank of England Staff Working Paper No. 1122.
[36] Financial Stability Board. (2024). “Evaluation of the Impact and Efficacy of the Basel III Reforms.” Basel Committee on Banking Supervision Publication No. 544.
[37] MITRE Corporation. (2024). “A Repeatable Process for Assuring AI-Enabled Systems.” MITRE Technical Report PR-24-1768.
[38] Jagdish, N. (2024). “From Prediction to Thinking: Enhancing LLMs with Simulation.” LinkedIn Pulse. https://www.linkedin.com/pulse/from-prediction-thinking-enhancing-llms-simulation-ninad-jagdish-rlk2c/
[39] Wise, R., Chen, T. & Zhu, D. (2024). “An innovative approach for optimising CCP default management through agent-based modelling.” Journal of Risk Management in Financial Institutions, 17(3), 286-293.
[40] Simudyne, Ltd. (2022). Method and system for efficient multi agent computer simulation (U.S. Patent No. 11,316,747). U.S. Patent and Trademark Office.