Paper deep dive
Recursive Multi-Agent Trading System: Iterative Optimized Portfolio Strategy Under Geopolitical Uncertainty
Jing Yang, Yichao Wu, Jianan Liu, Penghao Liang, Mengwei Yuan, Xianyou Li, Weiran Yan
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 7/8/2026, 7:59:28 PM
Summary
The paper introduces RMATS, a recursive multi-agent trading system designed for portfolio optimization under geopolitical uncertainty. It integrates Sentiment, Report, Analysis, and Risk agents coordinated by a Manager Agent with iterative feedback loops. Evaluated over 561 trading days across 24 assets, RMATS achieves a maximum drawdown of 9.62%, outperforming baselines like MVO and FinBERT in downside protection during geopolitical stress events, though it underperforms in sustained bull markets.
Entities (14)
Relation Signals (12)
RMATS → achieves → Maximum Drawdown
confidence 95% · RMATS achieves a maximum drawdown of 9.62%
Jing Yang → affiliatedwith → Washington University in St. Louis
confidence 95% · Jing Yang jing.y@wustl.edu Washington University in St. Louis
RMATS → coordinatedby → Manager Agent
confidence 95% · The Manager Agent coordinates all specialized agents through recursive orchestration.
RMATS → includes → Sentiment Agent
confidence 95% · The framework integrates four specialized agents—Sentiment, Report, Analysis, and Risk
RMATS → includes → Analysis Agent
confidence 95% · The framework integrates four specialized agents—Sentiment, Report, Analysis, and Risk
RMATS → includes → Risk Agent
confidence 95% · The framework integrates four specialized agents—Sentiment, Report, Analysis, and Risk
Analysis Agent → uses → Hidden Markov Model
confidence 92% · The Analysis Agent combines macroeconomic factor modeling with HMM-based regime classification.
Risk Agent → uses → CVaR
confidence 92% · The Risk Agent implements CVaR estimation using an EWMA-based dynamic covariance model
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recursive Multi-Agent Trading System (RMATS) integrates four specialized agents -- Sentiment, Report, Analysis, and Risk -- coordinated through a recursive Manager Agent with iterative feedback loops. Experimental evaluation over a 561-trading-day period (January 2023 to March 2025) across a 24-asset multi-class universe demonstrates that RMATS achieves a maximum drawdown of 9.62%, lower than MVO (15.49%) and FinBERT Sentiment (15.28%), and exhibits the lowest event-period drawdown in 3 of 5 geopolitical stress scenarios tested. While RMATS underperforms return-maximizing baselines in a sustained bull market environment, ablation studies confirm the individual contribution of each agent component to downside protection. These results position RMATS as a risk-control-oriented architecture suitable for institutions prioritizing capital preservation under geopolitical uncertainty.
Tags
Links
- Source: https://arxiv.org/abs/2605.25311v1
- Canonical: https://arxiv.org/abs/2605.25311v1
Trouble viewing inline? Open PDF directly →
Full Text
23,113 characters extracted from source content.
Expand or collapse full text
Recursive Multi-Agent Trading System: Iterative Optimized Portfolio Strategy Under Geopolitical Uncertainty Jing Yang jing.y@wustl.edu Washington University in St. LouisSt. LouisMOUSA , Yichao Wu wu.yicha@northeastern.edu Northeastern UniversityBostonMAUSA , Jianan Liu jiananliu2408@gmail.com Independent ResearcherChina , Penghao Liang liang.p@northeastern.edu Northeastern UniversityBostonMAUSA , Mengwei Yuan yuanmw1998@gmail.com Independent ResearcherChina , Xianyou Li xl4230@nyu.edu New York UniversityNew YorkNYUSA and Weiran Yan yanwr2016@gmail.com Independent ResearcherChina (2026) Abstract. This paper proposes a Recursive Multi-Agent Trading System (RMATS) for iterative portfolio optimization under geopolitical uncertainty. The framework integrates four specialized agents—Sentiment, Report, Analysis, and Risk—coordinated through a recursive Manager Agent with iterative feedback loops and formally specified typed message passing. Experimental evaluation over a 561-trading-day period (January 2023 to March 2025) across a 24-asset multi-class universe demonstrates that RMATS achieves a maximum drawdown of 9.62%, lower than MVO (15.49%) and FinBERT Sentiment (15.28%), and exhibits the lowest event-period drawdown in 3 of 5 geopolitical stress scenarios tested. Ablation studies confirm the individual contribution of each agent component. These results position RMATS as a risk-control-oriented architecture suitable for institutions prioritizing capital preservation under geopolitical uncertainty. multi-agent systems, portfolio optimization, geopolitical risk, large language models, reinforcement learning, drawdown control †copyright: none†journalyear: 2026†doi: X.X†conference: 2026 Asia-Pacific Conference on Artificial Intelligence and Machine Learning; 2026; Asia-Pacific†isbn: 978-1-4503-X-X/2026†ccs: Computing methodologies Machine learning†ccs: Theory of computation Algorithmic game theory 1. Introduction Geopolitical uncertainty has become a primary driver of financial market volatility. Events such as regional conflicts, trade sanctions, energy supply disruptions, and banking system stress can trigger rapid sector rotations and capital flight that overwhelm traditional quantitative trading frameworks (markowitz1952portfolio, ; engle1982autoregressive, ). These systems rely on historical statistical patterns and typically fail to adapt quickly to sudden structural changes in the information environment (bollerslev1986generalized, ). The emergence of Large Language Models (LLMs) and multi-agent systems has opened new avenues for financial intelligence architectures (vaswani2017attention, ; wooldridge2009introduction, ). Systems such as TradingAgents (tradingagents2025, ) and FinAgent (wang2024finagent, ) demonstrate that role-specialized agent collaboration can improve market signal quality. However, existing frameworks often adopt fixed sequential pipelines without iterative refinement. This paper proposes RMATS, a Recursive Multi-Agent Trading System designed with an explicit emphasis on downside risk control. The key innovation is the recursive coordination mechanism, wherein agents iteratively revise signals through multi-round communication until portfolio weights converge. The main contributions are: • A recursive multi-agent architecture with formally specified typed message passing (AgentMessage schema) and convergence guarantee ‖(r+1)−(r)‖2<ε\|w^(r+1)-w^(r)\|_2< . • An Analysis Agent incorporating HMM-based market regime classification and Kalman filter signal fusion. • A Risk Agent with CVaR-based estimation, geopolitical stress testing, and adaptive circuit breaker mechanisms. • Comprehensive empirical evaluation with ablation studies and honest analysis of limitations. 2. Related Work 2.1. Financial Large Language Models Domain-adapted foundation models including FinGPT (yang2023fingpt, ) and BloombergGPT (wu2023bloomberggpt, ) have demonstrated strong capabilities in financial text understanding and sentiment extraction. FinMA (xie2023pixiu, ) and instruction-tuned variants have improved financial reasoning (lopezlira2023chatgpt, ). RMATS extends these approaches by embedding LLM-inspired reasoning within a recursive multi-agent coordination framework. 2.2. Multi-Agent Trading Systems TradingAgents (tradingagents2025, ) demonstrates that role-specialized agent debate can improve trading signal generation. FinAgent (wang2024finagent, ) introduces tool-augmented financial decision-making, and FinMem (yu2023finmem, ) incorporates memory-augmented reasoning. Existing systems predominantly use fixed sequential workflows. RMATS contributes a convergence-guaranteed recursive communication protocol enabling dynamic signal revision. 2.3. Reinforcement Learning for Portfolio Management Deep reinforcement learning methods including DQN, PPO, and SAC have shown promise for sequential portfolio allocation (liu2020finrl, ). FinRL (liu2021finrl2, ) and FinRL-Meta (liu2022finrlmeta, ) provide open-source frameworks for DRL-based trading. RMATS integrates a simplified DRL-inspired optimization layer, focusing on risk-adjusted reward shaping. 2.4. Geopolitical Risk and Financial Markets Baker et al. (baker2016measuring, ) introduced the Economic Policy Uncertainty (EPU) index demonstrating predictive relationships between policy uncertainty and market volatility. Caldara and Iacoviello (caldara2022measuring, ) developed the Geopolitical Risk (GPR) index. These indices confirm that geopolitical risk constitutes a distinct and persistent source of market stress that cannot be fully captured by traditional volatility measures. 3. Multi-Agent Trading Architecture RMATS adopts a hierarchical collaborative architecture comprising four layers: Data Acquisition, Specialized Agent, Recursive Coordination, and Portfolio Optimization. Figure 1 illustrates the complete architecture. Figure 1. RMATS agent collaboration architecture. Solid arrows: BroadcastMessage MrM^r from Manager to all agents each round. Dashed arrows: AgentMessage irm_i^r replies. The convergence check gates continuation; the Risk Agent can issue a circuit-breaker override. Structural diagram showing Manager Agent coordinating four specialized agents (Sentiment, Report, Analysis, Risk) through typed message passing with convergence check and portfolio output. 3.1. Manager Agent The Manager Agent coordinates all specialized agents through recursive orchestration. It monitors agent signal quality using a health score: (1) Hi(t)=αAi(t)+βSi(t)+γRi(t)−δLi(t)H_i(t)=α A_i(t)+β S_i(t)+γ R_i(t)-δ L_i(t) where Ai(t)A_i(t), Si(t)S_i(t), Ri(t)R_i(t), and Li(t)L_i(t) denote predictive accuracy, signal stability, risk-adjusted profitability, and latency, respectively. 3.2. Sentiment Agent The Sentiment Agent generates geopolitical risk scores from market-derived signals using a Difference-in-Differences (DiD) causal estimator: (2) Yit=α+βDi+γTt+δ(Di×Tt)+εitY_it=α+β D_i+γ T_t+δ(D_i× T_t)+ _it where δ estimates causal effects of geopolitical events on sector returns. 3.3. Analysis Agent The Analysis Agent combines macroeconomic factor modeling with HMM-based regime classification. A Hidden Markov Model with three latent states (bull, bear, stress) classifies market regimes: (3) πk(t)=P(St=k∣O1:t,λ) _k(t)=P(S_t=k O_1:t,λ) A Kalman filter fuses signals from the Sentiment, Report, and Analysis agents into composite signal Z(t)Z(t). 3.4. Risk Agent The Risk Agent implements CVaR estimation using an EWMA-based dynamic covariance model: (4) CVaRα(t)=−[Rp(t)∣Rp(t)≤VaRα(t)]CVaR_α(t)=-E[R_p(t) R_p(t) _α(t)] A multi-level circuit breaker activates when drawdown, geopolitical risk, or volatility exceed adaptive thresholds: (5) CB(t)=[DDp(t)>θd]∨[GRS(t)>θgeo]∨[σp(t)>θvol]CB(t)=1[D_p(t)> _d] 1[GRS(t)> _geo] 1[ _p(t)> _vol] 3.5. Recursive Coordination Protocol Each inter-agent communication uses a formal AgentMessage schema: (6) ir=ir,cir,gir,sir,τir,δirm_i^r=\w_i^r,\;c_i^r,\;g_i^r,\;s_i^r,\; _i^r,\; _i^r\ where ir∈Δnw_i^r∈ ^n is the recommended weight vector, cir∈[0,1]c_i^r∈[0,1] is the confidence score, gir∈[0,1]g_i^r∈[0,1] is the geopolitical risk assessment, sir∈0,1,2s_i^r∈\0,1,2\ is the regime classification, and δir=‖ir−ir−1‖2 _i^r=\|w_i^r-w_i^r-1\|_2 is the intra-agent update magnitude. The Manager broadcasts a BroadcastMessage Mr=¯r,g¯r,s¯r,Hr,CBrM^r=\ w^r, g^r, s^r,H^r,CB^r\ and aggregates via confidence-and-health-weighted mean: (7) ¯r=∑icir⋅Hi⋅ir∑icir⋅Hi w^r= _ic_i^r· H_i·w_i^r _ic_i^r· H_i The recursion terminates when: (8) ‖¯(r+1)−¯(r)‖2<ε=0.008\| w^(r+1)- w^(r)\|_2< =0.008 Empirical evaluation over 27 rebalancing steps confirms convergence in a median of 2 rounds (mean 3.56, maximum 8), with 74.1% of steps converging within r≤2r≤ 2. 4. Portfolio Optimization 4.1. Risk-Aware Reward Function The reinforcement learning agent maximizes a risk-aware objective: (9) ℛt=rt−λ1σt−λ2max(0,DDt−θ)R_t=r_t- _1 _t- _2 (0,D_t-θ) with λ1=0.8 _1=0.8 and λ2=1.5 _2=1.5 calibrated to prioritize drawdown control. 4.2. Constrained Portfolio Optimization Portfolio weights satisfy: (10) max⊤−λ⊤Σ _w\; μ w- (11) s.t.‖1≤Lmax,ws≤cs∀s,⊤GRS(t)≤γgeos.t. \|w\|_1≤ L_ ,\;\;w_s≤ c_s\;∀ s,\;\;w GRS(t)≤ _geo The geopolitical risk constraint γgeo _geo tightens dynamically as aggregate GRS scores increase. 5. Experiments 5.1. Dataset and Setup Historical adjusted close price data were downloaded via yfinance for 24 ETFs covering US sector equities (XLK, XLE, XLF, XLV, XLI, XLP, XLY, XLU, XLB, XLRE), international equities (EWJ, EWG, EWU, FXI, EEM), US fixed income (TLT, IEF, LQD, EMB), and commodities (GLD, SLV, USO, DBC), plus the SPY benchmark. • Training: January 2016 – December 2020 • Validation: January 2021 – December 2022 • Test: January 2023 – March 2025 (561 trading days) 5.2. Baselines Four baselines are evaluated: (1) Mean-Variance Optimization (MVO); (2) DQN Single-Agent RL; (3) FinBERT Sentiment Proxy; (4) Multi-Factor Quant combining momentum, low volatility, mean reversion, and rolling Sharpe factors. All strategies use monthly rebalancing with 10 bps transaction costs. 5.3. Geopolitical Stress Events Five stress events in the 2023–2025 test period were selected: (1) SVB Banking Crisis (March 2023); (2) Israel-Hamas War escalation (October–November 2023); (3) US-China technology sanctions (January–February 2024); (4) Middle East military escalation (April–May 2024); and (5) Global rate cut pivot (August–September 2024). 6. Results and Analysis 6.1. Overall Performance Table 1 presents complete performance results. The test period coincides with a strong equity bull market (SPY ++ ≈ 60%), which systematically favors return-maximizing strategies. Table 1. Overall Performance (Test: Jan 2023 – Mar 2025) Strategy Ann.Ret. Sharpe MDD Calmar MVO 8.08% 0.397 15.49% 0.522 DQN RL 5.13% 0.150 8.57% 0.599 FinBERT Proxy 3.69% 0.048 15.28% 0.242 Multi-Factor 7.13% 0.342 9.33% 0.764 RMATS 0.22% −-0.297 9.62% 0.023 Figure 2 shows cumulative portfolio performance. RMATS achieves the second-lowest maximum drawdown (9.62%) after DQN (8.57%), substantially outperforming MVO (15.49%) and FinBERT Sentiment (15.28%). The negative Sharpe ratio (−0.297-0.297) reflects over-calibrated defensive positioning in the low-volatility bull market environment. Figure 2. Cumulative portfolio performance (normalized to 1.0). Red shading indicates geopolitical stress event windows. RMATS (red solid) exhibits the flattest trajectory, reflecting its capital-preservation orientation. Line chart comparing cumulative returns of RMATS and four baseline strategies from 2023 to 2025, with red-shaded geopolitical stress windows. Figure 3. Portfolio drawdown comparison. RMATS maintains the most contained drawdown profile, particularly during the Israel-Hamas War and Global Rate Cut Pivot events. Drawdown comparison showing RMATS with lower peak-to-trough losses than most baselines during geopolitical stress events. Figure 4. RMATS portfolio allocation over time. Red-shaded regions indicate geopolitical stress events; RMATS systematically increases fixed income and gold allocation during these periods. Stacked area chart showing RMATS portfolio weights over time with notable shifts toward defensive assets during geopolitical stress windows. 6.2. Geopolitical Stress Event Analysis Table 2 reports event-period cumulative returns. RMATS achieves the best outcome in three of five events: Israel-Hamas War (+1.15%+1.15\%, the only positive return), Middle East Escalation (−1.45%-1.45\%, lowest loss), and Global Rate Cut Pivot (−0.78%-0.78\%, lowest loss). Table 2. Event-Period Returns (%). Bold = best per event. Strategy SVB Israel US-CN Mid-E Rate MVO −-2.47 −-2.11 −-4.34 −-3.03 −-3.04 DQN RL −-4.22 −-2.81 −-3.30 −-4.33 −-3.42 FinBERT −-1.18 −-0.89 −-3.99 −-1.13 −-3.74 Multi-Factor −-5.53 −-0.32 −-1.97 −-3.87 −-3.33 RMATS −-4.19 +1.15 −-3.50 −-1.45 −-0.78 6.3. Ablation Study Table 3 presents the ablation study. Removing the Risk Agent has the largest impact on maximum drawdown (9.62%→14.76%9.62\%→ 14.76\%), confirming it as the primary driver of downside protection. Removing recursive coordination degrades all metrics most severely. Table 3. Ablation Study — Impact of Removing Components Configuration Sharpe MDD Ann.Ret. Avg.EDD Full RMATS −-0.297 9.62% 0.22% 2.17% w/o Recursive Coord. −-0.412 11.83% −-0.91% 3.41% w/o Sentiment Agent −-0.388 12.20% −-0.64% 3.79% w/o Risk Agent −-0.341 14.76% 0.08% 4.92% w/o Analysis Agent −-0.356 11.04% −-0.33% 3.18% w/o Causal Inf. (DiD) −-0.321 10.51% −-0.11% 2.83% 6.4. Convergence Analysis The recursive protocol achieves a median convergence of 2 rounds, with 74.1% of steps completing within r≤2r≤ 2 and a maximum of 8 rounds. The aggregate weight delta ‖¯r−¯r−1‖2\| w^r- w^r-1\|_2 decreases from 0.165 at r=1r=1 to 0.007 at r=2r=2, a 95.9% reduction in a single round. Figure 5. Convergence curve: mean ‖Δ‖2\| \|_2 per coordination round. The 95.9% drop from r=1r=1 to r=2r=2 confirms rapid inter-agent consensus formation. Line chart showing mean weight delta declining sharply from round 1 to round 2, then remaining near zero. Figure 6. Distribution of convergence rounds: 74.1% of steps converge at r=2r=2; 25.9% reach rmax=8r_ =8 during stress. Bar chart showing convergence round distribution with most steps converging at round 2. Figure 7. Normal vs. geopolitical stress convergence: stress periods show 80% convergence at r=2r=2 vs. 73% for normal periods. Grouped bar chart comparing convergence speed in normal vs stress market conditions. 7. Discussion 7.1. Interpretation of Results RMATS provides meaningful downside protection during geopolitically-driven stress events, at the cost of underperformance in sustained bull market conditions. The negative Sharpe ratio (−0.297-0.297) deserves direct discussion. The 2023–2025 test period was characterized by unusually strong equity market performance driven by AI-sector enthusiasm and soft-landing macroeconomic outcomes—an environment where defensive architectures are systematically penalized. 7.2. Limitations Several limitations constrain the current results. First, the Sentiment Agent uses market-derived proxy signals rather than true LLM-based geopolitical text reasoning. Second, the evaluation period of 561 trading days remains short for robust statistical inference. Third, the proxy-based ablation study provides approximate rather than precise component contribution estimates. 7.3. Future Work Priority improvements include: (1) integration of real-time FinBERT inference for true LLM-based geopolitical sentiment; (2) extension of the evaluation period to 2016–2025 to include the COVID-19 crash and Russia-Ukraine War; (3) development of a composite performance metric explicitly capturing the return-drawdown trade-off frontier; and (4) implementation of a retrieval-augmented memory module enabling the system to reference historical geopolitical analogues. 8. Conclusion This paper proposed RMATS, a Recursive Multi-Agent Trading System for portfolio optimization under geopolitical uncertainty. The framework integrates Sentiment, Report, Analysis, and Risk agents through a formally specified recursive coordination protocol with typed message passing and empirically validated convergence (median 2 rounds, 74.1% within r≤2r≤ 2). Experimental evaluation across 24 assets over 561 trading days demonstrated that RMATS achieves a maximum drawdown of 9.62%—substantially lower than MVO (15.49%) and FinBERT Sentiment (15.28%)—and records the lowest event-period loss in three of five geopolitical stress scenarios. Ablation studies confirm the individual contribution of the Risk Agent (primary drawdown reduction) and recursive coordination (overall stability improvement). RMATS underperforms return-maximizing baselines in the 2023–2025 bull market, reflecting an expected trade-off between capital preservation and return generation. This defines the target deployment context: risk-averse institutional investors for whom capital preservation under stress conditions takes priority. Acknowledgements.The authors thank the open-source communities behind yfinance, PyTorch, HuggingFace Transformers, and hmmlearn for providing the computational infrastructure used in this research. Use of AI. Large language models were used as writing assistance to improve grammar and clarity. All intellectual contributions, experimental design, results, and analysis are entirely the work of the human authors, who take full responsibility for the content of this publication. References (1) H. Markowitz, “Portfolio selection,” The Journal of Finance, vol. 7, no. 1, p. 77–91, 1952. (2) R. F. Engle, “Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation,” Econometrica, vol. 50, no. 4, p. 987–1007, 1982. (3) T. Bollerslev, “Generalized autoregressive conditional heteroskedasticity,” Journal of Econometrics, vol. 31, no. 3, p. 307–327, 1986. (4) A. Vaswani et al., “Attention is all you need,” in Advances in Neural Information Processing Systems, vol. 30, 2017. (5) M. Wooldridge, An Introduction to MultiAgent Systems, 2nd ed. Chichester: Wiley, 2009. (6) H. Yang, X.-Y. Liu, and C. D. Wang, “FinGPT: Open-source financial large language models,” arXiv:2306.06031, 2023. (7) S. Wu et al., “BloombergGPT: A large language model for finance,” arXiv:2303.17564, 2023. (8) Q. Xie et al., “PIXIU: A large language model, instruction data and evaluation benchmark for finance,” in Advances in Neural Information Processing Systems, 2023. (9) A. Lopez-Lira and Y. Tang, “Can ChatGPT forecast stock price movements? Return predictability and large language models,” arXiv:2304.07619, 2023. (10) Tauric Research, “TradingAgents: Multi-agent trading framework,” https://github.com/TauricResearch/TradingAgents, 2025. (11) B. Wang et al., “FinAgent: A multimodal foundation agent for financial trading,” arXiv:2402.18485, 2024. (12) P. Yu et al., “FinMem: A performance-enhanced LLM trading agent with layered memory and character design,” arXiv:2311.13743, 2023. (13) Z. Liu et al., “FinRL: A deep reinforcement learning library for automated stock trading in quantitative finance,” in NeurIPS Workshop on Deep Reinforcement Learning, 2020. (14) Y. Liu et al., “FinRL-Meta: Market environments and benchmarks for data-driven financial reinforcement learning,” in Advances in Neural Information Processing Systems, vol. 35, 2022. (15) Z. Liu, J. Yang, R. Gao, M. Weng, and A. Walid, “FinRL: Deep reinforcement learning framework to automate trading in quantitative finance,” in Proc. ACM Int. Conf. AI in Finance, 2021. (16) X. Liu et al., “FinRL-Podracer: High performance and scalable deep reinforcement learning for quantitative finance,” in Proc. ACM Int. Conf. AI in Finance, 2021. (17) J. He et al., “Towards a unified view of parameter-efficient transfer learning,” in Int. Conf. Learning Representations, 2022. (18) E. J. Hu et al., “LoRA: Low-rank adaptation of large language models,” in Int. Conf. Learning Representations, 2022. (19) N. Houlsby et al., “Parameter-efficient transfer learning for NLP,” in Int. Conf. Machine Learning, 2019. (20) M. Sang and J. H. L. Hansen, “UniPET-SPK: A unified framework for parameter-efficient tuning of pre-trained speech models for robust speaker verification,” IEEE/ACM Trans. Audio, Speech, Language Process., vol. 33, p. 1402–1414, 2025. (21) M. Sang and J. H. L. Hansen, “Efficient adapter tuning of pre-trained speech models for automatic speaker verification,” in Proc. IEEE ICASSP, p. 12131–12135, 2024. (22) M. Sang, Y. Zhao, G. Liu, J. H. L. Hansen, and J. Wu, “Improving transformer-based networks with locality for automatic speaker verification,” in Proc. IEEE ICASSP, 2023. (23) M. Sang and J. H. L. Hansen, “Multi-frequency information enhanced channel attention module for speaker representation learning,” in Proc. Interspeech, p. 321–325, 2022. (24) Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, p. 436–444, 2015. (25) N. Bloom, “Fluctuations in uncertainty,” Journal of Economic Perspectives, vol. 28, no. 2, p. 153–176, 2014. (26) S. R. Baker, N. Bloom, and S. J. Davis, “Measuring economic policy uncertainty,” Quarterly Journal of Economics, vol. 131, no. 4, p. 1593–1636, 2016. (27) D. Caldara and M. Iacoviello, “Measuring geopolitical risk,” American Economic Review, vol. 112, no. 4, p. 1194–1225, 2022. (28) R. F. Engle and S. Manganelli, “CAViaR: Conditional autoregressive value at risk by regression quantiles,” Journal of Business & Economic Statistics, vol. 22, no. 4, p. 367–381, 2004. (29) R. F. Engle, “Dynamic conditional correlation: A simple class of multivariate generalized autoregressive conditional heteroskedasticity models,” Journal of Business & Economic Statistics, vol. 20, no. 3, p. 339–350, 2002. (30) L. Monostori, J. Váncza, and M. A. Kumara, “Agent-based systems for manufacturing,” CIRP Annals, vol. 55, no. 2, p. 697–720, 2006. (31) S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed. Hoboken, NJ: Pearson, 2020.