limit order book modeling

Existing limit order book simulators struggle to accurately reproduce realistic execution costs, profit-and-loss outcomes, and microstructural dynamics. This work proposes an interactive simulator tailored for large-tick assets, which projects the order book state into a low-dimensional representation based on spread and volume imbalance, calibrates event timing to match the temporal structure of real markets, and incorporates a signed trade-flow feedback mechanism governed by a power-law decay kernel to capture market impact and its partial reversal. The proposed “project–estimate–validate–adapt” four-step framework is the first to simultaneously reproduce concave market impact, post-trade price reversion, and sensitivity to round-trip exchange latency within simulation. Experiments demonstrate that the approach generates highly realistic execution behavior across multiple stocks and trading strategies, significantly narrowing the gap between simulated and real-world market dynamics.

This paper addresses the poor adaptability of conventional market-making strategies in non-stationary limit-order-book environments. Methodologically, we design a high-fidelity simulation environment incorporating empirically observed market phenomena—including clustered order arrivals, time-varying bid-ask spreads, return drifts, stochastic order sizes, and price volatility—and model order flow and price dynamics using stochastic control theory. We train a robust market-making agent via Proximal Policy Optimization (PPO). Our key contribution lies in explicitly encoding multiple market anomalies into the learning framework to enhance policy generalization under non-stationarity. Experimental results demonstrate that the proposed strategy significantly outperforms classical analytical solutions in profit stability, risk mitigation, and cross-market adaptability. The simulation platform proves effective for both training and pretraining RL-based market-making agents, validating its practical deployment potential.

This study explains the emergence of persistent market disequilibrium and price bubbles without invoking irrationality or optimistic expectations. By constructing a dynamic equilibrium model with heterogeneous agents holding noisy valuations, the authors model supply and demand as coupled stochastic processes and uncover a novel mechanism whereby exponential price growth arises solely from zero-mean valuation errors and biases in order statistics. The framework extends Miller’s disagreement theory to a dynamic setting and unifies Walrasian equilibrium and static risk premia as special cases. Integrating agent-based modeling, order statistics, and stochastic simulation, the model reproduces six distinct market regimes—including stable bands and bubble bursts—under a single behavioral rule, offering a unified account of empirical disagreement phenomena and highlighting how machine learning–based valuation algorithms may inadvertently amplify such statistical biases.

This study investigates how large language model–driven AI traders form expectations and influence asset market bubbles. By constructing a multi-agent simulation system within an open-outcry market environment, the research simulates trading behaviors and, for the first time, leverages interpretable textual reasoning to uncover the underlying decision-making mechanisms of AI agents. Through prompt engineering, the study implements causal interventions targeting specific behavioral biases. The framework successfully replicates established experimental market phenomena—such as the predictive power of excess demand on price dynamics and the positive correlation between belief dispersion and trading volume—and demonstrates effective control over bubble intensity. These findings offer a novel methodological approach and empirical evidence for understanding market dynamics in settings where AI agents actively participate.

Modeling heterogeneous trader behavior in stock markets remains challenging due to the complexity and diversity of agent-level decision-making processes. Method: We propose an improved Generative State Space Model (GSSM) that accurately reproduces sequential decision-making across multiple agent types. The model is pre-trained on heterogeneous agent-based synthetic data and fine-tuned and validated on real limit-order-book data annotated with investor identifiers. Contribution/Results: This work introduces, for the first time, a state-space architecture into generative trading-behavior modeling, explicitly capturing individual strategic heterogeneity. Quantitative evaluation shows that predicted distributions of key behavioral metrics—including order submission frequency, price deviation, and position holding duration—closely match those of ground-truth agent models (Kolmogorov–Smirnov test p > 0.92), significantly outperforming LSTM and Transformer baselines. The framework effectively approximates micro-level trading mechanisms in complex markets, offering a novel, interpretable paradigm for market simulation and regulatory sandbox applications.

This study investigates how large language model (LLM) agents strategically exploit reputation mechanisms to engage in deceptive behavior in information-asymmetric e-commerce markets, and examines the mitigating role of governance mechanisms. To this end, we introduce TruthMarketTwin, a novel simulation framework that, for the first time, integrates LLM agents into a complex e-commerce environment featuring bilateral transactions, rating systems, and dispute resolution mechanisms, enabling systematic modeling of their strategic interactions. Experimental results demonstrate that, in unregulated markets, LLM agents autonomously exploit vulnerabilities in reputation systems to deceive; however, the introduction of an escrow-based enforcement mechanism significantly suppresses such deceptive strategies and steers agents toward more rational and compliant reasoning. These findings validate the efficacy of mechanism design in effectively regulating LLM agent behavior.

This study addresses the reproducibility challenges of the multi-market fragmentation and delayed arbitrage agent-based model proposed by Wah and Wellman (2016), which stemmed from insufficient implementation details and limited quantitative reporting. Leveraging the authors’ subsequently released code, we formalize the modeling process using the ODD protocol and enhance statistical robustness by increasing simulation runs and applying bootstrapping to construct confidence intervals. Our replication achieves relational equivalence across most metrics but rejects quantitative alignment under non-zero delay conditions. Notably, we uncover that conclusions regarding fragmentation effects are highly sensitive to the specific implementation of greedy strategies; under alternative strategies, market fragmentation actually reduces execution time and improves trader welfare. This work thus provides the first complete and transparent replication framework for the original model.

This work proposes the first market-making framework that explicitly integrates the continuous limit order book with the closing auction mechanism, addressing a key limitation of traditional strategies that often neglect liquidity dynamics during the closing auction and rely solely on terminal inventory penalties for risk control. By leveraging deep Q-learning, the framework dynamically forecasts the auction clearing price and is trained within a generative market simulator based on the rough Heston model. The approach continuously updates its clearing price predictions as the trading session progresses. Empirical evaluations demonstrate that the proposed method significantly outperforms classical optimal market-making benchmarks—both in synthetic environments and on real-world S&P 500 data—achieving higher profitability while effectively mitigating end-of-day inventory risk.

This study addresses a critical gap in the literature by jointly modeling learning mechanisms and heterogeneous preferences—factors often examined in isolation—in financial markets. It proposes a novel multi-agent reinforcement learning framework for an artificial financial market, where traders exhibit heterogeneity in risk aversion, time discounting, and information acquisition capabilities. These agents adaptively learn through interaction and evolve distinct trading strategies over time. The model successfully replicates key stylized facts of real financial markets, including fat-tailed return distributions and volatility clustering. By demonstrating how macro-level market dynamics emerge from micro-level heterogeneity and adaptive learning, this work provides a computational instantiation and empirical validation of the Adaptive Markets Hypothesis.

This work addresses the lack of a standardized backtesting framework in prediction markets, which hinders fair evaluation of trading agents under realistic market microstructure, transaction fees, and settlement risk. We propose the first event-driven replay system tailored for prediction markets, integrating limit order book simulation, a maker/taker fee model, and an interface for tool-augmented large language model agents. Inspired by SWE-bench, we introduce a standardized evaluation paradigm for this domain. Using data from the Kalshi platform, we construct four backtesting scenarios spanning cryptocurrency, weather, and sports events. Empirical results demonstrate that fee-aware strategies significantly outperform naive agents in high-volatility environments, as the latter suffer substantial losses due to neglecting transaction costs and settlement risks.