FinEvo: From Isolated Backtests to Ecological Market Games for Multi-Agent Financial Strategy Evolution Mingxi Zou, Jiaxiang Chen, Aotian Luo, Jingyi Dai, Chi Zhang, Dongning Sun, Zenglin Xu Abstract Conventional financial strategy evaluation relies on isolated backtests in static environments. Such evaluations assess each policy independently, overlook correlations and interactions, and fail to explain why strategies ultimately persist or vanish in evolving markets. We shift to an ecological perspective, where trading strategies are modeled as adaptive agents that interact and learn within a shared market. Instead of proposing a new strategy, we present FinEvo , an ecological game formalism for studying the evolutionary dynamics of multi-agent financial strategies. At the individual level, heterogeneous ML-based traders—rule-based, deep learning, reinforcement learning, and large language model (LLM) agents—adapt using signals such as historical prices and external news. At the population level, strategy distributions evolve through three designed mechanisms—selection, innovation, and environmental perturbation—capturing the dynamic forces of real markets. Together, these two layers of adaptation link evolutionary game theory with modern learning dynamics, providing a principled environment for studying strategic behavior. Experiments with external shocks and real-world news streams show that FinEvo is both stable for reproducibility and expressive in revealing context-dependent outcomes. Strategies may dominate, collapse, or form coalitions depending on their competitors—patterns invisible to static backtests. By reframing strategy evaluation as an ecological game formalism, FinEvo provides a unified, mechanism-level protocol for analyzing robustness, adaptation, and emergent dynamics in multi-agent financial markets, and may offer a means to explore the potential impact of macroeconomic policies and financial regulations on price evolution and equilibrium. 1 1 footnotetext: Equal contribution. 2 2 footnotetext: Corresponding authors. 3 3 footnotetext: Project lead. 1 Introduction Financial markets are among the most complex and adaptive systems, where outcomes depend not only on external signals but also on the collective behavior of participants. Most evaluation paradigms abstract away such interactions. Deep learning emphasizes predictive accuracy, from attention-based architectures such as TimeMixer ( Wang et al. 2024 ) to foundation-style models like Chronos ( Ansari et al. 2024 ) . Reinforcement learning extends to sequential decision-making, with recent advances in offline Reinforcement learning(RL) for data-efficient policy learning ( Kumar et al. 2020 ) and risk-sensitive RL for robustness under uncertainty ( Duan et al. 2016 ) . Meanwhile, LLM-based agents extend trading logic with reasoning and tool use: FinCon frames trading as structured financial reasoning with CVaR ( Yu et al. 2024 ) , FinGPT adapts foundation models for forecasting and decision support ( Liu et al. 2023 ) , and TradingAgents explore general-purpose planning and execution capabilities ( Xiao et al. 2024 ) . Yet most existing approaches optimize in isolation, neglecting the ecological interdependencies that shape real-world market performance. Rather than proposing new trading strategies or predictive models, our goal is to develop a principled ecological game formalism—serving as a “financial wind tunnel” to stress-test systemic risks and market shocks within a controlled environment(see Appendix H ). Simulation infrastructures such as ABIDES ( Byrd et al. 2020 ) and Mesa ( Masad et al. 2015 ) support market microstructure and agent-based experimentation, yet they operate primarily as execution environments without modeling population-level adaptation or long-horizon strategic dynamics. PyMarketSim ( Mascioli et al. 2024 ) further offers high-fidelity order-book execution for deep RL agents, but remains centered on single-agent optimization rather than collective ecological behavior. TraderTalk ( Vidler & Walsh 2024 ) explores LLM-driven behavioral trading agents, yet focuses on bilateral interactions without a framework for systemic market evaluation. Beyond these platforms, GPU-accelerated limit order book simulators such as JAX-LOB ( Frey et al. 2023 ) and fast agent-based frameworks for trading latency and RL ( Belcak et al. 2021 ) advance high-fidelity training on historical order-flow data, while ecosystem-style models such as HFTE ( Mahdavi-Damghani 2017 ) examine multi-species interactions among high-frequency strategies. However, none of these lines of work offer a unified evolutionary game-theoretic formulation for characterizing adaptation, population flows, and systemic impact in heterogeneous trading ecologies—a gap that FinEvo focuses on (see Table 6 ). To realize this vision, we introduce FinEvo , an ecological game formalism for studying the evolutionary dynamics of multi-agent financial strategies. The agent population spans rule-based, deep learning, reinforcement learning, and large language model (LLM) traders, each operating under parameterized policies and capital constraints, with prices endogenously determined by continuous double-auction clearing. Inspired by the replicator–mutator equation ( Nowak 2006 ) , FinEvo formalizes population dynamics through selection, innovation, and environmental perturbation, capturing adaptation, extinction, and ecosystem-level phenomena such as dominance cycles and regime shifts, enabling principled evaluation of robustness and adaptation under realistic market interaction. We provide a more detailed review of related work in Appendix A . In summary, our contributions are: 1) We introduce FinEvo ,an ecological game formalism that integrates heterogeneous traders with evolutionary population dynamics, formalized by the FinEvo SDE . Its mechanism decomposition unifies selection, innovation, and environmental perturbation, elevating evaluation from instance-level returns to mechanism-level behavior. 2) Under mild conditions, we establish minimal theoretical guarantees for the FinEvo SDE—simplex invariance, positivity, and existence/uniqueness—and derive a macro-level variance decomposition that attributes system volatility to selection, innovation, and perturbation. 3) We demonstrate through experiments with shocks and real-world signals that FinEvo reveals context-dependent performance, robustness, and emergent coalition dynamics that static backtests cannot capture. Together, these contributions establish FinEvo as a principled basis for studying robustness and adaptation in multi-agent financial markets. 2 Ecological Games: Problem Formulation and Evolutionary Dynamics We model financial strategy evaluation as an ecological game , where a population of N heterogeneous trader agents interact in a shared, dynamically evolving market. Each agent follows a distinct policy and adapts through feedback between market outcomes and external signals. This formulation allows us to study both the performance of individual strategies and the collective dynamics of selection, innovation, and perturbation. Agent behavior and market feedback. At each time step t , agent i selects an action a_{i,t}\sim\pi_{k(i)}(\mathcal{I}_{t}), where \pi_{k(i)} is the policy of type k , conditioned on recent prices \{P_{t-\tau},\dots,P_{t}\} and exogenous signals E_{t} . And the action space includes market buy, market sell, limit buy, limit sell, and hold . The market aggregates all actions and shocks \xi_{t} into the next price P_{t+1} , with the detailed matching mechanism given in Appendix G.6 , P_{t+1}=F\big(P_{t},\{a_{i,t}\}_{i=1}^{N},\xi_{t}\big). Executed trades then update portfolio holdings. Specifically, let C_{i,t} denote the cash holdings and S_{i,t} the assets holdings of agent i at time t , while P_{t} denotes the market-clearing price. C_{i,t+1}=C_{i,t}-a_{i,t}^{\text{buy}}P_{t}+a_{i,t}^{\text{sell}}P_{t}, and S_{i,t+1}=S_{i,t}+a_{i,t}^{\text{buy}}-a_{i,t}^{\text{sell}}. For strategy \pi_{k} , the realized payoff over [t,t{+}1] is \hat{f}_{k,t}=\frac{1}{N_{k,t}}\sum_{i:\pi_{k(i)}=\pi_{k}}\Big[C_{i,t+1}+S_{i,t+1}P_{t+1}-(C_{i,t}+S_{i,t}P_{t})\Big], representing the average incremental wealth change of its adopters. Individual adaptation. Each strategy maintains an internal parameter state \Psi_{k,t} that governs how it maps information \mathcal{I}_{t} into actions. This state evolves through a generic adaptation operator \Psi_{k,t+1}=\mathcal{A}\big(\Psi_{k,t},\hat{f}_{k,t},E_{t}\big), where \mathcal{A} abstracts diverse mechanisms of learning and adjustment. In practice, this update may correspond to temporal-difference or policy-gradient rules in reinforcement learning, gradient-based parameter tuning in deep learning models, in-context adaptation in large language models, or heuristic adjustment in rule-based strategies. Beyond reacting to immediate rewards, traders typically attempt to estimate a long-term value of their strategies. To capture this, we associate with each \pi_{k} a forward-looking value process V_{k}(t) , modeled as dV_{k}(t)=\lambda_{k}\big(f_{k}(E_{t})-V_{k}(t)\big)\,dt+\nu_{k}\,dB_{k}(t) , where f_{k}(E_{t}) denotes an environment-conditioned target payoff, \lambda_{k}>0 controls the adjustment speed, and \nu_{k}dB_{k}(t) introduces stochastic exploration. This Ornstein–Uhlenbeck process represents convergence of long-term value estimates toward external signals while retaining variability due to uncertainty, serving as a forward-looking anchor that guides subsequent population-level dynamics. Population evolution (FinEvo SDE). Let X_{t}=(x_{1}(t),\ldots,x_{K}(t))^{\top}\in\Delta^{K-1} denote the population distribution over K strategies, with \sum_{k}x_{k}(t)=1 . The evolution of X_{t} can be abstractly represented by an operator x_{t+1}=\mathcal{G}\big(x_{t},V(t),m_{t},\xi_{t}\big), where V(t)=(V_{1}(t),\ldots,V_{K}(t))^{\top} collects the forward-looking values estimated at the individual level, m_{t} denotes the innovation distribution, and \xi_{t} represents stochastic perturbations. Concretely, we formulate \mathcal{G} as a continuous-time stochastic differential equation, the FinEvo SDE , which integrates the three designed mechanisms—selection, innovation, and perturbation—into a unified dynamic: \displaystyle dX_{t} \displaystyle=b(X_{t},t)\,dt+\Sigma(X_{t},t)\,dW_{t} \displaystyle=\underbrace{\beta\mathrm{diag}(X_{t})\,\big(V(t)-\bar{V}(t)\mathbf{1}\big)}_{\textbf{Selection}}\,dt+\underbrace{\mu\,(m_{t}-X_{t})}_{\textbf{Innovation}}\,dt+\underbrace{\gamma\,\mathrm{diag}(X_{t})\,\sigma\,P(X_{t})\,dW_{t}}_{\textbf{Perturbation}}, where \bar{V}(t)=X_{t}^{\top}V(t) is the population mean payoff, P(X_{t})=I-\mathbf{1}X_{t}^{\top} projects perturbations onto the simplex tangent space, \sigma=\mathrm{diag}(\sigma_{1},\ldots,\sigma_{K}) , \gamma denotes the global perturbation scaling factor and W_{t} is a K -dimensional Brownian motion. The target distribution m^{t}\sim\mathrm{Dir}(\alpha_{1},\ldots,\alpha_{K}) models social influence and random experimentation, ensuring m^{t}\in\Delta^{K-1} and maintaining diversity across strategies. This formulation preserves the simplex, ensures positivity, and admits invariant measures under mild conditions (Appendix E ). While related to replicator–mutator equations, FinEvo grounds selection in payoff signals V(t) and explicitly incorporates innovation and perturbations, yielding a more flexible and robust model of population evolution. Independence Assumptions. The independence assumptions (A1–A3) used in our variance decomposition serve only to obtain a closed-form expression. They are not required for the FinEvo SDE itself: the dynamics are unchanged when payoff and environmental shocks are correlated. In the correlated case, the decomposition simply acquires additional covariance terms (see Appendix D ). Mechanism decomposition. In discrete approximation, the short-run change in share \Delta x_{k} is shaped by three forces: \Delta x_{k}\approx\beta x_{k}^{t}\big(V_{k}^{t}-\bar{V}^{t}\big)+\mu\,(m_{k}^{t}-x_{k}^{t})+x_{k}^{t}\Big(\gamma\sigma_{k}\eta_{k}^{t}-\gamma\sum_{j}x_{j}^{t}\sigma_{j}\eta_{j}^{t}\Big),\quad\eta_{k}^{t}\sim\mathcal{N}(0,1). Correspondingly, their contributions to short-run volatility are \mathrm{Var}(\Delta x_{k})\approx\underbrace{\beta^{2}(x_{k}^{t})^{2}\tfrac{\nu_{k}^{2}}{2\lambda_{k}}}_{\text{Selection}}+\underbrace{\mu^{2}\tfrac{m_{k}(1-m_{k})}{\alpha_{0}+1}}_{\text{Innovation}}+\underbrace{\gamma^{2}\Big((x_{k}^{t})^{2}\sigma_{k}^{2}-2x_{k}^{t}\sigma_{k}\sum_{j}x_{j}^{t}\sigma_{j}+\Big(\sum_{j}x_{j}^{t}\sigma_{j}\Big)^{2}\Big)}_{\text{Perturbation}}. Together, these decompositions show how selection, innovation, and perturbation jointly shape both the direction and volatility of evolutionary adaptation in financial markets.Detailed proof can be seen in the Appendix C . 3 Experiments and Analysis of Evolutionary Dynamics 3.1 Experimental Setup To study how financial strategies evolve in an ecological setting, we build a multi-agent simulation environment. Within this environment, we test two scenarios: one with artificial shocks and another with real-world news events.We simulate a synthetic market with news shocks, multi-indicator redistribution, and eco-evolutionary updates (see Appendix G.6.1 ). A full discussion of the motivation for these scenarios is provided in Appendix G.7 . Agent Population The market is composed of 20 trader agent archetypes, each with distinct behaviors, information sources, and decision-making logic. For clarity, we group these agents into four broad categories: (1) Rule-based Agents , which include traditional strategies such as trend-following, mean-reversion ( Jegadeesh & Titman 1993 ; De Bondt & Thaler 1985 ) , and noise trading, along with fundamental (informational) traders; we further endow them with heterogeneous trading styles (e.g., aggressive and conservative ) to reflect real-world diversity. (2) Deep Learning Agents , supervised models such as Informer and TimeMixer ( Zhou et al. 2021 ; Wang et al. 2024 ) , which use neural networks to forecast price direction from market indicators. (3) Reinforcement Learning (RL) Agents , adaptive strategies (e.g., DoubleDQN , PPO ) ( Van Hasselt et al. 2016 ; Schulman et al. 2017 ) that learn policies online by interacting with the market to maximize cumulative returns. (4) Large Language Model (LLM) Agents , which parse textual information such as news and announcements to infer market sentiment and generate trading signals (e.g., FinCon , TradingAgents ) ( Yu et al. 2024 ; Xiao et al. 2024 ) ,with GPT-4o-mini ( Hurst et al. 2024 ) serving as the backbone language model. Details of each agent’s design are presented in Appendix F.2.1 . Abbreviations are adopted throughout the main text, with the full mapping summarized in Appendix F.1 , and complete experimental parameter configurations can be found in Appendix G.6.2 . Simulation Scenarios Each scenario is evaluated through Monte Carlo experiments, with 128 independent runs per configuration. We consider two settings: (1) Artificial Shocks , where predefined perturbations are injected into the market to examine system stability and resilience (see Appendix G.6.4 ); and (2) Real-World News , simulating July 2023–July 2025 with actual historical news streams(e.g., GDELT and Reuters Markets; details in Appendix G.6.5 ). The market operates as a continuous double auction from 09:00 to 17:00 (slightly longer than real trading hours to allow market states to converge) each day, enabling event-driven and LLM agents to process real information and interact under realistic intraday constraints. Analysis Metrics Our evaluation framework follows a micro–meso–macro structure (detailed in Appendix G.6.3 ). At the micro level , we assess individual performance and diversity through population proportion (market shares of each agent), agent win rate (frequency with which an agent’s return is ranked within the top three performers across evaluation trials), and financial metrics including return with confidence interval, Sharpe ratio, maximum drawdown, and turnover. At the meso level , we capture population diversity and interaction structures via concentration (HHI), strategy entropy, modularity (community strength), synergy vs. antagonism (correlation heatmaps), co-occurrence frequency, and mutual-information networks reflecting dependency and alliance reconfiguration. At the macro level , we evaluate system volatility and regime shifts by decomposing volatility into performance pressure ( V_{\text{select}} ), innovation/social mixing ( V_{\text{innovation}} ), and environmental Perturbation ( V_{\text{perturbation}} ),as well as measuring phase changes , which captures shifts in the dominant strategy cluster. A formal definition is provided in Appendix G.6.3 . For visualization, raw time-series data (e.g., population shares and volatility metrics) are smoothed using a Savitzky–Golay filter. In addition, following standard practice in empirical finance, we report Excess Kurtosis , Skewness , Sharpe Dispersion , and Vol-of-Vol to quantify fat tails, return asymmetry, cross-strategy performance dispersion, and volatility clustering. These metrics, widely used to characterize Stylized Facts of real markets, provide an external validation of the realism of the endogenous dynamics generated by FinEvo. 3.2 Simulation Under Artificial Shocks We test how the market ecology adapts under large external artificial disturbances. In our synthetic market, we inject news shocks between 12:00 and 13:30, considering both a strongly positive and a strongly negative scenario. This setting allows us to examine adaptation at the micro, meso, and macro levels. Micro: Market Share Dynamics. Figure 1: Agent population shares over time. Left: positive shock; Right: negative shock. Figure 1 shows that responses are strongly asymmetric. Positive shocks increase diversity by reducing CoT’s dominance and redistributing market share across multiple agents (e.g., CIT, ReAct), whereas negative shocks reinforce concentration, leading to LLM-dominated equilibria (FinCon, ReAct). Thus, shocks drive regime-dependent transitions in dominance concentration, highlighting dynamics not captured by static benchmarks. Meso: Interaction Structures and Stability. Figure 2: Correlation matrices over time under positive (top) and negative (bottom) shocks. Red = cooperation, Blue = competition. Figure 2 tracks the reorganization of alliances. Before the shock, large red clusters indicate coordination groups where many strategies move together. During the shock, these clusters fragment as blue bands appear, reflecting stronger competition and divergence. After the shock, new cooperative clusters form, but their structures diverge: pluralistic in the positive case, concentrated around LLMs in the negative case. This contrast highlights the path-dependent nature of adaptation. (a) Strategy-pair correlations under shocks. (b) Volatility decomposition under shocks. Figure 3: Responses under external shocks: Left = positive shock, Right = negative shock. In (b), Blue = selection, Red = innovation, Green = perturbation. Zooming into individual pairs (Figure 3(a) ), we observe three consistent patterns: (1) some pairs maintain persistent cooperation, (2) others maintain persistent competition, and (3) shocks induce only localized reorganization. Thus, while alliances shift, the broader cooperation–competition backbone remains intact, providing structural stability. Macro: Volatility Decomposition. Figure 3(b) decomposes volatility into three components: selection ( \mathbf{V_{\mathrm{selection}}} ), innovation ( \mathbf{V_{\mathrm{innovation}}} ), and perturbation ( \mathbf{V_{\mathrm{perturbation}}} ). Across both scenarios, shocks trigger a spike in perturbation, destabilizing alliances. Afterward, selection dominates: in the positive case, it supports a pluralistic equilibrium, while in the negative case, it drives convergence toward LLM dominance. Innovation remains steady, acting as a background force that preserves diversity rather than driving short-term change. Summary. External shocks reshape the market at all levels: they break monopolies or reinforce concentration (micro), fragment and reorganize alliances (meso), and amplify volatility before selection restores order (macro). Overall, adaptation is governed by the interplay of selection, innovation, and perturbation. Shocks amplify perturbation in the short run, but long-run equilibria are primarily shaped by selection, with innovation maintaining diversity. For completeness, we also analyze the evolution of co-occurrence matrices and mutual-information networks under shocks, which provide complementary evidence of robustness from the perspective of networked interactions. The detailed results are reported in Appendix G.1 . 3.3 Empirical Analysis of Intra-Day Evolutionary Dynamics After examining the adaptive dynamics under simulated shocks, we next turn to real-world data and analyze the intra-day evolutionary dynamics within a single trading day. This allows us to validate whether the eco-evolutionary patterns observed in controlled experiments persist under actual market conditions. Figure 4: Intra-day evolution on July 7, 2025: (left) agent population proportions, (middle) theoretical volatility metrics, and (right) cumulative returns under a 0.5% transaction cost (including slippage). Intra-Day Adaptation and Profitability. Figure 4 shows that intra-day dynamics follow a multi-phase cycle: strong selection initially concentrates capital in dominant strategies, sustained innovation later restores diversity and balance, and heightened perturbation near the close disrupts alliances before a new equilibrium emerges. Informer and TimeMixer ultimately prevail, highlighting how evolutionary pressures jointly shape regime transitions within a single trading day. Beyond survival, a natural question is which agents convert adaptation into profitability under market frictions. Table 1 reports averages over the same rounds (Return with 95% CI, Sharpe, MaxDrawdown, Turnover). Most agents lose, but a few achieve positive returns (shaded) despite large drawdowns, enabled by diversity-preserving innovation. Higher turnover further indicates that profitable agents trade more actively, highlighting a market ecology that is both selective and tolerant. Table 1: Performance metrics by agent (Return with 95% CI, Sharpe, MaxDrawdown, Turnover;2025-07-07, N=128). Positive returns are shaded in gray. Agent Return Sharpe MaxDD Turn. Agent Return Sharpe MaxDD Turn. Agent Return Sharpe MaxDD Turn. MT -0.832 ± 0.064 -2.016 -0.908 0.157 ReAct -0.685 ± 0.025 15.972 -0.978 0.207 Informer 4.333 ± 0.022 26.928 -0.771 0.362 RSI -0.849 ± 0.023 -3.546 -0.888 0.122 CoT -0.560 ± 0.046 5.118 -0.815 0.434 TimeMixer 6.738 ± 0.031 30.392 -0.790 0.463 BT 0.034 ± 0.010 10.804 -0.802 0.309 ITT -0.896 ± 0.066 -1.228 -0.910 0.119 PPO -0.431 ± 0.033 6.641 -0.859 0.123 CIT -0.544 ± 0.032 6.265 -0.844 0.163 Trading -0.746 ± 0.048 -2.244 -0.979 0.890 A2C -0.733 ± 0.004 0.847 -0.837 0.197 AIT -0.807 ± 0.026 -3.121 -0.838 0.126 FinCon -0.685 ± 0.045 -2.605 -0.958 0.845 DQN -0.885 ± 0.034 13.436 -0.955 0.139 SIT -0.673 ± 0.057 0.624 -0.948 0.334 SAIT -0.866 ± 0.010 -0.933 -0.909 0.099 OIT -0.915 ± 0.032 2.265 -0.942 0.130 Dynamic Alliance Reconfiguration. Beyond individual outcomes, we next examine the structural layer of interactions. Figure 5 illustrates the intra-day evolution of agent-agent correlations, showing a dynamic interplay between synergistic alliances and strategic antagonisms . The market initially features multiple competing clusters with strong internal coordination, which are subsequently fragmented and reorganized under rising environmental perturbation. Toward the close, these transient shifts consolidate into a new regime characterized by restructured alliances and altered power balances. We further report the intra-day evolution of co-occurrence matrices and mutual-information networks in Appendix G . Figure 5: Dynamic evolution of intra-day correlation matrices on July 7, 2025. Red blocks represent synergistic alliances , while blue blocks indicate strategic antagonism . Performance Rankings in Bull vs. Bear Markets. Finally, moving from intra-day fluctuations to market cycles, we evaluate whether these patterns persist across long-term bull and bear regimes. We approximate January–October 2022 as a bear market phase , objectively defined by persistent downward dynamics in the synthetic environment, and October 2024–July 2025 as a bull market phase , characterized by sustained recoveries and new highs. Figure 6 compares podium finishes and top-3 win rates across these regimes. Figure 6: Agent performance rankings in bull vs. bear markets. Our analysis shows that agent performance is highly regime-dependent : bull markets produce a concentrated dominance by a few strategies, whereas bear markets foster a more dispersed competitive landscape. Notably, LLM-based agents consistently rank higher across both regimes , suggesting that their ability to integrate diverse informational cues and adapt flexibly confers a significant robustness advantage over strategies based on fixed rules or purely quantitative signals. Evolutionary Game Modeling vs. Agent-Based Simulation in Market Environments Table 2: Comparison of Ecological Metrics Across Evolutionary Game Modeling and Agent-Based Simulation Platform & Phase Entropy {}_{\text{mean}} HHI {}_{\text{mean}} Top1 {}_{\text{mean}} AvgAbsCorr {}_{\text{mean}} Modularity {}_{\text{mean}} PhaseChanges {}_{\text{mean}} FinEvo (bull) 0.732 \pm 0.016 0.226 \pm 0.032 0.453 \pm 0.024 0.441 \pm 0.019 0.091 \pm 0.028 3 \pm 0.437 FinEvo (bear) 1.167 \pm 0.036 0.175 \pm 0.029 0.241 \pm 0.067 0.376 \pm 0.039 0.063 \pm 0.018 4 \pm 0.348 ABIDES (bull) 0.516 \pm 0.045 0.539 \pm 0.011 0.746 \pm 0.031 0.653 \pm 0.046 0.153 \pm 0.023 1 \pm 0.439 ABIDES (bear) 0.623 \pm 0.011 0.481 \pm 0.012 0.615 \pm 0.045 0.541 \pm 0.022 0.132 \pm 0.032 1 \pm 0.344 PyMarketSim (bull) 0.539 \pm 0.005 0.564 \pm 0.019 0.773 \pm 0.046 0.662 \pm 0.039 0.159 \pm 0.034 1 \pm 0.332 PyMarketSim (bear) 0.662 \pm 0.037 0.461 \pm 0.015 0.719 \pm 0.023 0.575 \pm 0.007 0.144 \pm 0.019 1 \pm 0.329 Table 2 compares the performance of FinEvo, ABIDES, and PyMarketSim during bull and bear markets. FinEvo exhibits a dynamic, adaptive environment where agents continuously adjust strategies, promoting greater diversity and competition. In contrast, ABIDES and PyMarketSim are more rigid, with fewer strategies dominating over time, limiting adaptability and diversity. This demonstrates FinEvo’s strength in capturing market evolution, while the other platforms show more static behavior. 3.4 Ablation Study Ablation on Evolutionary Pressures To disentangle the roles of the three evolutionary pressures (selection, innovation, perturbation), we remove each component in turn and compare the resulting population dynamics (Figure 7 ). Without \mathbf{V_{\mathrm{selection}}} (left), competitive reinforcement vanishes and all agents remain at comparable proportions, producing a static, inefficient ecology. Without \mathbf{V_{\mathrm{innovation}}} (middle), diversity collapses as FinCon and Trading rapidly dominate, leading to an oligopolistic lock-in. Without \mathbf{V_{\mathrm{perturbation}}} (right), dynamics become smoother but rigid: FinCon and Trading still dominate, while adaptive reorganization is muted. Overall, selection drives efficiency, innovation sustains diversity, and disturbances from the environment can amplify shocks within the system, preventing premature lock-in and enabling regime shifts.. Figure 7: Ablation study of evolutionary pressures. From left to right: removing selection , innovation , and perturbation . Table 3: Ablation study results with global ecological metrics (mean \pm 95% CI). Setting Entropy {}_{\text{mean}} HHI {}_{\text{mean}} Top1 {}_{\text{mean}} AvgAbsCorr {}_{\text{mean}} Modularity {}_{\text{mean}} PhaseChanges {}_{\text{mean}} FinEvo 0.828\pm 0.14 0.189\pm 0.012 0.344\pm 0.015 0.384\pm 0.021 0.074\pm 0.009 4\pm 0.663 FinEvo (w/o selection) 1.793\pm 0.25 0.081\pm 0.108 0.074\pm 0.009 0.139\pm 0.018 0.121\pm 0.012 14\pm 0.767 FinEvo (w/o innovation) 0.535\pm 0.13 0.477\pm 0.014 0.531\pm 0.033 0.553\pm 0.024 0.023\pm 0.007 1\pm 0.483 FinEvo (w/o perturbation) 0.719\pm 0.04 0.245\pm 0.011 0.376\pm 0.017 0.474\pm 0.019 0.036\pm 0.008 3\pm 0.591 We complement the single-day dynamics in Figure 7 with aggregated statistics over the full evaluation period (Table 3 ). These global metrics provide a more robust picture: removing selection yields inflated entropy and frequent phase transitions, indicating unstable fragmentation; removing innovation collapses diversity and locks the ecology into an oligopoly; removing perturbation leads to higher concentration but fewer regime shifts, reflecting rigidity. Together, these results substantiate the earlier case study: selection promotes efficiency, innovation preserves diversity, and perturbation enables adaptive reorganization. Figure 8: Robustness to random seeds (July 7, 2025). Left: trade price trajectories; Right: strategy entropy. Curves show means; shaded bands indicate 95% confidence intervals across 128 runs. Ablation on Robustness(Random seeds). We conducted an ablation study based on the simulation carried out on July 7, 2025, with 128 rounds, in order to evaluate the robustness of our framework under different random seeds. As shown in Figure 8 , we report the cross-run mean of trade price trajectories and strategy entropy, together with the shaded bands representing 95% confidence intervals across runs. Despite minor run-to-run variations, the core patterns remain consistent: (i) on the price side, the market exhibits a stable upward trend following external shocks with bounded fluctuations; (ii) on the strategy side, the entropy remains persistently high, indicating sustained diversity and heterogeneity. The relatively narrow shaded regions suggest that different random seeds lead to similar dynamics, thereby demonstrating the stability and robustness of our framework under stochastic perturbations. In addition, we report trade price trajectories under simulated exogenous shocks, which further demonstrate the robustness of our framework(see Appendix G.1 ) for detailed results. Ablation on Parameter Sensitivity. Figure 9: Sensitivity analysis on evolutionary parameters (simulation of July 7, 2025). Left: effect of selection strength \beta on market concentration . Middle: effect of innovation rate \mu on strategy diversity . Right: effect of perturbation intensity \gamma on volatility contribution .Error bars indicate 95\% CI across 128 runs. We further conduct sensitivity experiments on the three key parameters of our evolutionary framework: the selection strength \beta , the mutation rate \mu , and the environmental perturbation intensity \gamma . We report system-level outcomes using three complementary metrics: (i) average concentration (Herfindahl–Hirschman Index, HHI), (ii) average diversity (Shannon entropy), and (iii) average perturbation contribution V_{\text{perturbation}} . Figure 9 illustrates the results. As \beta increases, the average concentration rises significantly, suggesting that stronger selection pressure accelerates the dominance of a few strategies. In contrast, higher \mu leads to increased diversity, confirming that mutation and social mixing mechanisms help maintain heterogeneity. Meanwhile, increasing \gamma monotonically elevates V_{\text{perturbation}} , demonstrating the independent contribution of exogenous perturbations to volatility. We additionally conducted a \pm 20\% perturbed-parameter robustness test; the results (Appendix G.3 ) show that all stylized-fact metrics remain stable, confirming that FinEvo’s dynamics are not sensitive to initial agent parameters. Ablation on Market Size N . We examine the robustness of FinEvo with respect to the market size N , i.e., the number of agents participating in the ecology. Table 4 reports stylized-fact and performance metrics under four market sizes N\in\{60,80,100,120\} , for both bull and bear regimes (128 runs per setting). Table 4: Ablation on market size N (128 runs). Stylized-fact metrics remain stable across N . Values are mean \pm 95% CI. Regime N Ex. Kurtosis {}_{\text{mean}} \uparrow Skewness {}_{\text{mean}} \downarrow Sharpe Disp. {}_{\text{mean}} \uparrow Vol-of-Vol {}_{\text{mean}} \uparrow Bullish 120 3.143\pm 0.141 -1.539\pm 0.132 1.673\pm 0.056 0.009\pm 0.0004 100 3.163\pm 0.138 -1.532\pm 0.129 1.669\pm 0.061 0.008\pm 0.0004 80 3.178\pm 0.165 -1.541\pm 0.158 1.657\pm 0.073 0.008\pm 0.0005 60 3.159\pm 0.178 -1.516\pm 0.162 1.659\pm 0.067 0.007\pm 0.0005 Bearish 120 4.326\pm 0.169 -1.805\pm 0.152 1.759\pm 0.091 0.012\pm 0.0006 100 4.259\pm 0.172 -1.793\pm 0.149 1.732\pm 0.089 0.013\pm 0.0007 80 4.115\pm 0.175 -1.765\pm 0.161 1.741\pm 0.138 0.013\pm 0.0007 60 4.251\pm 0.197 -1.772\pm 0.183 1.725\pm 0.119 0.011\pm 0.0007 Across all values of N , FinEvo consistently maintains heavy-tailed PnL (high excess kurtosis), strong downside asymmetry (negative skewness), and comparable Sharpe dispersion and Vol-of-Vol, with mean returns fluctuating only within a narrow band. 3.5 Stylized-Fact Validation A key question is whether FinEvo’s endogenous price formation—despite Gaussian exogenous shocks—can reproduce classical market stylized facts ( Cont 2001 ) . To address this, we compare stylized-fact metrics across FinEvo , ABIDES , and PyMarketSim , running 128 simulations in both bull and bear regimes and computing standard econometric measures (excess kurtosis, skewness, Sharpe dispersion, and Vol-of-Vol). FinEvo matches the heavy tails, negative skewness, and elevated volatility-of-volatility more closely than the baselines(Table 5 ), and this behavior is robust: a correlated, heavy-tailed shock variant yields similar stylized-fact metrics (Appendix G.4 , Table 9 ).We also provide a small cross-asset robustness check in Appendix G.5 . Table 5: Comparison of stylized-fact metrics across simulators (128 runs). Values report mean \pm 95% CI. Simulator / Regime Excess Kurtosis {}_{\text{mean}} \uparrow Skewness {}_{\text{mean}} \downarrow Sharpe Dispersion {}_{\text{mean}} \uparrow Vol-of-Vol {}_{\text{mean}} \uparrow FinEvo (bull) 3.163\pm 0.138 -1.532\pm 0.129 1.669\pm 0.061 0.008\pm 0.0004 FinEvo (bear) 4.259\pm 0.172 -1.793\pm 0.149 1.732\pm 0.089 0.013\pm 0.0007 ABIDES (bull) 0.315\pm 0.041 -0.209\pm 0.036 0.675\pm 0.028 0.004\pm 0.0002 ABIDES (bear) 0.562\pm 0.053 -0.474\pm 0.047 0.836\pm 0.033 0.007\pm 0.0003 PyMarketSim (bull) 0.446\pm 0.044 -0.186\pm 0.032 0.512\pm 0.026 0.003\pm 0.0002 PyMarketSim (bear) 0.591\pm 0.058 -0.438\pm 0.041 0.906\pm 0.034 0.005\pm 0.0003 4 Conclusion and Future Work In this work, we introduce an ecological game formalism for modeling market ecology, which unifies selection, innovation, and environmental perturbation into a single mathematical formulation. We implement the framework in a multi-agent financial market environment and, through large-scale simulations, demonstrate its ability to capture the evolution of stability, concentration, and diversity across different market phases, thereby helping to narrow the gap between theoretical analysis and empirical dynamics. Overall, our study establishes a three-layer mapping from mechanisms to metrics to behaviors, highlighting both the interpretability and robustness of the proposed framework. FinEvo is designed with modular components that naturally support further extensions. A brief discussion of potential future directions is provided in Appendix I . 5 Ethics statement FinEvo provides an ecological game formalism for studying the evolution of financial strategies. The datasets used consist of publicly available financial market data and synthetic news streams, which do not involve personal or sensitive information; thus, privacy concerns are minimal. To reduce potential risks, our framework is intended for scientific investigation of market dynamics and policy evaluation, rather than for the development of deployable trading algorithms. 6 Reproducibility statement We are committed to ensuring the reproducibility of our results. All datasets used in this work are publicly available (see Appendix G.6.5 ), and we will release our source code, model implementations, and configuration files upon publication. To guarantee robustness and stability, each experiment was repeated with 128 Monte Carlo runs, and we report averaged results along with 95% confidence intervals (see Appendix G.6.1 ). Detailed descriptions of model architectures F.2.1 , hyperparameters G.6.2 , and evaluation matrics G.6.3 are provided in the main text and appendix to facilitate independent verification. Appendix A Related Work A.1 Financial Agents The concept of financial agents abstracts market participants as decision-making entities with heterogeneous rules, objectives, and information sets. This framing emphasizes (i) heterogeneity — agents differ in behavioral assumptions such as trend-following, mean-reversion, or news sensitivity; and (ii) interaction — outcomes depend not only on external data but also on the actions of other participants. By treating strategies as agents, prior research provides a unifying abstraction that accommodates diverse modeling approaches. Within this perspective, several families have emerged. Rule-based agents include momentum, mean-reversion, breakout, and pairs trading rules ( Brock et al. 1992 ; Gatev et al. 2006 ) , as well as extensions with overlays for volatility or stop-loss control ( Harvey et al. 2018 ) . Data-driven agents employ modern sequence models: TCN and N-BEATS for long-horizon forecasting ( Oreshkin et al. 2019 ) , Informer and FEDformer for efficient attention ( Zhou et al. 2021 ; Zhou et al. 2022 ) , and foundation-style models such as TimeGPT and Chronos for zero-shot transfer ( Garza et al. 2023 ; Ansari et al. 2024 ) . News- and multimodal agents integrate text, prices, and sentiment, from FinBERT for financial language ( Huang et al. 2023 ) to multimodal benchmarks such as FinMME ( Luo et al. 2025 ) . LLM-based agents extend trading logic with reasoning and tool use, exemplified by FinCon, FinMem, and FinGPT ( Yu et al. 2024 ; Yu et al. 2025 ; Liu et al. 2023 ) , as well as hybrid systems that combine forecasting models with LLM reasoning ( Gan et al. 2025 ) . These developments illustrate how financial agents have evolved into adaptive, data-rich decision makers. Yet most evaluations still focus on individual performance rather than population-level adaptation, motivating our ecological perspective where heterogeneous agents interact and co-adapt within shared markets. A.2 Reinforcement Learning Reinforcement learning (RL) provides a complementary perspective to our ecological formulation in FinEvo. Both view markets as sequential decision processes under uncertainty: RL formalizes policy optimization over state–action trajectories ( Sutton et al. 1998 ) , while FinEvo frames the dynamics of strategy populations as stochastic replicator–mutator processes. The parallel is clear: maximizing long-horizon return in RL ( Mnih et al. 2015 ; Schulman et al. 2017 ) mirrors survival and dominance of strategies in evolutionary games, though RL emphasizes a single adaptive policy while FinEvo studies heterogeneous populations. RL draws on several traditions. Dynamic programming and optimal control inspire value-based methods ( Bellman 1966 ; Puterman 2014 ) , bandit and online learning emphasize regret minimization ( Auer et al. 2002 ; Bubeck et al. 2012 ) , and control-as-inference links planning to probabilistic modeling ( Levine 2018 ) . These abstractions parallel FinEvo’s payoff-driven selection, innovation, and environmental perturbation modeling. Methodological advances further connect the two domains. Scalable actor–critic algorithms such as A3C ( Mnih et al. 2016 ) highlight the role of parallelism and sample efficiency in training adaptive policies. Offline RL and off-policy evaluation ( Kumar et al. 2020 ; Uehara et al. 2020 ) echo our concern with reproducible evaluation under counterfactual conditions. Risk-sensitive RL ( Mihatsch & Neuneier 2002 ) aligns with our treatment of shocks as ecological stressors. Preference- and feedback-driven RL ( Christiano et al. 2017 ) resembles our innovation mechanism, where external signals redirect evolutionary trajectories. Meanwhile, sequence-decision formulations such as decision transformers ( Chen et al. 2021 ) illustrate how supervised learning and RL converge, analogous to FinEvo’s integration of forecasting with adaptation. From this vantage point, RL and FinEvo can be seen as two instantiations of a common paradigm: learning under uncertainty with delayed feedback, balancing exploitation, exploration, and robustness. Yet, RL typically optimizes an isolated policy, leaving open how heterogeneous strategies evolve collectively under market pressures. FinEvo addresses this gap by extending RL principles to agent populations, bridging individual-level optimization with ecological dynamics. A.3 Ecological Games Ecological game theory extends classical game theory by modeling the adaptation and survival of strategy populations rather than equilibrium among rational individuals ( Hofbauer & Sigmund 1998 ; Sandholm 2010 ) . Originating in biology, dynamics such as the replicator equation ( Taylor & Jonker 1978 ) capture how successful strategies propagate, while mutation and drift introduce continual innovation. This perspective has influenced machine learning through evolutionary algorithms ( Bäck et al. 1997 ; Deb et al. 2002 ) , multi-agent evolutionary games ( Tuyls et al. 2006 ) , and population-based training ( Jaderberg et al. 2017 ) . Recent work emphasizes open-endedness, as in AlphaStar’s league training ( Vinyals et al. 2019 ) , autocurricula for exploration ( Leibo et al. 2019 ) , and open-ended strategy ecology ( Grill et al. 2020 ) . All highlight that robustness arises from continual adaptation within diverse populations. In finance, ecological models explain heterogeneous traders’ interactions ( Lux & Marchesi 1999 ; Hommes 2006 ) and are instantiated in agent-based markets ( LeBaron 2006 ; Farmer & Foley 2009 ) . More recently, high-fidelity platforms like ABIDES ( Byrd et al. 2020 ; Amrouni et al. 2021 ) provide infrastructure for market microstructure research, though they lack formal guarantees on population dynamics. Closer in spirit, the HFTE model ( Mahdavi-Damghani 2017 ) proposes a multi-species predator–prey ecosystem for high-frequency quantitative strategies, but relies on fixed topologies and genetic algorithms rather than a stochastic replicator–mutator dynamic.FinEvo complements this line of work by providing an explicit evolutionary game-theoretic SDE that links heterogeneous strategy populations with CDA microstructure. FinEvo advances beyond infrastructure by introducing a principled ecological formalism with stochastic replicator–mutator dynamics, enabling reproducible and theory-grounded evaluation of robustness and adaptation. Dimension ABIDES Mesa PyMarketSim TraderTalk FinEvo (ours) Goal / Target Market microstructure simulation General-purpose agent-based modeling RL benchmark for trading strategies LLM-driven bilateral trading behavior Mechanism-level evaluation of evolving financial ecosystems Theoretical Guarantees None None Limited (training stability only) None Simplex invariance, positivity, OU-process convergence Mechanism Layering Agent-level only Agent-level only Agent-policy interactions Dialogue-driven actions Selection / Innovation / Perturbation decomposition Research Use Stress-test specific microstructure Generic agent dynamics demos RL strategy benchmarking LLM role-playing for traders Systematic evaluation, stress-test, mechanism attribution Reproducibility Depends on configuration High (but non-financial focus) Medium (sensitive to seeds) Low (LLM variability) High, with cross-seed robustness and theoretical consistency Table 6: Comparison of FinEvo with existing agent-based financial market platforms. Appendix B From FinEvo SDE to Discrete Simulation Update This appendix derives the discrete-time update rule implemented in our simulator from the continuous-time FinEvo SDE . The goal is to make explicit how theoretical dynamics map to a practical update. B.1 Continuous-time formulation Let X_{t}=(x_{1}(t),\ldots,x_{K}(t))^{\top}\in\Delta^{K-1} be the population shares with \sum_{k}x_{k}(t)=1 . The FinEvo dynamics are dX_{t}=\underbrace{\beta\,\mathrm{diag}(X_{t})\!\big(V(t)-\bar{V}(t)\mathbf{1}\big)}_{\text{Selection}}\,dt+\underbrace{\mu\,(m_{t}-X_{t})}_{\text{Innovation}}\,dt+\underbrace{\gamma\,\mathrm{diag}(X_{t})\,\sigma\,P(X_{t})\,dW_{t}}_{\text{Perturbation}}, (1) where \bar{V}(t)=X_{t}^{\top}V(t) , m_{t}\in\Delta^{K-1} is the innovation distribution, \sigma=\mathrm{diag}(\sigma_{1},\ldots,\sigma_{K}) , and P(X)=I-\mathbf{1}X^{\top} projects perturbation onto the simplex tangent space. B.2 Step 1: Selection–perturbation flow Consider first equation 1 without innovation. For each coordinate, define y_{k}=\log x_{k} . By Itô’s lemma, to leading order on a short interval with V(t) frozen, dy_{k}\;\approx\;\beta\,(V_{k}-\bar{V})\,dt+\gamma\,\sigma_{k}\,dB_{k}, where B_{k} are independent Brownian motions (after projection). Integrating over \Delta t gives y_{k}(t+\Delta t)-y_{k}(t)\;\approx\;\beta\,(V_{k}-\bar{V})\,\Delta t+\gamma\,\sigma_{k}\,\sqrt{\Delta t}\,\xi_{k},\quad\xi_{k}\sim\mathcal{N}(0,1). (2) Exponentiating equation 2 yields the multiplicative-weights update \hat{x}_{k}(t+\Delta t)=\frac{x_{k}(t)\,\exp\!\big(\beta V_{k}(t)\Delta t+\gamma\sigma_{k}\sqrt{\Delta t}\,\xi_{k}\big)}{\sum_{j}x_{j}(t)\,\exp\!\big(\beta V_{j}(t)\Delta t+\gamma\sigma_{j}\sqrt{\Delta t}\,\xi_{j}\big)}. (3) Normalization ensures \hat{X}_{t+\Delta t}\in\Delta^{K-1} . B.3 Step 2: Innovation injection Reintroducing the innovation term in equation 1 , a forward Euler step gives X_{t+\Delta t}=(1-\mu\Delta t)\,\hat{X}_{t+\Delta t}+\mu\Delta t\,m_{t}. (4) Thus the population is a convex combination of the selection–perturbation update and the innovation distribution. B.4 Final discrete update For simulation we take \Delta t=1 , yielding x_{k}^{t+1}=(1-\mu)\,\frac{x_{k}^{t}\,\exp\!\big(\beta V_{k}^{t}+\gamma\sigma_{k}\xi_{k}^{t}\big)}{\sum_{j}x_{j}^{t}\,\exp\!\big(\beta V_{j}^{t}+\gamma\sigma_{j}\xi_{j}^{t}\big)}+\mu\,m_{k}^{t},\qquad\xi_{k}^{t}\sim\mathcal{N}(0,1). (5) Equation equation 5 is the update used in practice. It preserves positivity and the simplex by construction. B.5 Payoff dynamics Each payoff V_{k} evolves as an Ornstein–Uhlenbeck (OU) process, dV_{k}(t)=\lambda_{k}(\bar{f}_{k}-V_{k}(t))\,dt+\nu_{k}\,dB_{k}(t), (6) whose exact discretization is V_{k}^{t+1}=V_{k}^{t}e^{-\lambda_{k}\Delta t}+\bar{f}_{k}(1-e^{-\lambda_{k}\Delta t})+\nu_{k}\sqrt{\tfrac{1-e^{-2\lambda_{k}\Delta t}}{2\lambda_{k}}}\,\zeta_{k}^{t},\quad\zeta_{k}^{t}\sim\mathcal{N}(0,1). (7) B.6 Summary - Selection \to exponential weighting by payoff V_{k} , - Perturbation \to log-normal shocks via \sigma_{k}\xi_{k}^{t} , - Innovation \to convex mixing with m_{t} , - Payoffs V_{k} follow OU discretization equation 7 . Together, equation 5 and equation 7 provide the discrete-time simulator consistent with the continuous-time FinEvo SDE. Appendix C Mechanism Decomposition We aim to compute the second moment of the perturbation term under the per-strategy perturbation intensities \sigma_{k} and a global scale \gamma . With the tangent (mean-subtracted) construction used in the main text, the per-component discrete perturbation is \text{N}_{k}^{t}\;=\;\gamma\,x_{k}^{t}\Big(\sigma_{k}\,\eta_{k}^{t}\;-\;\sum_{j}x_{j}^{t}\sigma_{j}\,\eta_{j}^{t}\Big),\qquad\eta_{k}^{t}\sim\mathcal{N}(0,1)\ \text{i.i.d. across $k,t$}. We compute \mathbb{E}\!\left[(\text{N }_{k}^{t})^{2}\right]=\gamma^{2}(x_{k}^{t})^{2}\,\mathbb{E}\!\left[\Big(\sigma_{k}\,\eta_{k}^{t}\;-\;\sum_{j}x_{j}^{t}\sigma_{j}\,\eta_{j}^{t}\Big)^{2}\right]. Expanding the square and using independence with unit variance, \mathbb{E}[\eta_{i}^{t}\eta_{j}^{t}]=\delta_{ij} , we obtain \displaystyle\mathbb{E}\!\left[\Big(\sigma_{k}\,\eta_{k}^{t}\;-\;\sum_{j}x_{j}^{t}\sigma_{j}\,\eta_{j}^{t}\Big)^{2}\right] \displaystyle=\sigma_{k}^{2}\;-\;2\,\sigma_{k}\sum_{j}x_{j}^{t}\sigma_{j}\,\mathbb{E}[\eta_{k}^{t}\eta_{j}^{t}]\;+\;\sum_{j,\ell}x_{j}^{t}x_{\ell}^{t}\sigma_{j}\sigma_{\ell}\,\mathbb{E}[\eta_{j}^{t}\eta_{\ell}^{t}] \displaystyle=\sigma_{k}^{2}\;-\;2\,x_{k}^{t}\sigma_{k}^{2}\;+\;\sum_{j}(x_{j}^{t})^{2}\sigma_{j}^{2}. Therefore, \mathbb{E}\!\left[(\text{N}_{k}^{t})^{2}\right]=\gamma^{2}(x_{k}^{t})^{2}\!\left(\sigma_{k}^{2}-2x_{k}^{t}\sigma_{k}^{2}+\sum_{j}(x_{j}^{t})^{2}\sigma_{j}^{2}\right). (8) This per-strategy expression replaces the uniform- \sigma formula \sigma^{2}(x_{k}^{t})^{2}\!\left(1-2x_{k}^{t}+\|x^{t}\|_{2}^{2}\right) . It reduces to the uniform case by setting all \sigma_{k}\equiv\sigma and \gamma\equiv 1 , yielding \mathbb{E}\!\left[(\text{N}_{k}^{t})^{2}\right]=\sigma^{2}(x_{k}^{t})^{2}\!\left(1-2x_{k}^{t}+\|x^{t}\|_{2}^{2}\right). C.1 Independence and variance additivity. To justify the variance decomposition used in the mechanism analysis, we separate the discrete update into three centered (zero-mean) random fluctuation components around a deterministic drift: \Delta x_{k}^{t}=\underbrace{S_{k}^{t}}_{\text{selection fluct.}}+\underbrace{I_{k}^{t}}_{\text{innovation fluct.}}+\underbrace{N_{k}^{t}}_{\text{perturbation fluct.}}\;+\;\text{(deterministic drift)}. Only the centered fluctuations contribute to short-run variance. We work conditional on X_{t} and impose the following assumptions. Assumptions. (A1) ( Selection fluctuations ) Write the payoff signal as V_{k}^{t}=\bar{V}_{k}^{t}+\zeta_{k}^{t} , where \bar{V}_{k}^{t}:=\mathbb{E}[V_{k}^{t}\mid X_{t}] and \zeta_{k}^{t} is the centered fluctuation. Define S_{k}^{t}\;:=\;\beta x_{k}^{t}\big(\zeta_{k}^{t}-\bar{\zeta}^{t}\big),\qquad\bar{\zeta}^{t}:=\sum_{j}x_{j}^{t}\zeta_{j}^{t}, so that \mathbb{E}[S_{k}^{t}\mid X_{t}]=0 . We assume \zeta^{t} has diagonal conditional covariance \mathrm{Cov}(\zeta_{k}^{t},\zeta_{j}^{t}\mid X_{t})=\delta_{kj}\,\nu_{k}^{2}/(2\lambda_{k}) . (A2) ( Innovation fluctuations ) Let m_{t}\sim\mathrm{Dir}(\alpha) be independent of (\zeta^{t},\eta^{t}) given X_{t} , with mean m:=\mathbb{E}[m_{t}] and centered fluctuation \tilde{m}_{t}:=m_{t}-m . Define I_{k}^{t}\;:=\;\mu\,\tilde{m}_{k}^{t}, so that \mathbb{E}[I_{k}^{t}\mid X_{t}]=0 and \mathrm{Var}(I_{k}^{t}\mid X_{t})=\mu^{2}\,\tfrac{m_{k}(1-m_{k})}{\alpha_{0}+1} . (A3) ( Perturbation fluctuations ) The environmental shocks \eta^{t}=(\eta_{1}^{t},\ldots,\eta_{K}^{t})^{\top} are i.i.d. standard normal and independent of (\zeta^{t},m_{t}) given X_{t} . Define N_{k}^{t}\;:=\;\gamma\,x_{k}^{t}\Big(\sigma_{k}\eta_{k}^{t}-\sum_{j}x_{j}^{t}\sigma_{j}\eta_{j}^{t}\Big), so that \mathbb{E}[N_{k}^{t}\mid X_{t}]=0 and \mathrm{Var}(N_{k}^{t}\mid X_{t}) is given in Eq. ( 8 ). Lemma (variance additivity). Under (A1)–(A3), \mathrm{Var}(\Delta x_{k}^{t}\mid X_{t})=\mathrm{Var}(S_{k}^{t}\mid X_{t})+\mathrm{Var}(I_{k}^{t}\mid X_{t})+\mathrm{Var}(N_{k}^{t}\mid X_{t}). Proof. By construction, \mathbb{E}[S_{k}^{t}\mid X_{t}]=\mathbb{E}[I_{k}^{t}\mid X_{t}]=\mathbb{E}[N_{k}^{t}\mid X_{t}]=0 . Independence in (A2)–(A3) implies the cross-covariances vanish: \mathrm{Cov}(S_{k}^{t},I_{k}^{t}\mid X_{t})=0 , \mathrm{Cov}(S_{k}^{t},N_{k}^{t}\mid X_{t})=0 , \mathrm{Cov}(I_{k}^{t},N_{k}^{t}\mid X_{t})=0 . Therefore, \mathrm{Var}(\Delta x_{k}^{t}\mid X_{t})=\mathrm{Var}(S_{k}^{t}+I_{k}^{t}+N_{k}^{t}\mid X_{t})=\sum_{\bullet\in\{S,I,N\}}\mathrm{Var}(\bullet_{k}^{t}\mid X_{t}).\qquad\square Remarks. (i) The deterministic drift (e.g., \beta x_{k}^{t}(\bar{V}_{k}^{t}-\bar{V}^{t}) and \mu(m_{k}-x_{k}^{t}) ) affects the conditional mean but not the centered variance. (ii) If one wishes to allow correlated payoff and environment shocks, replace (A1)–(A3) by a joint Gaussian with block covariance; the variance then includes explicit cross terms \mathrm{Cov}(S_{k}^{t},N_{k}^{t}\mid X_{t}) , etc. Our empirical results use (A1)–(A3), which are standard in evolutionary/replicator SDEs and lead to the closed forms reported in the main text. C.2 Parameter summary. Symbol Description Typical usage/role x_{k}^{t} Proportion of agents using \pi_{k} at time t \sum_{k}x_{k}^{t}=1 f_{k}^{t} Empirical mean payoff of strategy k at t Observed fitness \beta Selection intensity Tuning parameter \gamma_{kj} Ecological effect of j on k Set to 0 in baseline; >0 competition \sigma_{k} perturbation intensity for strategy k Calibrated or fixed \eta_{k}^{t} Standard normal noise Simulation draw \mu Innovation probability Tuning parameter m_{k} Innovation distribution User-defined \Delta t Time step (usually 1) Simulation round C.3 Conclusion. This derivation demonstrates that the agent-level elimination, selection, and replacement mechanism implemented in our simulation is a numerically consistent discretization of the generalized Lotka–Volterra stochastic ecology, augmented with innovation. The result is a unified framework with theoretical foundations in evolutionary dynamics and practical applicability for complex agent-based market studies. Appendix D Variance Decomposition under Correlated Payoff and Environmental Shocks In this section, we extend the variance analysis in Appendix C.1 to the case where payoff shocks and environmental shocks are correlated. We emphasize that correlation does not alter the FinEvo SDE itself; only the decomposition gains additional covariance terms. FinEvo SDE dX_{t}=\underbrace{\beta\,X_{t}\odot(V_{t}-\bar{V}_{t})}_{\text{Selection}}\,dt\;+\;\underbrace{\mu(m_{t}-X_{t})}_{\text{Innovation}}\,dt\;+\;\underbrace{\sigma\,P(X_{t})\,dB_{t}}_{\text{Perturbation}}, (C.1) where P(X_{t}) is the tangent-space projection ensuring simplex invariance, and B_{t}\in\mathbb{R}^{K} is a vector-valued Brownian motion. D.1 Allowing Correlated Shocks We now assume the Brownian increments satisfy \mathrm{Cov}(dB_{t})\;=\;\Sigma\,dt, (9) where \Sigma is a symmetric positive semidefinite correlation matrix: \Sigma=\begin{pmatrix}1&\rho_{12}&\cdots&\rho_{1K}\\ \rho_{21}&1&\cdots&\rho_{2K}\\ \vdots&&\ddots&\vdots\\ \rho_{K1}&\cdots&\rho_{K(K-1)}&1\end{pmatrix}. This relaxation affects only the variance decomposition, not the SDE dynamics. D.2 Variance of a Single Component For any strategy population component X_{k,t} , Itô’s lemma yields: \displaystyle\mathrm{Var}(dX_{k,t}) \displaystyle=\underbrace{\mathrm{Var}\!\left[\beta X_{k,t}(V_{k,t}-\bar{V}_{t})\right]}_{\text{Selection}}\;+\;\underbrace{\mathrm{Var}\!\left[\mu(m_{k,t}-X_{k,t})\right]}_{\text{Innovation}} \displaystyle\quad+\;\underbrace{\sigma^{2}\left(P\Sigma P^{\top}\right)_{kk}}_{\text{Perturbation}}\;+\;\underbrace{2\sigma\,\mathrm{Cov}\!\left(\beta X_{k,t}(V_{k,t}-\bar{V}_{t}),\;(P\,dB_{t})_{k}\right)}_{\text{Selection--Perturbation Cross Term}}. (C.2) D.3 Full Matrix Variance For the full vector process X_{t} , we have: \displaystyle\mathrm{Var}(dX_{t}) \displaystyle=\beta^{2}\,\mathrm{Var}\!\left[X_{t}\odot(V_{t}-\bar{V}_{t})\right]\;+\;\mu^{2}\,\mathrm{Var}(m_{t}-X_{t}) \displaystyle\quad+\;\sigma^{2}\,P\Sigma P^{\top}\;+\;2\beta\sigma\;\mathrm{Cov}\!\left(X_{t}\odot(V_{t}-\bar{V}_{t}),\;P\,dB_{t}\right). (C.3) D.4 Relation to the Uncorrelated Case Under independence assumptions (A1–A3) in Appendix C.1, we have \Sigma=I,\qquad\mathrm{Cov}\big(\text{selection},\;dB_{t}\big)=0, and Equations (C.2)–(C.3) reduce to the clean additive decomposition in the main text. When shocks are correlated, the FinEvo SDE itself remains unchanged, and only additional covariance terms appear in the variance decomposition, consistent with the remark in Appendix C.1. Appendix E Minimal theoretical guarantees E.1 Continuous-time FinEvo SDE (for reference). We explicitly write the continuous-time dynamics underlying our discrete updates: \displaystyle dX_{i} \displaystyle=X_{i}\!\left(f_{i}(X,t)-\sum_{j}X_{j}f_{j}(X,t)\right)\,dt (10) \displaystyle+\mu\,(m_{i}-X_{i})\,dt \displaystyle+\sigma_{i}X_{i}\!\left(dW_{i}-\sum_{j}X_{j}\,dW_{j}\right),\qquad i=1,\dots,K. where X(t)\in\Delta^{K}:=\{x\in\mathbb{R}_{\geq 0}^{K}:\sum_{i}x_{i}=1\} are strategy shares, m\in\mathrm{int}(\Delta^{K}) with \min_{i}m_{i}>0 , \mu\geq 0 , and \{W_{i}\} are independent Brownian motions. The diffusion is tangent to \Delta^{K} . E.2 Assumptions. (i) f_{i}(\cdot,t) are locally Lipschitz with linear-growth bounds in X , uniformly on compact time sets; (ii) m\in\mathrm{int}(\Delta^{K}) , \mu\geq 0 , \sigma_{i}\geq 0 ; (iii) Brownian motions \{W_{i}\} are independent. When we couple with exogenous value processes V_{i}(t) , we assume V_{i} are OU with unique stationary laws and bounded moments; the discrete OU approximation is Eq. equation 7 . Lemma 1 (Well-posedness) Under the above assumptions, the SDE equation 10 admits a unique strong solution on [0,\infty) for any X(0)\in\Delta^{K} . Proposition 1 (Simplex invariance and positivity) Let X(0)\in\Delta^{K} . Then the solution to equation 10 satisfies \sum_{i}X_{i}(t)\equiv 1 for all t\geq 0 and X_{i}(t)\geq 0 almost surely. If, in addition, X(0)\in\mathrm{int}(\Delta^{K}) , \mu>0 , and \min_{i}m_{i}>0 , then X_{i}(t)>0 for all t>0 almost surely (the boundary is non-absorbing). Proof sketch. Summing equation 10 over i cancels both the selection drift and the tangent diffusion; the innovation term sums to \mu(\sum_{i}m_{i}-\sum_{i}X_{i})=0 , hence mass is conserved. Multiplicative diffusion vanishes on the boundary, implying non-negativity. With \mu>0 and \min_{i}m_{i}>0 , the drift \mu(m_{i}-X_{i}) points strictly inward. \square Proposition 2 (Invariant measure and quasi-stability) Consider the joint Markov process (X(t),V(t)) with X governed by equation 10 and V OU. If \mu>0 and \min_{i}m_{i}>0 , then at least one invariant probability measure exists on \Delta^{K}\times\mathbb{R}^{K} . Moreover, in the small-perturbation/weak-innovation regime ( \sigma\to 0 , \mu\to 0 with \mu=O(\sigma^{2}) ), invariant measures concentrate near the attractor set of the deterministic replicator–mutator ODE \dot{X}_{i}=X_{i}\,(f_{i}-\sum_{j}X_{j}f_{j})+\mu(m_{i}-X_{i}) , yielding quasi-stable ecological compositions. Proof sketch. OU has a unique stationary law with bounded moments. By Prop. 1 , X(t) remains in the compact simplex; the joint process is Feller and tight, so Krylov–Bogoliubov yields existence of an invariant measure. Small-perturbation concentration follows from standard Freidlin–Wentzell arguments for coupled systems. \square E.3 Discrete-time counterpart (consistency). Our simulation update implements exponential selection followed by innovation/mutation. The selection-normalization step is Eq. equation 3 , and the innovation mixing is the final update Eq. equation 5 . Together, these define a positivity-preserving, simplex-invariant map. In particular, \sum_{k}x_{k}^{t+1}=1,\quad x_{k}^{t+1}\geq 0,\quad x_{k}^{t+1}>0\ \text{for }t\geq 1\ \text{if }\mu>0\ \text{and}\ \min_{k}m_{k}>0. Coupled value signals use the OU discretization in Eq. equation 7 . Under \Delta t\to 0 , the discrete scheme is a first-order consistent approximation to Eq. equation 10 . Appendix F Agent Library and Initialization. F.1 Agent Name Mapping Table F.1 summarizes the mapping between full agent names and their abbreviations used throughout the paper. Agents are organized by category (rule-based, informed, reinforcement learning, time-series models, and LLM-based traders). This mapping provides a concise reference to ensure consistency across figures, tables, and analyses in the main text. Table 7: Agent Name Mapping Category Full Name Abbreviation Technical Traders MomentumTrader MT RSITrader RSI BreakoutTrader BT Informed Traders ConservativeInformedTrader CIT AggressiveInformedTrader AIT SpreadAwareInformedTrader SAIT ScaledInformedTrader SIT OpportunisticInformedTrader OIT Reinforcement Learning DoubleDQNTrader DoubleDQN PPO_Trader PPO A2C_Trader A2C Time-Series Models iTransformer_Trader ITT Informer_Trader IT TimeMixer_Trader TimeMixer LLM-based Agents LLM_ReAct_Trader ReAct LLM_CoT_Trader CoT LLM_Trading_Trader Trading LLM_FinCon_Trader FinCon F.2 Agent library This section describes the full design of each agent archetype in our framework, including their trading objectives, information sources, decision rules, and execution mechanisms. By detailing both classical strategies (e.g., technical and informed traders) and modern adaptive agents (e.g., deep learning, reinforcement learning, and LLM-based models), the agent library highlights the heterogeneity of behaviors underlying FinEvo’s ecological simulations. F.2.1 Informational Traders Agent • Trading Objective Exploit discrepancies between the market price P_{t} and an oracle‐provided fundamental value P^{*}_{t} to generate profitable trades without long‐term directional bias. • Signal Computation At each time t , compute the relative deviation \delta_{t}=\frac{P_{t}-P^{*}_{t}}{P^{*}_{t}}\,,\quad P_{t}=\frac{P^{b}_{t}+P^{a}_{t}}{2} • Decision Rule Given a deviation threshold \theta>0 , the agent’s trading decision a_{t}^{i}\in\{\text{buy},\text{sell},\text{hold}\} is defined as: a_{t}^{i}=\begin{cases}\text{buy},&\text{if }\delta_{t}<-\theta\text{ and }h_{t}^{i}\leq 0\\ \text{sell},&\text{if }\delta_{t}>\theta\text{ and }h_{t}^{i}>0\\ \text{hold},&\text{otherwise}\end{cases} where h_{t}^{i} denotes agent i ’s current inventory position at time t . • Execution Mechanism Depending on variant, the agent submits orders as follows: – Conservative : place limit orders at P^{\text{buy}}=P^{b}_{t}+\epsilon,\quad P^{\text{sell}}=P^{a}_{t}-\epsilon – Aggressive : place market orders immediately at the best ask/bid. – Opportunistic : place market orders of size q_{t}=\begin{cases}\lambda\,q,&\text{if liquidity is low}\\ q,&\text{otherwise}\end{cases} where q is the base trade quantity and \lambda>1 is an amplification factor. F.2.2 Market Maker Agent • Trading Objective Maintain liquidity in the limit order book by continuously submitting limit orders on both bid and ask sides across multiple price levels, aiming to profit from the bid-ask spread. • Signal Computation At each decision time t , observe the best bid and ask prices: P^{b}_{t},\quad P^{a}_{t} Calculate the mid-price and half-spread: P^{m}_{t}=\frac{P^{b}_{t}+P^{a}_{t}}{2},\quad s_{t}=\frac{P^{a}_{t}-P^{b}_{t}}{2} Sample the total order size per side from a predefined range: q_{t}\sim[q_{\min},q_{\max}] Determine the number of price levels L to quote on each side, sampled randomly. • Decision Rule For each level i=0,1,\dots,L-1 , compute limit order prices: P^{b}_{t,i}=P^{m}_{t}-s_{t}-i,\quad P^{a}_{t,i}=P^{m}_{t}+s_{t}+i Allocate order sizes at each level according to weight vector w=(w_{1},\dots,w_{L}) , satisfying \sum_{i=1}^{L}w_{i}=1,\quad q_{t,i}=w_{i}\cdot q_{t} • Execution Mechanism – Polling mode (subscribe=False) : Query current market spread and place symmetric limit orders around the mid-price with uniform size at each level. – Subscription mode (subscribe=True) : Subscribe to order book updates, randomly select L\in\{1,\ldots,5\} levels, and allocate volumes per the predefined weight vector w from the levels quote dictionary. F.2.3 Noise Trader Agent • Trading Objective Operates without access to fundamental information or technical indicators. Makes randomized decisions to simulate irrational or uninformed trading behavior. • Decision Rule at time t With equal probability p=0.5 , the agent executes one of two actions: – Aggressive Execution (Market Order) : Submits a market buy or sell order with equal probability: q_{t}\sim\mathcal{U}\{10,20,...,100\},\quad a_{t}=\begin{cases}\text{MarketBuy}(q_{t}),&\text{prob. }0.5\\ \text{MarketSell}(q_{t}),&\text{prob. }0.5\end{cases} – Passive Execution (Limit Order) : Submits a limit buy or sell order offset from the mid-price: P^{m}_{t}=\frac{P^{b}_{t}+P^{a}_{t}}{2},\quad\epsilon_{t}\sim\mathcal{U}(0.001,0.005) P^{\text{limit}}_{t}=\begin{cases}\lfloor P^{m}_{t}(1-\epsilon_{t})\rfloor,&\text{if buy}\\ \lceil P^{m}_{t}(1+\epsilon_{t})\rceil,&\text{if sell}\end{cases} a_{t}=\begin{cases}\text{LimitBuy}(q_{t},P^{\text{limit}}_{t}),&\text{prob. }0.5\\ \text{LimitSell}(q_{t},P^{\text{limit}}_{t}),&\text{prob. }0.5\end{cases} • Next Decision Time The next wake-up time is sampled from: t_{\text{next}}=t+\Delta t,\quad\Delta t\sim\mathcal{U}(55,65)\text{ seconds} F.2.4 Event-Driven News Reaction Agent • Trading Objective React only to significant market news events (e.g., macroeconomic shocks or asset-specific ratings) by placing immediate market orders in response to the sentiment direction of the news. • Information Source Subscribes to an EventBus that broadcasts discrete news events. Each event includes: – event_type \in { POSITIVE , NEGATIVE , UPGRADE , DOWNGRADE , …} – content : textual description (ignored by logic) • Decision Logic on Event Reception At time t , when a news event e_{t} is received: \text{Action}_{t}=\begin{cases}\text{MarketBuy}(q),&\text{if }\texttt{event\_type}\in\{\texttt{POSITIVE},\texttt{UPGRADE}\}\\ \text{MarketSell}(q),&\text{if }\texttt{event\_type}\in\{\texttt{NEGATIVE},\texttt{DOWNGRADE}\}\\ \text{NoAction},&\text{otherwise}\end{cases} where q is the fixed order quantity (e.g., q=10 shares). • Timing and Wake-up Frequency The agent wakes up every \Delta t=10 seconds to maintain kernel activity and ensure one-time subscription to the event stream, but trading decisions are purely event-driven. • Execution Mechanism All trades are submitted as market orders immediately upon signal reception to ensure fastest reaction to major events. F.2.5 Reinforcement Learning Trader Agent • Trading Objective The RL-based agent aims to maximize cumulative portfolio returns by interacting with the environment in an online manner. At each time step, it observes the current market state, selects an action (buy/sell/hold), and updates its policy parameters through continual learning (online fine-tuning). • State Representation The market state s_{t} is represented as follows for different learning schemes: \begin{array}[]{ll}\textbf{Q-Learning:}&s_{t}=(x_{t}^{(1)},\ x_{t}^{(2)},\ x_{t}^{(3)})\ \text{(discretized)}\\ \textbf{PPO:}&s_{t}=\left[x_{t}^{(1)},\ x_{t}^{(2)},\ \ldots,\ x_{t}^{(n)}\right]\ \text{(continuous and normalized)}\end{array} 1 • Action Policy The agent selects actions from a discrete space: \mathcal{A}=\{0=\text{HOLD},\ 1=\text{BUY},\ 2=\text{SELL}\} Each action triggers a market order of fixed quantity Q in the corresponding direction. • Reward Function The agent is rewarded based on changes in portfolio value: r_{t}=PV_{t}-PV_{t-1}-\lambda\cdot\mathbb{I}(a_{t}\in\{\text{BUY},\text{SELL}\}) where PV_{t} is portfolio value and \lambda is a trade penalty coefficient. • Learning Mechanism 1. Q-Learning (Tabular, Online) The agent maintains a Q-table Q(s,a) and updates it using the standard Bellman equation: Q(s_{t},a_{t})\leftarrow Q(s_{t},a_{t})+\alpha\left[r_{t}+\gamma\max_{a^{\prime}}Q(s_{t+1},a^{\prime})-Q(s_{t},a_{t})\right] Parameters (\alpha,\gamma) denote the learning rate and discount factor, respectively. Updates occur online every N steps from recent experience buffer D . 2. Proximal Policy Optimization (PPO) The agent learns a stochastic policy \pi_{\theta}(a|s) using clipped surrogate objective: L^{\text{PPO}}=\mathbb{E}_{t}\left[\min\left(r_{t}(\theta)A_{t},\ \text{clip}(r_{t}(\theta),1-\epsilon,1+\epsilon)A_{t}\right)\right] where r_{t}(\theta)=\frac{\pi_{\theta}(a_{t}|s_{t})}{\pi_{\theta_{\text{old}}}(a_{t}|s_{t})} is the probability ratio, and A_{t} is the advantage function. The agent collects N steps of interaction and performs K epochs of gradient updates for online fine-tuning. F.2.6 Technical Trader Agent • Trading Objective Execute market orders based on common technical analysis signals derived from historical price data, aiming to capture trends or mean reversion. • State Observation At each decision time t , the agent observes a recent price window: S_{t}=\{P_{t-n},\ldots,P_{t}\} where P_{t} is the latest trade price. • Trading Quantity Fixed order size Q for all trades. • Trading Strategies and Decision Rules The agent implements one of the following strategies: 1. Momentum Agent (Moving Average Crossover) Compute short-term and long-term moving averages: MA^{short}_{t}=\frac{1}{N_{s}}\sum_{i=t-N_{s}+1}^{t}P_{i},\quad MA^{long}_{t}=\frac{1}{N_{l}}\sum_{i=t-N_{l}+1}^{t}P_{i} The trading action a_{t} is: a_{t}=\begin{cases}\text{BUY},&\text{if }MA^{short}_{t}>MA^{long}_{t}\text{ and position}\leq 0\\ \text{SELL},&\text{if }MA^{short}_{t}\theta_{\text{high}}\\ \text{HOLD},&\text{otherwise}\end{cases} Market order of size Q placed in the direction of a_{t} (buy low RSI, sell high RSI). 3. Breakout Agent (Channel Breakout) Define breakout levels: R_{t}=\max(P_{t-L},\ldots,P_{t-1}),\quad S_{t}=\min(P_{t-L},\ldots,P_{t-1}) The trading action a_{t} is: a_{t}=\begin{cases}\text{BUY},&P_{t}>R_{t}\\ \text{SELL},&P_{t}{} 12%). In the MI network, node size reflects the dominance of a strategy and edge thickness encodes co-dependence strength. Figure 15: Intraday evolution of co-occurrence matrices on July 7, 2025 (dominance threshold > 12%). Early in the day ( T{=}0\% ) the matrix is near-empty; mid-day ( T{=}50\% ) a few pairs begin to co-dominate; by the close ( T{=}100\% ) the structure consolidates around a small number of persistent alliances while many strategies remain peripheral. Figure 16: Intraday evolution of the mutual-information network on July 7, 2025. Node size indicates dominance; edge thickness indicates co-dependence strength. The network transitions from a sparse, weakly connected configuration in the morning to a hub-centric structure by mid-day, followed by selective pruning and stabilization toward the close. Detailed analysis. (Early, T{=}0\%-25\% ) Co-occurrence is negligible and the MI network is sparse, indicating a fragmented ecology with weak coordination. Most strategies operate independently. (Mid-day, T{=}50\% ) The first co-dominant pairs emerge in the co-occurrence matrix, and the MI network becomes hub-centric : one or two strategies act as connectors linking otherwise disjoint clusters. This marks a regime shift from exploratory behavior to coordinated specialization. (Late, T{=}75\%-100\% ) A small set of alliances persists and strengthens, while peripheral strategies decouple. The MI network shows selective pruning : superfluous links vanish and high-weight ties concentrate around the main hub(s). This coincides with the end-of-day re-concentration seen in population shares. Takeaways. Across the day, the ecology moves from diffuse (low co-occurrence, sparse MI) to organized (stable alliances, hub-centric MI), then stabilizes toward the close. This intraday reorganization is consistent with our macro findings: exploration predominates early, while selection gradually consolidates alliances into a small, persistent backbone. G.2 Additional Intraday Trajectories Across Multiple Real-Market Days In the main text (Figure 6), we evaluated FinEvo on two long real-market periods: January–October 2022 (bear regime) and October 2024–July 2025 (bull regime). These experiments already aggregate ecosystem statistics over many months of real data. However, due to space constraints, only a single representative intraday trajectory was shown in the main text. Therefore, we provide additional intraday population-evolution trajectories from multiple dates across both bull and bear markets. These plots demonstrate that FinEvo’s ecological dynamics are not specific to a single trading day: dominance shifts, strategy turnover, and phase-change patterns remain consistent across a wide range of trading conditions. Figure 17: Intraday population evolution on 2025-07-02, 07-05, and 07-06. Figure 18: Intraday population evolution on 2025-07-08, 07-09, and 07-10. These results confirm that FinEvo’s real-data performance is robust across days, months, and market regimes , G.3 Robustness to Perturbations of Initial Agent Parameters To assess whether FinEvo’s ecological dynamics depend sensitively on the initial agent-level hyperparameters, we conducted a Perturbed-Parameter robustness experiment. For each Monte Carlo run, all key hyperparameters (e.g., RL learning rates, window lengths for rule-based agents, and LLM prompting parameters) were randomly perturbed by \pm 20\% around their baseline values, sampled from a uniform distribution. This produces substantially different heterogeneous initializations for every simulation. Table 8: Robustness under perturbed agent parameters (128 runs). Values report mean \pm 95% CI. Stylized-fact metrics remain stable under \pm 20\% perturbations. Regime Model Ex. Kurtosis {}_{\text{mean}} \uparrow Skewness {}_{\text{mean}} \downarrow Sharpe Disp. {}_{\text{mean}} \uparrow Vol-of-Vol {}_{\text{mean}} \uparrow Bullish FinEvo 3.163\pm 0.138 -1.532\pm 0.129 1.669\pm 0.061 0.008\pm 0.0004 FinEvo (Perturbed) 3.141\pm 0.152 -1.596\pm 0.138 1.638\pm 0.067 0.008\pm 0.0004 Bearish FinEvo 4.259\pm 0.172 -1.793\pm 0.149 1.732\pm 0.089 0.013\pm 0.0007 FinEvo (Perturbed) 4.196\pm 0.181 -1.692\pm 0.157 1.684\pm 0.094 0.012\pm 0.0008 Across both bull and bear regimes, the perturbed-parameter version closely matches the baseline across all stylized-fact metrics. This confirms that heavy tails, downside asymmetry, Sharpe dispersion, and volatility clustering are stable emergent properties of FinEvo’s evolutionary dynamics , not artifacts of specific hyperparameter choices. G.4 Robustness to Correlated, Heavy-Tailed Shocks In the main text, our analytical results are derived under the standard assumption of independent Gaussian shocks. To verify that our empirical findings are not an artifact of this simplification, we conduct an additional robustness experiment in which payoff and environmental shocks are both correlated and heavy-tailed . Concretely, we modify the noise driving the value processes V_{k}(t) as follows. At each time step t , instead of drawing independent standard normal increments, we construct strategy-level shocks \epsilon_{k}(t)\;=\;\rho\,Z_{\text{common}}(t)\;+\;\sqrt{1-\rho^{2}}\,Z_{k,\text{idio}}(t), (11) where \rho\in(0,1) controls the correlation strength (we use \rho=0.5 in our experiments), Z_{\text{common}}(t) is a common macro factor shared by all strategies, and Z_{k,\text{idio}}(t) is an idiosyncratic component specific to strategy k . Both Z_{\text{common}}(t) and Z_{k,\text{idio}}(t) are drawn from a standardized Student- t distribution with \nu degrees of freedom (we use \nu=5 ), rescaled to have unit variance. This yields positively correlated, heavy-tailed shocks at the strategy level: \text{Cov}\bigl(\epsilon_{k}(t),\epsilon_{\ell}(t)\bigr)\;=\;\rho^{2}\quad(k\neq\ell), (12) in contrast to the independent Gaussian setting in the main analysis. We then drive the value process of each strategy with these correlated heavy-tailed shocks: dV_{k}(t)\;=\;\lambda_{k}\bigl(m_{t}-V_{k}(t)\bigr)\,dt\;+\;\nu_{k}\,\epsilon_{k}(t)\,\sqrt{dt}, (13) and, for the environmental shock, we take \eta_{t}\;=\;Z_{\text{common}}(t), (14) so that strategy-level payoffs are explicitly correlated both with each other and with the environment shock. We rerun the main set of simulations under this correlated, heavy-tailed specification using the same market configurations as in Section 3 . The resulting markets continue to exhibit (i) fat-tailed and negatively skewed PnL distributions, (ii) pronounced volatility clustering and elevated volatility-of-volatility, and (iii) similar ecological phase changes in the strategy population. Quantitatively, the stylized-fact metrics change only moderately; a detailed comparison is reported in Table 9 . These results confirm that the independence/normality assumptions are primarily used for analytical tractability in the SDE derivation and are not critical for the qualitative behavior of FinEvo. Table 9: Robustness to correlated, heavy-tailed shocks (128 runs). Values report mean \pm 95% CI. Regime Model Ex. Kurtosis {}_{\text{mean}} \uparrow Skewness {}_{\text{mean}} \downarrow Sharpe Disp. {}_{\text{mean}} \uparrow Vol-of-Vol {}_{\text{mean}} \uparrow Bullish FinEvo 3.163\pm 0.138 -1.532\pm 0.129 1.669\pm 0.061 0.008\pm 0.0004 FinEvo (Correlated) 5.278\pm 0.151 -1.632\pm 0.133 1.727\pm 0.075 0.011\pm 0.0007 Bearish FinEvo 4.259\pm 0.172 -1.793\pm 0.149 1.732\pm 0.089 0.013\pm 0.0007 FinEvo (Correlated) 6.399\pm 0.214 -1.932\pm 0.175 1.844\pm 0.101 0.015\pm 0.0009 G.5 Cross-Asset Robustness Setup. FinEvo is designed to be asset-agnostic: the evolutionary SDE and CDA mechanism do not rely on any asset-specific assumption. To examine robustness across heterogeneous markets, we instantiate the same FinEvo ecology under four widely studied equities—AAPL, TSLA, MSFT, and XOM. Across all assets, the evolutionary parameters (\beta,\mu,\gamma) , agent library, value-learning dynamics, and market microstructure remain identical . The only asset-dependent components are the empirical volatility scale and the asset-specific news schedule used to parameterize exogenous shocks. Thus, each asset represents the same evolutionary system under different empirical environments , rather than a separately calibrated model. Results. Table 10 reports the mean and 95% confidence intervals of four stylized-fact metrics of strategy PnL distributions. Across all assets, FinEvo consistently produces positive excess kurtosis (fat tails), negative skewness (downside asymmetry), elevated Vol-of-Vol (volatility clustering), and substantial Sharpe-ratio dispersion (heterogeneous strategy performance). The magnitudes also vary in economically intuitive ways: TSLA exhibits the strongest heavy tails, skewness, Sharpe-dispersion, and Vol-of-Vol; AAPL lies in the intermediate range; while MSFT and XOM display more moderate tails and lower regime-switching intensity. These findings demonstrate that FinEvo’s evolutionary dynamics are robust across heterogeneous assets without any per-asset tuning. All assets are evaluated using the same empirical bear-market period as in Section 3.3 (January–October 2022), so that cross-asset differences arise only from volatility scales and news patterns rather than different market regimes. Table 10: Cross-asset robustness of FinEvo. Metrics are reported as mean \pm 95% CI over 128 runs. Asset Ex. Kurtosis Skewness Sharpe Disp. Vol-of-Vol AAPL 4.354\pm 0.41 -1.656\pm 0.18 1.633\pm 0.12 0.011\pm 0.0015 TSLA \mathbf{5.578\pm 0.53} \mathbf{-2.344\pm 0.24} \mathbf{1.975\pm 0.18} \mathbf{0.015\pm 0.0023} MSFT 3.176\pm 0.35 -0.967\pm 0.15 1.158\pm 0.09 0.008\pm 0.0016 XOM 3.741\pm 0.38 -0.573\pm 0.12 1.392\pm 0.17 0.008\pm 0.0011 In our experiments, the same FinEvo configuration produces qualitatively similar eco- logical signatures across AAPL, TSLA, MSFT, and XOM (Table 10 ), suggesting that the frame- work is not tightly tied to any single underlying asset. While we only test equities here, the evolu- tionary SDE and CDA mechanism do not rely on equity-specific assumptions, so in principle the same formalism could be instantiated for other electronically traded markets (e.g., futures, FX, cryp- tocurrencies) with appropriate data and calibration. G.6 Experiment Configuration Price Matching Mechanism (Continuous Double Auction) Order book and priority. We adopt a continuous double auction with price–time priority. At time t , let the best bid/ask be b_{t} and a_{t} with spread s_{t}=a_{t}-b_{t} . Depth at level \ell on the bid/ask sides are B_{t}(\ell),A_{t}(\ell) , tick size \delta p . Orders are queued FIFO within price. Order types. Agents submit limit orders (p,q) (positive q for buy, negative for sell) or market orders ( p=\pm\infty ). A limit order is marketable if p\geq a_{t} (buy) or p\leq b_{t} (sell). Matching rule and fills. Incoming marketable quantity q>0 (buy) consumes the ask ladder \{(A_{t}(1),a_{t}),(A_{t}(2),a_{t}+\delta p),\dots\} until filled or book exhausted. If the order walks L levels with per-level trades v_{\ell} at prices p_{\ell} , the volume-weighted execution price (VWAP) is \overline{p}_{t}\;=\;\frac{\sum_{\ell=1}^{L}v_{\ell}\,p_{\ell}}{\sum_{\ell=1}^{L}v_{\ell}},\qquad\sum_{\ell=1}^{L}v_{\ell}\;=\;q_{\mathrm{filled}}\leq q. Any unfilled remainder \Delta q=q-q_{\mathrm{filled}} (for a limit order) is posted at its limit price p on the respective side with timestamp t . The last trade price P_{t} is set to the price of the final match within the event; the book then updates depths and best quotes. Mid/micro-price and slippage. Define mid-price m_{t}=\frac{a_{t}+b_{t}}{2} and a depth-weighted micro-price \tilde{m}_{t}=\frac{A_{t}(1)\,b_{t}+B_{t}(1)\,a_{t}}{A_{t}(1)+B_{t}(1)} . For a buy, the instantaneous slippage (vs. pre-trade mid) is \mathrm{Slip}_{t}^{\mathrm{buy}}\;=\;\overline{p}_{t}-m_{t^{-}},\qquad\mathrm{Slip}_{t}^{\mathrm{sell}}\;=\;m_{t^{-}}-\overline{p}_{t}. Trading frictions (costs). We account for realistic frictions via a per-trade proportional cost c=0.5\% (including slippage/impact setting). For a trade of notional N_{t}=|\overline{p}_{t}\,q_{\mathrm{filled}}| , the fee component is c\,N_{t} . Cash, inventory, and mark-to-market. For agent k , let inventory be x_{k,t} and cash be C_{k,t} . A filled buy of size q>0 at \overline{p}_{t} updates C_{k,t}\leftarrow C_{k,t}-\overline{p}_{t}\,q-c\,|\overline{p}_{t}q|,\qquad x_{k,t}\leftarrow x_{k,t^{-}}+q. A sell symmetrically yields C_{k,t}\leftarrow C_{k,t}+\overline{p}_{t}\,|q|-c\,|\overline{p}_{t}q|,\ x_{k,t}\leftarrow x_{k,t^{-}}-|q| . Mark-to-market portfolio value at event-time t is V_{k,t}\;=\;C_{k,t}+x_{k,t}\,m_{t}. PnL and return. Incremental PnL is \Delta\mathrm{PnL}_{k,t}=V_{k,t}-V_{k,t^{-}} . Cumulative return over a session is \mathrm{Ret}_{k}\;=\;\frac{V_{k,T}-V_{k,0}}{V_{k,0}}. We report \mathrm{Ret}_{k} with a 95% confidence interval (CI) across runs: \mathrm{CI}_{95}=\bar{r}\pm 1.96\,\hat{\sigma}/\sqrt{n} . Sharpe, drawdown, turnover. Using event- or bar-level returns \{r_{k,t}\} , the (annualized) Sharpe is \mathrm{Sharpe}_{k}\;=\;\frac{\mu(r_{k,t})}{\sigma(r_{k,t})}\,\sqrt{A}, with A the annualization factor (e.g., bars-per-year). Max drawdown is \mathrm{MDD}_{k}=\max_{t\leq u}\bigl(1-\frac{V_{k,u}}{\max_{s\leq t}V_{k,s}}\bigr) . Turnover over the session is measured as traded notional over average equity: \mathrm{Turnover}_{k}\;=\;\frac{\sum_{t}|\overline{p}_{t}\,q_{k,t}|}{\frac{1}{T}\sum_{t}V_{k,t}}. Interpretation. • Price–time priority ensures fair queueing: better prices fill first; within a price level, earlier timestamps precede later ones. • VWAP execution captures multi-level consumption and embeds microstructure slippage relative to m_{t^{-}} . • 0.5% cost per trade penalizes overtrading and brings profitability closer to realistic conditions. • Mark-to-market PnL at mid links inventory risk to contemporaneous liquidity via m_{t} (or \tilde{m}_{t} if desired). • Sharpe/MDD/Turnover jointly describe risk-adjusted performance, path risk, and trading intensity, complementing raw returns. G.6.1 Experiment Setup All experiments are conducted using the simulation market engine, which provides a continuous double auction market with realistic order book dynamics (see Appendix G.6 for details of the price matching mechanism). We extend the environment with three key components to capture eco-evolutionary dynamics: 1. News Release Line. External information shocks are injected via a news-release channel that delivers common signals to all agents at scheduled times. Shocks can be positive or negative and directly affect agent beliefs and trading aggressiveness. 2. Multi-Metric Distribution Mechanism. Beyond simple price updates, the simulation tracks and redistributes capital based on multiple metrics including realized PnL, volatility-adjusted returns, and strategy diversity. This allows performance pressure, innovation, and environmental perturbation to jointly influence ecological adaptation. 3. Evolutionary Rules. We implement the replicator–mutator dynamics introduced in Section 2. At each epoch (set to 5 minutes of simulated time), agent population shares are updated according to selection strength \beta , mutation rate \mu , and perturbation intensity \sigma . Capital is then reallocated consistently with these updated proportions, ensuring both ecological realism and conservation of market value. Unless otherwise specified, we run each configuration with 128 Monte Carlo seeds with a 8-hour trading session per day. Transaction costs (including slippage) are set to 0.5\% per trade. All agents start with equal initial capital and interact in the same synthetic market environment. G.6.2 Agent Parameters For clarity, we summarize the design of the infrastructure agents used in our experiments. Since the trading logic of each agent type has already been detailed in Appendix F.2.1 , here we only describe their functional roles and key configuration parameters. DataCenterAgent. Acts as the unique global data hub and indicator engine. It collects trades and liquidity snapshots from the exchange, computes a wide set of microstructure and technical indicators (e.g., bid–ask spread, order book imbalance, RSI, MACD, ATR, Bollinger Bands), and maintains historical records via bounded deques to ensure memory efficiency. Key parameters include: • wakeup frequency: 5s (sampling interval for updates). • max history length: 200 ticks. • minimum history for service: 30 ticks. • output: indicator_history.pkl and liquidity_history.pkl. EventBusAgent. Implements a centralized publish/subscribe hub. It manages subscriptions from other agents and forwards broadcast events (e.g., news). The design is lightweight: subscribers are stored in a set, ensuring de-duplication, and events are distributed in real time to all registered agents. NewsAgent. Publishes exogenous events (e.g., news shocks) according to a pre-specified schedule. At each scheduled time, the event is broadcast via the EventBusAgent. The schedule is defined as a sorted list of (t,\text{event}) pairs, ensuring reproducibility and precise timing of information shocks. NoiseTraderAgent. A baseline stochastic trader that injects random liquidity into the market. At each wakeup, it places either a market order (immediate execution) or a limit order (near the mid-price with small random offsets), each with 50% probability. Key parameters: traded symbol , random order size (10–100 shares), wakeup interval (55–65 seconds, uniformly drawn). MarketMakerAgent. Provides liquidity symmetrically on both sides of the order book. At each wakeup, it cancels its resting orders and posts new ones across multiple price levels, ensuring spread coverage. It can operate in two modes: (i) polling mode (places orders around mid-price using current spread), (ii) subscription mode (subscribes to order book updates and allocates volumes according to a predefined level-quote distribution). Key parameters: initial cash = 10M, order size range = [10, 50], wakeup frequency = 5s, number of levels = 1–5, order size split follows DEFAULT_LEVELS_QUOTE_DICT. NewsReactionAgent. An event-driven trader that reacts to exogenous news released by the NewsAgent via the EventBus. Upon receiving a broadcast, it submits a market order consistent with the sentiment (buy if POSITIVE/UPGRADE , sell if NEGATIVE/DOWNGRADE ). Key parameters: initial cash = 1M, trade size = 10 shares per reaction, wakeup frequency = 10s (used only for subscription maintenance). MomentumAgent. A trend-following trader based on moving average crossovers. It buys when the short-term average exceeds the long-term average, and exits when the reverse occurs. Key parameters: initial cash = 1M, short window = 5, long window = 20, trade size = 10, wakeup interval = 55–65s. RSIAgent. A mean-reversion trader using the Relative Strength Index (RSI). It buys when RSI drops below the oversold threshold and sells when it exceeds the overbought threshold. Key parameters: initial cash = 1M, RSI window = 14, thresholds = [30, 70], trade size = 10, wakeup interval = 55–65s. BreakoutAgent. A channel breakout trend-follower. It buys when the current price exceeds the recent maximum and sells when it falls below the recent minimum. Key parameters: initial cash = 1M, lookback window = 50, trade size = 10, wakeup interval = 55–65s. InformedTraderAgent (Base). This family of agents trades on deviations between market price and fundamental value as observed from the oracle. If deviation exceeds a threshold, subclasses determine execution style. Key parameters: initial cash = 1M, trade size = 10, deviation threshold = 2%, wakeup interval = 2–3 min. • Conservative Informed. Places passive limit orders when value deviations are large. • Aggressive Informed. Executes immediate market orders upon detecting deviations. • Opportunistic Informed. Adapts trade size to available liquidity (e.g., larger buys if ask depth is thin). • Scaled Informed. Scales order size proportionally with deviation magnitude (up to 3x baseline). • Spread-Aware Informed. Chooses between limit and market orders depending on the bid–ask spread width. DoubleDQNTradingAgent. An online reinforcement learning trader that implements Double DQN with experience replay. The state space is discretized via RSI, order book imbalance (OBI), and bid–ask spread bins. The agent learns Q-values for three actions (HOLD, BUY, SELL) with periodic target network updates. Key parameters: learning rate = 10^{-3} , discount = 0.9, \epsilon=0.1 , buffer size = 1000, batch = 32, target update = 50 steps, trade size = 100, wakeup interval = 2–3 min. LLM Agents (CoT, ReAct, Trading). Our generic LLM-based agents rely on structured prompting to integrate technical indicators, news events, and portfolio states. They differ in reasoning style: CoT uses step-by-step deliberation, ReAct interleaves reasoning with action, and Trading employs a hybrid evidence-aggregation workflow. When a news event arrives, the agents immediately request indicators and issue a trading decision; in the absence of news, they wake up on randomized intervals (5–15 minutes) to avoid synchronization artifacts. Trade sizes are discretized into multiples of 10 units, ensuring reproducibility and comparability across agents. These design choices keep agents responsive to exogenous shocks while preserving heterogeneous but tractable behavior, and robustness checks confirm stability under alternative sampling intervals and trade granularities. iTransformerTradingAgent. A supervised-learning trader using the iTransformer architecture. It predicts short-term price direction (down, neutral, up) from streaming indicators, with labels generated via delayed price changes. The model is fine-tuned online using a replay buffer. Key parameters: input features = {RSI, MACD, spread, OBI, position}, output = 3 actions, buffer size = 1000, batch = 32, train frequency = 10, trade size = 10, wakeup interval = 2–3 min. InformerTradingAgent. A supervised agent based on the Informer architecture for time-series forecasting. It takes streaming indicators from the DataCenter and predicts future price directions (down, flat, up). Predictions are used to place trades, while labels are generated ex-post using realized returns, enabling online fine-tuning with a replay buffer. Key parameters: input = {RSI, MACD, spread, OBI, position}, prediction horizon = 10 min, buffer = 1000, batch = 32, train freq = 10, trade size = 10, wakeup interval = 2–3 min. TimeMixerTradingAgent. A supervised agent using the TimeMixer sequence model, similarly trained online. It processes technical features and position state, predicts directional moves, executes trades, and fine-tunes periodically. Key parameters: input = {RSI, MACD, spread, OBI, position}, hidden dim = 64, 2 layers, buffer = 1000, batch = 32, train freq = 10, trade size = 10, wakeup interval = 2–3 min. RL Agents (A2C, PPO). Our reinforcement learning agents adopt canonical hyperparameters commonly used in the financial RL literature. For instance, the discount factor is fixed at \gamma=0.99 , entropy coefficient 0.01 , and value coefficient 0.5 , which follow standard practice in continuous control benchmarks. Training updates are based on short n -step trajectories ( n=20 ), balancing responsiveness and stability under noisy market feedback. Action space is deliberately simplified to \{\text{Buy},\text{Sell},\text{Hold}\} with fixed trade size (10 units), ensuring comparability across agents. We further verified that moderate perturbations of these hyperparameters do not materially change the qualitative dynamics. LLM FinCon (Risk-Aware). In addition to evidence-based reasoning, FinCon incorporates a lightweight risk module. We approximate Conditional Value-at-Risk (CVaR) using the last 50 realized returns and flag a risk alert when estimated CVaR falls below -2\% . This threshold is chosen empirically as a conservative proxy for downside fragility. To avoid overfitting to a single specification, we verified robustness across alternative window sizes (20–100) and thresholds ( -1\% to -3\% ), which yield consistent qualitative patterns. When a risk alert is active, trade sizes are scaled down to reflect more cautious behavior. Population Manager. The ecological dynamics are governed by a stochastic replicator–mutator mechanism. We adopt parameters within standard ranges of evolutionary game theory: selection strength \beta=2.0 , mutation rate \mu=0.05 , and environmental perturbation \sigma=0.1 . Strategy-specific Ornstein–Uhlenbeck parameters (\lambda,\nu) are used to smooth empirical payoffs, analogous to mean-reverting fitness dynamics in ecological modeling. The Dirichlet prior \alpha=10 encourages moderate baseline diversity. Although these values are not calibrated to a specific market, qualitative behaviors—such as coexistence, dominance cycles, and extinction events—are robust under moderate perturbations of (\beta,\mu,\sigma,\alpha) , as confirmed by sensitivity checks. G.6.3 Analysis Metrics Our evaluation framework adopts a micro–meso–macro structure, with metrics defined as follows. Micro (Individual Performance and Diversity). We track each agent’s financial performance and survival: • Return : \text{Return}_{i}=\tfrac{V_{i}^{\text{final}}-V_{i}^{\text{init}}}{V_{i}^{\text{init}}} , with 95\% confidence intervals via bootstrapping. • Sharpe ratio : \text{Sharpe}_{i}=\frac{\mathbb{E}[r_{i,t}]}{\sigma(r_{i,t})} . • Maximum drawdown : \text{MaxDD}_{i}=\max_{t0 indicates a heavy right tail. • Sharpe dispersion (heterogeneity of strategic performance). While excess kurtosis and skewness are computed at the market level, Sharpe dispersion measures how differently individual strategy classes perform within the same environment. Let S_{k} denote the Sharpe ratio of strategy class k , and \bar{S}=\tfrac{1}{K}\sum_{k=1}^{K}S_{k} their cross-sectional mean. Then \text{SharpeDisp}=\sqrt{\frac{1}{K}\sum_{k=1}^{K}(S_{k}-\bar{S})^{2}}. A low SharpeDisp means most strategies earn similar risk-adjusted returns (the environment does not strongly discriminate between them), whereas a high SharpeDisp indicates that the ecosystem sharply separates robust “winners” from fragile “losers” and thus provides a stringent stress test for trading rules. • Vol-of-Vol (instability and clustering of volatility). Volatility in financial markets is time-varying and exhibits persistent clustering: periods of high volatility tend to be followed by high volatility, and calm periods by calm periods ( Andersen et al. 2003 ) . To capture this, we first construct a rolling-window estimator of volatility (e.g., with window size W ): \sigma_{t}=\sqrt{\frac{1}{W}\sum_{i=t-W+1}^{t}(r_{i}-\mu_{t})^{2}},\quad\mu_{t}=\tfrac{1}{W}\sum_{i=t-W+1}^{t}r_{i}. We then measure the standard deviation of this volatility process: \text{VoV}=\sqrt{\frac{1}{T-W}\sum_{t=W}^{T}(\sigma_{t}-\bar{\sigma})^{2}},\qquad\bar{\sigma}=\tfrac{1}{T-W}\sum_{t=W}^{T}\sigma_{t}. Larger VoV indicates more unstable volatility and stronger volatility clustering (frequent transitions between calm and turbulent regimes), whereas smaller VoV corresponds to a more stationary, homoskedastic market. In heterogeneous-agent and agent-based models, the accepted evaluation philosophy is not to fit price trajectories, but to examine whether the model endogenously generates the Stylized Facts observed in real markets ( Lux & Marchesi 1999 ; LeBaron 2006 ; Hommes 2006 ) . Theoretical and microstructure studies further emphasize that these non-Gaussian features arise from interaction among market participants, not exogenous noise or curve fitting ( Bouchaud et al. 2009 ) . Thus, reproducing Stylized Facts is the most meaningful and widely accepted measure of “closeness” to real financial behavior. Economic interpretation. This metric captures a structural reorganization of the market ecology rather than a statistical effect. It is distinct from (i) volatility clustering, which is a price-level phenomenon, and (ii) strategy regime switching, which typically refers to a micro-level change within a single agent. A Phase Change reflects a macro-level transition in the ecosystem equilibrium, driven by the evolutionary forces of the FinEvo SDE (selection, innovation, perturbation). Trading indicators observed by agents. To ensure realism, agents condition their actions on standard microstructure and technical signals. Liquidity and flow. Bid–ask spread, order-book imbalance (OBI), market depth (buy/sell), and tick-rule trade flow imbalance (TFI). Momentum and oscillators. ROC, RSI, MACD (line, signal, histogram), Aroon (up/down), KDJ, and Williams-%R. Volatility and OHLC-based. ATR, Bollinger Bands (upper/middle/lower), and ADX/DMI (ADX, + DI, - DI). These features ground agent behavior in market observables rather than oracle information. Overall, this micro–meso–macro design provides complementary views: profitability and diversity of individuals, relational structures within populations, and systemic variance decomposition at the ecology level. Together they enable both fine-grained and aggregate analysis of evolving market games. G.6.4 Artificial Shocks We inject controlled exogenous shocks by simultaneously perturbing (i) the news stream processed by event-driven/LLM agents and (ii) the latent fundamental-value process observed by informational traders, while (iii) tightening microstructure conditions to mimic liquidity stress. Unless otherwise stated, shocks are scheduled during 12:00–13:30 (local time) and we report both positive and negative cases. News channel. Let \lambda_{\text{base}} be the baseline Poisson rate for news arrivals and z_{t} the scalar sentiment payload. During the shock window [t_{s},t_{e}] we amplify both frequency and magnitude : \lambda_{\text{shock}}=\kappa\,\lambda_{\text{base}},\qquad z_{t}\sim\mathcal{N}(\mu_{s},\sigma_{\text{news}}^{2}),\ \ s\in\{+1,-1\}, where s=+1 denotes a positive shock ( \mu_{+}>0 ) and s=-1 a negative shock ( \mu_{-}<0 ). All agents that subscribe to the news bus (event-driven, LLM) ingest these messages in real time; the higher-rate, large-magnitude signals create an immediate directional impulse. Fundamental channel. Informational traders observe an oracle-provided fundamental anchor P_{t}^{\star} , which we use to maintain a stable link between transaction prices and real-world valuation. Shocks are injected on top of this anchor rather than replacing it. Let P^{\star}_{\text{base}}(t) denote the baseline fundamental curve (before shocks). At the news-shock start t_{s} we overlay a deterministic bump aligned with the shock, so the effective anchor becomes \tilde{P}_{t}^{\star}\;=\;P^{\star}_{\text{base}}(t)\,\Big(1+s\,A_{f}\,e^{-(t-t_{s})/\tau_{f}}\,\mathbf{1}\{t\geq t_{s}\}\Big), where s\in\{+1,-1\} indicates positive/negative shocks, A_{f} controls impact size ( percentage of the baseline anchor), and \tau_{f} sets the decay horizon. Informational traders then act on mispricing relative to the effective anchor as in our implementation: with \delta_{t}=(P_{t}-\tilde{P}_{t}^{\star})/\tilde{P}_{t}^{\star} , they buy when \delta_{t}<-\theta and sell when \delta_{t}>\theta , thereby transmitting fundamental shocks to order flow while preserving the anchor linkage.:contentReference[oaicite:1]index=1 G.6.5 Data Details To ground our intraday simulations in realistic conditions, we construct a multi-source financial dataset that integrates heterogeneous information streams across assets, firms, and macroeconomic environments. We deliberately cover two contrasting market regimes—the global bear market of January–October 2022 and the bull market of October 2024–July 2025—to ensure robustness across structurally distinct dynamics. All signals are aligned at sub-daily resolution, enabling high-frequency event-driven trading and ecological adaptation. Market events and news. High-frequency event streams are obtained from GDELT , Reuters Markets , and the Bloomberg Terminal , each providing thousands of timestamped market-relevant headlines per trading day at minute-level granularity. These data form the primary source of exogenous shocks that drive intraday volatility. Macroeconomic and policy signals. Macroeconomic indicators from FRED (e.g., GDP, CPI, unemployment, Treasury yields) are modeled as scheduled events and injected at their official release timestamps. Federal Reserve policy documents (FOMC minutes, statements, and press conferences) are treated as discrete intraday shocks aligned to release time, providing textual signals about monetary stance and forward guidance. Firm-level and analyst data. Corporate earnings reports, financial statements, and analyst forecasts (EPS, target prices, sector outlooks) are included at their official disclosure times. Since these variables evolve more slowly, they are interpolated into daily anchors that inform fundamental-based agents. Integration and routing. Each data type is dispatched to specialized analytical agents according to its temporal resolution and structural format: high-frequency event streams to intraday news agents, scheduled macro releases to shock-sensitive agents, and low-frequency fundamentals to balance-sheet interpreters. This routing ensures that heterogeneous signals are jointly exploited for both event-driven trading and long-horizon adaptation. Source Content Resolution GDELT / Reuters / Bloomberg Market news, corporate events Minute-level intraday FRED / Macro releases GDP, CPI, unemployment, yields Intraday at release Firm reports / Analyst forecasts Earnings, targets Quarterly (daily anchors) Fed / FOMC Policy statements, press conf. Intraday at release Table 11: Data sources and temporal resolution aligned to the trading clock for intraday simulation. G.7 Motivation and Rationale of Simulation Scenarios We provide here the motivations and empirical grounding of the two scenarios used in Section 3. Scenario 1: Artificial Shocks (Controlled Mechanism Validation). This scenario is designed as a controlled stress test of the FinEvo evolutionary dynamics. By injecting exogenous positive and negative shocks into an otherwise clean synthetic market, we can isolate how disturbances propagate through the selection, innovation, and perturbation components of the evolutionary SDE. This controlled setup allows for transparent tracing of shock amplification, ecological transitions, and re-stabilization patterns, which would be difficult to interpret if tested only under noisy real-world conditions. Such controlled perturbation tests are standard in the study of heterogeneous-agent models and systemic risk. Scenario 2: Real-World News (Empirical Relevance and Robustness). After validating the mechanisms in a controlled environment, we evaluate FinEvo under news-driven real-market conditions using high-frequency GDELT/Reuters releases. This tests the framework’s ability to remain stable and informative when exposed to unstructured, economically meaningful signals that influence short-horizon returns, order flow, and trading behavior. This scenario also assesses whether heterogeneous agents—in particular LLM-based traders—can extract useful signals from real news and survive evolutionary selection. Bull vs. Bear Market Regimes. Within the Real-World News scenario, we further embed two major macro environments: a bear phase (January–October 2022) and a bull phase (October 2024–July 2025). These distinct regimes allow us to examine how evolutionary equilibria change across volatility and sentiment conditions. As shown in Figure 6, the stable ecological configurations differ substantially across market regimes, illustrating that FinEvo does not collapse to a single fixed structure, but responds to macroeconomic context in a realistic manner. Summary. Together, the controlled Artificial Shocks scenario and the empirically grounded Real-World News scenario form a coherent validation path: from mechanism-level stress testing to real-market robustness analysis. This design demonstrates that FinEvo is capable of capturing both theoretical evolutionary dynamics and realistic market adaptations, providing a comprehensive evaluation of the framework. G.8 Economic Motivation for News-Driven Synthetic Markets This section provides additional clarification on the economic foundations of our news-driven experiments and the role of LLM-based agents in FinEvo. No Structural Bias Toward LLM Dominance. FinEvo does not assume, nor structurally enforce, the dominance of LLM-based agents, and their performance is not hard-coded through the sentiment–return mapping. As shown in Figure 6 , LLM agents are not uniformly superior: in bullish regimes, the Informer agent outperforms FinCon and Trading, while in bearish regimes DQN achieves higher returns than Trading and CoT. This indicates that relative performance is determined by the evolving ecology and market conditions, rather than by any built-in preference for a particular model class. Economic Basis for the Sentiment–Return Mapping. The mapping from news sentiment to value/return shocks is grounded in established results in market microstructure and behavioral finance, rather than chosen heuristically. In the spirit of Kyle’s model of information-based trading ( Kyle 1985 ) , information shocks affect order flow and prices through informed and liquidity traders. Behavioral models show that sentiment and investor belief distortions can drive short-horizon predictability and crash risk (e.g., Barberis et al. 1998 ). Empirically, major news announcements are known to increase volatility, jump intensity, and variance dynamics ( Engle & Ng 1993 ; Andersen et al. 2003 ) . In FinEvo, news shocks enter by shifting agents’ perceived value processes V_{k}(t) , which is consistent with these mechanisms: sentiment does not directly dictate prices, but perturbs perceived fundamentals and thereby indirectly affects order flow and returns. Why Use Synthetic Endogenous Price Dynamics. We intentionally use synthetic, endogenously generated price series in the news experiments, rather than replaying a fixed historical tape. Our goal is to study closed-loop evolutionary feedback between heterogeneous strategies and the market, which real historical prices cannot capture because the true price path was generated under an unknown and unmodeled ecology. In FinEvo, a news shock acts as an exogenous information input, while the resulting price dynamics are fully endogenous, jointly determined by all interacting agents (LLM, RL, DL, and rule-based). This design enables us to analyze how different strategy populations respond to the same news process and how evolutionary selection reshapes the ecology over time. G.9 Purpose of the Simulation Section The goal of Section 3 is to demonstrate that FinEvo provides a general dynamical framework for studying market-level evolutionary phenomena induced by interacting heterogeneous agents. The simulations serve three purposes: (1) validating that the proposed evolutionary SDE coupled with a continuous double auction generates endogenous ecological dynamics (selection, dominance cycles, phase transitions); (2) illustrating system-wide responses to perturbations such as sentiment shocks or liquidity disturbances; and (3) enabling controlled experiments for market design and policy analysis. Appendix H Scientific Investigation vs. Trading-Model Construction In this appendix, we clarify what we mean by “scientific investigation of markets” in the context of FinEvo. While trading-model construction is one possible use of simulation, FinEvo is designed to support a broader research agenda focused on the emergent and systemic properties of market ecosystems. These questions cannot be addressed by optimizing a single agent’s PnL in a replayed historical environment. Systemic Risk and Stability. FinEvo enables the study of how market-wide stability changes when the strategy composition evolves. For example, the emergence of a new “species” (e.g., LLM-based agents) may alter ecological stability, volatility persistence, or crowding-induced transitions such as flash-crash–like dynamics. These phenomena arise from endogenous selection, innovation, and perturbation within the evolutionary SDE and cannot be replicated by training a single agent. Policy and Regulation Analysis. Regulators focus on the behavior of the overall market rather than on individual trading profits. FinEvo provides a controlled virtual environment to evaluate how policies—such as transaction taxes, liquidity constraints, or rule changes—affect innovation rates, species survival, and ecological diversity. This positions FinEvo as a policy laboratory . Market Design. FinEvo also enables the analysis of how auction formats, matching rules, or latency structures influence ecosystem-level coexistence among heterogeneous agents (rule-based, RL, LLM). These are structural questions about how the market functions as an adaptive game, rather than how to construct a profitable strategy. Summary. Thus, the primary research use of FinEvo is to study the adaptive, emergent, and systemic dynamics of market ecosystems—a long-standing objective in market-microstructure and financial-economics research. Appendix I Future Work: Enhancing Empirical Realism This appendix outlines several natural extensions to increase the empirical realism of FinEvo. Our current contribution focuses on establishing the evolutionary formalism—the FinEvo SDE and its selection–innovation–perturbation dynamics—together with a microstructure-aware CDA engine. The framework was deliberately designed to support the realism enhancements described below. L2-Level Microstructure Calibration. A key direction is incorporating empirical limit-order-book statistics (e.g., LOBSTER) to calibrate depth profiles, spread dynamics, order-arrival processes, and queueing effects. FinEvo’s modular separation between the CDA mechanism and evolutionary dynamics makes this integration straightforward. Hybrid Replay + Endogenous Evolution. We plan to introduce a hybrid regime in which external liquidity is replayed from historical LOB data, while strategy populations evolve endogenously under the FinEvo SDE. This provides both empirical fidelity and evolutionary tractability. Richer Economic Feedback Mechanisms. Additional realism can be achieved by introducing inventory constraints, market-impact feedback, transaction taxes, throttling rules, and adaptive market-design modules. These components allow FinEvo to be used not only as an evolutionary model but also as a platform for evaluating policy interventions and systemic risk. High-Fidelity Interactive Environment for Agent Research. A more realistic market ecology benefits research on RL agents, LLM agents, and hybrid systems. An enhanced FinEvo environment provides non-stationary competitive pressure, realistic order-book interactions, and emergent collective behaviors that cannot be obtained from isolated backtests. Scientific Applications. Improved realism enables deeper investigation of systemic fragility, shock propagation, strategy concentration, and regulatory effects on long-term ecological equilibria. This aligns FinEvo with broader market-microstructure and financial-economics research agendas. Open-Ended Strategy Evolution. An additional extension is to allow the strategies themselves to evolve over time, rather than keeping the strategy set fixed. This could be achieved through genetic algorithms, neuroevolution, or policy hybridization mechanisms that mutate or recombine existing strategies to create new ones. While such open-ended evolution substantially enriches the expressive power of the ecosystem, it also leads to a rapidly expanding strategy space that complicates theoretical analysis. In this work, we focus on establishing the ecological game formalism and its population-level dynamics, but the FinEvo framework is fully compatible with incorporating strategy-level evolution in future research. Network-Based Propagation and Local Diffusion. While FinEvo models population-level diffusion through its innovation and perturbation terms—consistent with the mean-field formulation of evolutionary game theory )—an important future direction is to incorporate micro-level propagation mechanisms based on local interactions. Such dynamics include diffusion of strategies through adjacent nodes in an agent network, local imitation on a graph, and contagion driven by neighborhood structures. Introducing network-based propagation would allow FinEvo to capture spatially heterogeneous diffusion patterns, clustered adoption, and localized cascades, complementing the current mean-field evolutionary dynamics. The modular design of the FinEvo SDE makes it straightforward to extend the evolutionary mechanism with graph-based transition operators or network-weighted innovation kernels, enabling richer models of behavioral contagion and microstructural propagation in future work. Appendix J Limitations While FinEvo provides a unified ecological game formalism with endogenous market dynamics, several modeling assumptions introduce limitations that also serve as directions for future improvement. Scope of the Evolutionary Mechanism. The evolutionary dynamics are formulated at the population level following the mean-field tradition of evolutionary game theory. This provides analytical clarity but abstracts away from network-based local propagation. As discussed in Appendix I , the framework can be naturally extended with graph-based transition operators to model diffusion through adjacent nodes. External Knowledge of LLM Agents. LLM-based agents may have encountered textual descriptions of certain historical events during pretraining. Since FinEvo generates endogenous market dynamics and the LLMs do not access future prices or internal microstructure, this overlap is un- likely to materially distort the comparative ecological outcomes we report. Appendix K Use of Large Language Models Large language models (LLMs), such as ChatGPT, were employed solely for minor language polishing. All conceptual development, technical contributions, experiments, and analyses were conducted exclusively by the authors.