ABIDES: Towards High-Fidelity Multi-Agent Market Simulation

In one sentence

Summary

The paper addresses the lack of accessible, high-fidelity market simulation environments for academic research, noting that proprietary tools are unavailable and existing open-source platforms often lack the realism required for studying complex agent interactions. The authors introduce ABIDES, an open-source Agent-Based Interactive Discrete Event Simulation environment designed to support agent-based research in market applications. The core problem is enabling researchers to conduct controlled "what if" studies, such as analyzing market impact or co-location benefits, which are impossible with historical data alone due to the inability to isolate variables and observe individual agent intent. ABIDES aims to fill this gap by providing a platform where strategies can be endogenously chosen by learning agents interacting in a realistically structured market. The method involves a Python-based discrete event simulation kernel that models the NASDAQ ITCH and OUCH message-based protocols. Key features include nanosecond time resolution, configurable pairwise network latency with cubic jitter, and computation delays that affect message timing. The architecture consists of a simulation kernel, an exchange agent, and a hierarchy of agent classes (basic, exchange, trading). Experiments validate the environment using background agents (Zero Intelligence and Heuristic Belief Learning) to reproduce historical price movements and an "impact agent" to study market impact. The system supports deterministic execution via global virtual time and per-agent random number generators, ensuring reproducibility while allowing for A/B testing of specific agent changes. Findings demonstrate that ABIDES can accurately reproduce historical intra-day transaction histories using background agents and effectively model the mechanical and strategic market impact of large orders. The market impact case study shows that profit per share declines as trade size increases, with a correlation of r = -0.31 across trials. The platform also supports non-finance applications, such as simulating secure multiparty federated learning. Limitations include the current single-threaded nature of the kernel, which may limit scalability for extremely large populations, and the reliance on historical data for background agents, which may not capture all emergent behaviors. The tool is positioned as a foundational infrastructure for developing and evaluating ML trading algorithms and studying market microstructure.

Research problem

The paper addresses the scarcity of open-source, high-fidelity market simulation environments that allow for controlled experimentation with trading agents. Existing proprietary simulations are inaccessible, and prior open-source tools often lack the realism needed to study complex phenomena like market impact, co-location advantages, or the behavior of learning agents. The specific problem is enabling researchers to conduct controlled "what if" studies where variables can be isolated, and agent intent can be observed, which is impossible with historical data alone.

Core contributions

• Introduction of ABIDES, an open-source, high-fidelity agent-based discrete event simulation environment for equity markets.
• Implementation of a simulation kernel that replicates NASDAQ's ITCH and OUCH protocols with nanosecond time resolution.
• Design of configurable network latency and computation delay models to simulate realistic electronic trading conditions.
• Development of a deterministic execution framework using Global Virtual Time and per-agent PRNGs for reproducible A/B testing.
• Demonstration of the platform's utility through case studies on market impact modeling and historical market replay.

Method

• Discrete event simulation kernel with Global Virtual Time (GVT) for chronological event processing.
• Message-based architecture modeled after NASDAQ ITCH and OUCH protocols.
• Configurable pairwise network latency matrix with cubic jitter model.
• Per-agent computation delay tracking to simulate processing time effects on order submission.
• Deterministic random number generation using a global seed and per-agent PRNGs.
• Historical data injection via a "data oracle" for background agent behavior.
• Order book matching logic with partial execution and price-time priority.
• Event logging and serialization for detailed post-simulation analysis.

Results & findings

• ABIDES can simulate tens of thousands of trading agents interacting with an exchange agent.
• The simulator processes over 10,000 events per second on a standard laptop.
• Background agents (ZI and HBL) can reproduce historical price histories with statistical similarity.
• Market impact experiments show profit per share declines as trade size increases (r = -0.31).
• The impact trader's total profit increased from $2,633 to $12,502 as bid size grew, despite lower profit per share.
• The platform supports deterministic A/B testing by isolating stochastic perturbations.
• Computation delays and network latency are modeled to affect order timing realistically.

Data & experiments

• Simulation of historical equity trading data (e.g., IBM, MSFT) using background agents to reproduce intra-day price movements.
• Market impact experiments using an "impact agent" placing large market orders to measure price displacement and profit per share.
• Comparison of simulated trades against historical trades to validate the realism of background agent populations.
• Evaluation of network latency and jitter models through co-location experiments.
• Non-finance validation via simulation of secure multiparty federated learning protocols.

Limitations & caveats

• The simulation kernel is currently single-threaded, which may limit scalability for very large agent populations.
• Background agents rely on historical data or simple heuristics, which may not capture all emergent market behaviors.
• The platform does not yet implement parallel distributed simulation at the logical process level.
• Market replay simulations may not fully replicate the complexity of real-world market dynamics.
• The open-source nature means it may lack the proprietary optimizations of institutional-grade simulators.

Curator notes

Discovery status: institutional_trading_round2_promoted. Source links: StockSim: A Dual-Mode Order-Level Simulator for Evaluating Multi-Agent LLMs in Financial Markets