Paper deep dive
SwingGuard: Risk-Conditioned Multi-Agent Swing Trading with Dual-Timescale Governance
Junzhou Su, Yichen Luo, Yebo Feng, Teng Li, Jiahua Xu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 10/12/2026, 5:41:15 AM
Summary
The paper introduces SwingGuard, a dual-timescale, long-only multi-agent framework for cryptocurrency swing trading that conditions strategy generation on real-time risk context. It employs specialized analysts for technical, sentiment, and macroeconomic deliberation, a Portfolio Architect for position formation, and a five-minute Risk Controller for persistent governance and execution constraints. Evaluated on Bitcoin and Ether, SwingGuard achieves a Sharpe ratio of 2.98, outperforming a post-hoc risk variant (2.79) by integrating risk assessment before strategy finalization and maintaining high-frequency governance.
Entities (13)
Relation Signals (11)
SwingGuard → evaluatedon → Ether
confidence 99% · Evaluations on Bitcoin (BTC) and Ether (ETH) show that SwingGuard preserves positive returns...
SwingGuard → evaluatedon → Bitcoin
confidence 99% · Evaluations on Bitcoin (BTC) and Ether (ETH) show that SwingGuard preserves positive returns...
SwingGuard → outperforms → Post-hoc Risk
confidence 95% · The exposure-capped joint portfolio achieves a Sharpe ratio of 2.98, compared with 2.79 for an otherwise identical Post-hoc Risk variant.
SwingGuard → uses → Portfolio Architect
confidence 95% · The resulting evidence and current risk context are then combined by a Portfolio Architect to form a target position.
SwingGuard → uses → Risk Controller
confidence 95% · At the governance timescale, a five-minute Risk Controller performs conditional diagnosis...
SwingGuard → usesbackbone → GPT-4o
confidence 95% · The controlled SwingGuard variants share the GPT-4o backbone
SwingGuard → uses → Bearish Manager
confidence 90% · The Bullish and Bearish Managers receive the same reports and construct opposing long-only interpretations
SwingGuard → uses → Bullish Manager
confidence 90% · The Bullish and Bearish Managers receive the same reports and construct opposing long-only interpretations
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Cryptocurrency swing trading spans mismatched timescales: positions remain active while market and event risk can change within minutes. However, many predictive and large language model (LLM)-based trading systems assess risk after strategy generation or run both functions at the same cadence, leaving proposals unconditioned by current risk and exposure weakly governed between revisions. We propose SwingGuard, a dual-timescale, long-only multi-agent framework for risk-conditioned strategy generation. At the strategy timescale, specialized analysts synthesize point-in-time technical, sentiment, on-chain, and macroeconomic evidence through bullish-bearish deliberation. The resulting evidence and current risk context are then combined by a Portfolio Architect to form a target position. At the governance timescale, a five-minute Risk Controller performs conditional diagnosis, maintains persistent risk states with staged recovery, audits the proposal, and constrains execution until the next strategy revision. This design allows risk to shape the feasible strategy before proposal finalization while continuing to govern the active position afterward. Evaluations on Bitcoin (BTC) and Ether (ETH) show that SwingGuard preserves positive returns after transaction costs while maintaining shallow drawdowns and strong risk-adjusted performance on both assets. The exposure-capped joint portfolio achieves a Sharpe ratio of 2.98, compared with 2.79 for an otherwise identical Post-hoc Risk variant. Structured traces further link risk evidence and state transitions to strategy constraints and executed position changes, providing an auditable framework for asynchronous cryptocurrency swing trading. Overall, the results indicate that conditioning strategy formation on current risk, together with persistent high-frequency governance, improves risk-adjusted performance without sacrificing auditability.
Tags
Links
- Source: https://arxiv.org/abs/2610.12988v1
- Canonical: https://arxiv.org/abs/2610.12988v1
Trouble viewing inline? Open PDF directly →
Full Text
47,009 characters extracted from source content.
Expand or collapse full text
SwingGuard: Risk-Conditioned Multi-Agent Swing Trading with Dual-Timescale Governance Junzhou Su Yichen Luo Yebo Feng Teng Li Jiahua Xu Abstract Cryptocurrency swing trading spans mismatched timescales: positions remain active while market and event risk can change within minutes. However, many predictive and large language model (LLM)-based trading systems assess risk after strategy generation or run both functions at the same cadence, leaving proposals unconditioned by current risk and exposure weakly governed between revisions. We propose SwingGuard, a dual-timescale, long-only multi-agent framework for risk-conditioned strategy generation. At the strategy timescale, specialized analysts synthesize point-in-time technical, sentiment, on-chain, and macroeconomic evidence through bullish–bearish deliberation. The resulting evidence and current risk context are then combined by a Portfolio Architect to form a target position. At the governance timescale, a five-minute Risk Controller performs conditional diagnosis, maintains persistent risk states with staged recovery, audits the proposal, and constrains execution until the next strategy revision. This design allows risk to shape the feasible strategy before proposal finalization while continuing to govern the active position afterward. Evaluations on Bitcoin (BTC) and Ether (ETH) show that SwingGuard preserves positive returns after transaction costs while maintaining shallow drawdowns and strong risk-adjusted performance on both assets. The exposure-capped joint portfolio achieves a Sharpe ratio of 2.98, compared with 2.79 for an otherwise identical Post-hoc Risk variant. Structured traces further link risk evidence and state transitions to strategy constraints and executed position changes, providing an auditable framework for asynchronous cryptocurrency swing trading. Overall, the results indicate that conditioning strategy formation on current risk, together with persistent high-frequency governance, improves risk-adjusted performance without sacrificing auditability. 1Xidian University 2University College London 3The University of Hong Kong 4Nanyang Technological University yebo.feng@ntu.edu.sg Introduction Cryptocurrency swing trading creates an asynchronous control problem: a strategy formed from a completed daily research window remains active while prices, liquidity, on-chain activity, and event risk can change within minutes. A medium-horizon thesis may therefore remain directionally valid after its planned exposure becomes unsafe. Trading systems must coordinate what position to seek with how much of that position remains executable and when exposure may be restored (Liu and Tsyvinski 2021; Liu et al. 2022; Makarov and Schoar 2020; Corbet et al. 2019; Fang et al. 2025; Omrane et al. 2025). Forecasting and reinforcement-learning methods primarily learn price or return patterns and map predictive signals to trading decisions or portfolio positions (Jiang and Liang 2017; Lahmiri and Bekiros 2019; Tharun et al. 2023; Cakici et al. 2024; Liu et al. 2023b; Jing and Kang 2024). These representative methods mainly optimize prediction or strategy decisions without explicitly modeling the independent fast risk-governance layer studied here. Sequence forecasters model long-horizon and general temporal structure (Zhou et al. 2021; Wu et al. 2021; Nie et al. 2023; Wu et al. 2023a). Large language models (LLMs) support increasingly structured forms of reasoning and financial-domain adaptation (Wei et al. 2022; Xiong et al. 2025; Xiong et al. 2026b; Liu et al. 2023a; Xie et al. 2023), while LLM-based multi-agent trading systems use complementary mechanisms including specialized analysis, agent coordination, deliberation, and memory (Luo et al. 2025; Yu et al. 2024; Xiao et al. 2024). Yet proposal-only risk review neither shapes the preceding deliberation nor governs the position at a faster cadence. The central distinction is intervention timing. For strategy information XtX_t, policy π, and risk context RtR_t, a post-hoc operator g(π(Xt),Rt)g(π(X_t),R_t) modifies a completed proposal, whereas π(Xt,Rt)π(X_t,R_t) lets risk alter direction, exposure, and parameters during deliberation. We ask whether this conditioning improves the return–risk trade-off over an otherwise identical post-hoc design while preserving an auditable path to execution. We propose SwingGuard, a dual-timescale multi-agent architecture for risk-conditioned cryptocurrency swing trading. Technical, sentiment, and fundamentals analyzers inform bullish–bearish deliberation, after which a Portfolio Architect forms a long-only target. A coupled five-minute risk chain diagnoses threats, maintains a persistent state with staged recovery, and constrains execution. Its current risk context enters the Portfolio Architect before proposal finalization; a deterministic audit subsequently enforces the controller boundaries. The strategy and governance layers operate at distinct timescales: strategy reasoning occurs over the completed research window, while risk governance updates every five minutes. We evaluate Bitcoin (BTC) and Ether (ETH) from August 2024 through January 2025 against passive, rule-based, machine-learning, and single-agent references using cost-adjusted performance and trace-level evidence. The main contributions of this paper are summarized as follows: • Risk-conditioned strategy generation: the current System Risk Context constrains the feasible strategy before the Portfolio Architect finalizes a proposal. • Dual-timescale risk governance: daily structural reasoning is coordinated with five-minute detection, persistent state transitions, staged recovery, and execution constraints. • Auditable risk-to-execution linkage: structured records connect risk evidence and state transitions to strategy constraints and executed positions. Related Work Cryptocurrency forecasting and reinforcement learning estimate returns or map signals to positions (Zhang et al. 2024; John et al. 2024; Jiang and Liang 2017; Liu et al. 2023b; Jing and Kang 2024), while technical, news, and momentum evidence remains regime sensitive (Hudson and Urquhart 2021; Anamika and Subramaniam 2022; Asness et al. 2013; Cakici et al. 2024). Volatility-managed and regime-switching methods instead treat exposure and persistent market state as control objects (Moreira and Muir 2017; Harvey et al. 2018; Hamilton 1989). SwingGuard builds on this distinction between a directional thesis and executable exposure. Financial LLMs combine domain resources (Li et al. 2023; Liu et al. 2023a; Xie et al. 2023) with trading-oriented reflection and specialized reasoning (Koa et al. 2024; Li et al. 2024; Wang et al. 2024; Ding et al. 2024). Multi-agent frameworks add structured roles (Wu et al. 2023b; Hong et al. 2024), debate (Du et al. 2023; Liang et al. 2024), and scalable collaboration (Qian et al. 2025). Beyond role specialization, recent work on general LLM reasoning increasingly introduces explicit structure and control into long-horizon inference. Structure-aware planning organizes intermediate reasoning around structured state transitions and verification (Xiong et al. 2025), multi-level reasoning decomposes long trajectories into high-level planning and low-level execution (Xiong et al. 2026b), and adaptive information control regulates the extent and granularity of evidence acquisition in search-augmented reasoning (Xiong et al. 2026a). These approaches motivate structured reasoning and controlled information flow, but address general reasoning or retrieval rather than financial risk governance. Financial systems apply role-based and deliberative mechanisms to research and trading (Yu et al. 2024; Xiao et al. 2024), while recent cryptocurrency applications employ multi-agent reasoning in portfolio allocation and adversarial trading environments (Luo et al. 2025; Luo et al. 2026). SwingGuard instead addresses the asynchronous case: faster risk updates condition strategy generation and govern the active swing position. Beyond post-hoc attribution and the broader interpretable-ML literature (Lundberg and Lee 2017; Rudin 2019; Yeo et al. 2023), it links evidence, risk states, constraints, and realized positions through an auditable strategy–risk interface. Methodology We first formulate the asynchronous decision problem, then describe the strategy and risk chains and their mapping to executed exposure. Problem Formulation Let d index a completed daily strategy-generation window and τ a subsequent five-minute governance instant. The strategy input Xd=(Md,Nd,Fd)X_d=(M_d,N_d,F_d) contains technical market, news and event sentiment, and on-chain and macroeconomic evidence observable by the close of d. Let Zd,τZ_d,τ be the latest completed five-minute observation. From it, the controller assigns Sd,τS_d,τ to NORMAL, ALERT, SURVIVAL, or RECOVERY and exposes the System Risk Context Rd,τ=(Sd,τ,qd,τ,w¯d,τ,ξd,τ,ℓd),R_d,τ=(S_d,τ,q_d,τ, w_d,τ, _d,τ, _d), (1) where q is the composite risk score, w¯ w the position cap, ξ supporting short-horizon evidence, and ℓ the long-horizon market regime. Let DdD_d denote analysis and deliberation, and UdU_d the carried position and strategy feedback. The Portfolio Architect produces a raw long-only target w^d w_d and executable parameters Θd _d: (w^dfull,Θdfull) ( w_d^full, _d^full) =π(Dd,Rd,0,Ud), =π(D_d,R_d,0,U_d), (2) (w^dpost,Θdpost) ( w_d^post, _d^post) =π(Dd,Ud). =π(D_d,U_d). Let p(d,τ)p(d,τ) denote the preceding execution instant and wp(d,τ)execw^exec_p(d,τ) its realized position. Both variants then use the same deterministic audit-and-arbitration mapping: wd,τ=g(w^d,Θd,Rd,τ,wp(d,τ)exec),0≤wd,τ≤w¯d,τ.w_d,τ=g( w_d, _d,R_d,τ,w^exec_p(d,τ)), 0≤ w_d,τ≤ w_d,τ. (3) Here g abbreviates the two-stage deterministic application of audit h followed by arbitration a, detailed below. Thus the ablation changes when risk enters reasoning, not the downstream safety mechanism. Targets formed from completed observations apply only to the next return interval. The distinction can be stated as a difference in feasible strategy spaces. Let Θd _d collect entry, exit, stop, validity-horizon, and sizing parameters, let Γd=(w^d,Θd) _d=( w_d, _d) denote the complete proposal, and let ℱ(Rd,0)F(R_d,0) denote the set admitted by the current risk state. Risk-conditioned generation and post-hoc control can be represented as Γdfull _d^full =π(Dd,Ud,ℱ(Rd,0)), =π(D_d,U_d;F(R_d,0)), (4) Γdpost _d^post =ℱ(Rd,0)[π(Dd,Ud,ℱ0)], =P_F(R_d,0)[π(D_d,U_d;F_0)], where ℱ0F_0 is the risk-agnostic strategy space under the fixed long-only rules and P is the downstream audit projection. These operators are not generally equivalent: projection can enforce safety bounds, but cannot reconstruct the rationale and coupled parameters that would have been formed inside the risk-conditioned feasible set. Dual-Timescale SwingGuard Framework Figure 1: Overall SwingGuard architecture. The Data Team provides market, news, and on-chain transaction evidence to the analytical agents and adds the latest observable U.S. M2 broad-money context before deliberation. The Risk Controller supplies pre-generation risk context to the Portfolio Architect, audits the proposed target position, and continues five-minute governance before the validated target reaches the Execution Manager. Figure 1 shows two coupled loops. The deterministic Data Team aligns point-in-time sources for the daily analytical loop, which combines multimodal evidence and bullish–bearish debate into w^d w_d. The five-minute loop supplies Rd,0R_d,0 before generation, audits the proposal, and updates the executable cap until the next daily cycle without rerunning the strategy LLMs. The deterministic Execution Manager converts approved position changes into exchange actions. This manager–specialist organization follows established role-based agent coordination, while the explicit controller boundary restricts open-ended conversation with a deterministic execution interface (Wu et al. 2023b; Hong et al. 2024; Qian et al. 2025). Daily Multi-Agent Strategy Generation The Technical Analyzer, Sentiment Analyzer, and Fundamentals Analyzer transform XdX_d into technical-market, news-event-sentiment, and on-chain-macroeconomic reports, Ad=(ϕM(Md),ϕN(Nd),ϕF(Fd)).A_d=( _M(M_d), _N(N_d), _F(F_d)). (5) The Bullish and Bearish Managers receive the same reports and construct opposing long-only interpretations: the former argues for exposure, while the latter argues for reduction or cash rather than short selling. This deliberation exposes competing reasoning paths before synthesis (Du et al. 2023; Liang et al. 2024). Their trace is Bd=β(Ad,Ud),Dd=(Ad,Bd).B_d=β(A_d,U_d), D_d=(A_d,B_d). (6) The Portfolio Architect internally uses a Research Manager to synthesize DdD_d and a Trader to convert that plan into a target and executable parameters Θd _d (entry, stop, take-profit, validity, and size). Full SwingGuard supplies Rd,0R_d,0 to both internal stages, so the risk state can alter the direction, feasible size, and parameters considered before finalization. Post-hoc Risk omits this input while holding reports, debate, strategy memory, and downstream control fixed. Five-Minute Hierarchical Risk Diagnosis (a) Three-Stage Risk Detection and Diagnosis. (b) Persistent Risk-State Transition and Strategy Conditioning. Figure 2: Hierarchical risk diagnosis and control framework. Panel (a) uses rule screening and six-hour persistence to trigger retrieval-augmented diagnosis. Panel (b) combines confirmed evidence and market conditions into persistent state transitions, producing the System Risk Context for strategy generation. “LLM CoT” records Stage 3’s bounded diagnostic rationale. As summarized in Figure 2, for every completed five-minute bar, Stage 1 screens market, on-chain, and news anomalies with deterministic rules. Stage 2 evaluates multi-source accumulation, recent density, and source diversity over a six-hour buffer. Only an accumulated trigger γd,τ _d,τ or designated severe event bd,τb_d,τ invokes Stage 3: d,τ=ℒ(Zd,τ,d,τ,d,τ),γd,τ∨bd,τ,∅,otherwise,D_d,τ= casesL(Z_d,τ,C_d,τ,K_d,τ),& _d,τ b_d,τ,\\ ,&otherwise, cases (7) where d,τC_d,τ is recent context and d,τK_d,τ historical context. The LLM classifies the current condition as normal, gray-rhino risk, or black-swan risk, with confidence and supporting evidence; a deterministic fallback handles invalid outputs. Conditional invocation concentrates LLM reasoning on persistent or severe events. For historical context, each archived snapshot eie_i stores symbol sis_i, metadata time timetat_i^meta, diagnostic fields, and a vector i v_i. A fixed, normalized 14-dimensional vector summarizes market, transaction, and news conditions. For traded symbol sds_d, the current vector d,τ v_d,τ retrieves only earlier same-symbol snapshots: d,τ=ei:i∈Top3i:si=sd,timeta<td,τcos(d,τ,i).K_d,τ= \e_i:i∈ i:\,s_i=s_d,\;t_i^meta<t_d,τTop3 ( v_d,τ, v_i) \. (8) Here, Top3Top3 returns the indices of the three highest cosine similarities. The retrieved snapshots provide precursors, causes, drawdown, and recovery profiles for diagnosis. This design draws on retrieval augmentation and episodic-memory mechanisms (Lewis et al. 2020; Shinn et al. 2023), while keeping the historical archive distinct from the online persistence state: the former supplies similar-event evidence for diagnosis, whereas the latter tracks whether current risk is accumulating, dissipating, or recovering. Persistent Risk State and Recovery The Risk Prediction interface normalizes event type, severity, confidence, and supporting evidence, while the Market Monitor supplies drawdown, volatility/average true range (ATR), momentum/trend, and current exposure. Their fusion produces qd,τq_d,τ. Given predecessor p(d,τ)p(d,τ), the state evolves with active events ℰE, hold counters u, and recovery evidence ρ: Sd,τ=T(Sp(d,τ),qd,τ,ℰd,τ,d,τ,ρd,τ).S_d,τ=T(S_p(d,τ),q_d,τ,E_d,τ, u_d,τ, _d,τ). (9) Hysteresis and minimum holds prevent a quiet bar from immediately erasing risk. Recovery requires jointly satisfied hold, price, volatility, quiet, and market-repair predicates: ρd,τ=[chold∧cprice∧cvol∧cquiet∧crepair]. _d,τ=I[c^hold c^price c^vol c^quiet c^repair]. (10) NORMAL permits the standard cap; ALERT compresses exposure; SURVIVAL forces zero; and RECOVERY restores exposure in stages. Renewed risk can re-escalate any recovery. The resulting control is (w¯d,τ,ad,τrisk)=ψ(Sd,τ,sd,τrec,ℰd,τ,Ed,τint).( w_d,τ,a^risk_d,τ)=ψ(S_d,τ,s^rec_d,τ,E_d,τ,E^int_d,τ). (11) Here srecs^rec is the staged-recovery level, EintE^int is the active set of intervention constraints, and ariska^risk is the resulting controller action. At τ=0τ=0, the controller first consumes the latest completed bar, updates the state, and serializes Eq. (1) before building either strategy prompt. This ordering implements pre-strategy conditioning without seeing the subsequent return. Risk-Conditioned Strategy Generation, Audit, and Execution After proposal generation, a hard-boundary auditor h validates long-only direction, size, stop, validity horizon, and controller limits; an arbiter a applies the current state action: (w~d,τ,Vd,τ) ( w_d,τ,V_d,τ) =h(w^d,Θd,Rd,τ), =h( w_d, _d,R_d,τ), (12) wd,τ w_d,τ =a(w~d,τ,Rd,τ,wp(d,τ)exec). =a( w_d,τ,R_d,τ,w^exec_p(d,τ)). (13) The audit verdict Vd,τV_d,τ records any validation or clipping decision. Until the next five-minute open, the preceding executed position remains active; at that open, wd,τexec=wd,τw^exec_d,τ=w_d,τ and ad,τexeca^exec_d,τ records the corresponding position change. Both Full and Post-hoc Risk retain this audit; only Eq. (2) differs. Between daily updates, Rd,τR_d,τ may reduce, liquidate, or stage re-entry without regenerating the strategy. The Execution Manager maps the change from the preceding executed position to OPEN, ADD, HOLD, REDUCE, or CLOSE and records d,τ=(CLOSE _d,τ=( Rd,τ,Θd,w^d,Vd,τ, R_d,τ, _d, w_d,V_d,τ, (14) OPENw~d,τ,wd,τ,ad,τexec,wd,τexec). w_d,τ,w_d,τ,a^exec_d,τ,w^exec_d,τ). This trace separates the LLM proposal, validated target, state-driven constraint, and realized position, linking risk evidence to execution. Experimental Setup We evaluate cost-adjusted value, cross-asset consistency, and risk-to-execution auditability. External baselines measure comparative performance; controlled variants test architectural effects. Data and protocol. We evaluate long-only BTC and ETH from August 1, 2024 through January 31, 2025. Each asset provides 52,992 five-minute observations; earlier data support indicator warm-up and Random Forest training. Decisions use only completed bars, preceding-seven-day news, observable on-chain records, and the latest released U.S. M2 broad-money observation (series code WM2NS); ex-post regimes serve analysis only. Single-asset tests start with unit wealth and zero exposure. The joint portfolio synchronizes executed BTC and ETH positions without another LLM decision. For wtB,wtE∈[0,1]w_t^B,w_t^E∈[0,1], (ωtB,ωtE)=(wtB,wtE),wtB+wtE≤1,(wtB,wtE)wtB+wtE,wtB+wtE>1.(ω^B_t,ω^E_t)= cases(w^B_t,w^E_t),&w^B_t+w^E_t≤ 1,\\[2.0pt] (w^B_t,w^E_t)w^B_t+w^E_t,&w^B_t+w^E_t>1. cases (15) Gross joint return is ωtBrt+1B+ωtErt+1E _t^Br_t+1^B+ _t^Er_t+1^E; costs are charged on both weight changes after scaling. Models and baselines. The controlled SwingGuard variants share the GPT-4o backbone, zero-temperature decoding, a 2,048-token response limit, identical prompts for retained components, and all other non-ablated settings; variants retaining the debate module use one round. Baselines include buy-and-hold and a 20-day moving average (MA); a 4-hour Donchian–ATR rule motivated by prior cryptocurrency technical-trading evidence (Hudson and Urquhart 2021) and the average true range (ATR) (Wilder 1978); a 4-hour ATR trend-exposure rule using momentum signals (Asness et al. 2013); a pre-evaluation 4-hour Random Forest strategy (Breiman 2001; Gu et al. 2020); a daily volatility-managed rule that scales exposure toward 35% annualized volatility estimated over 20 days (Moreira and Muir 2017; Harvey et al. 2018); an observable regime-switching rule combining a 120-day trend with 20-day volatility relative to its trailing 120-day median (Hamilton 1989); and a Daily Single-Agent LLM. The latter uses the same model, decoding settings, token limit, point-in-time modalities, and daily cadence, but a single tool-free prompt that directly outputs a long-only target exposure in [0,1][0,1], held until the next revision. Controlled comparisons. Table 1 supports two adjacent risk comparisons and two supporting component removals. Full versus Post-hoc Risk changes only whether Rd,0R_d,0 enters strategy generation while retaining the same subsequent audit and controller. Post-hoc Risk versus No Risk Controller then changes only the downstream audit and five-minute governance while keeping pre-generation conditioning absent; the strategy-native stop remains active. The final two variants remove the three analyst reports as one group or remove bullish–bearish deliberation. Point-in-time source windows and unaffected upstream outputs are held fixed, while affected downstream decisions are regenerated. Mode Pre Post Analysts Debate Full SwingGuard Yes Yes Yes Yes Post-hoc Risk No Yes Yes Yes No Risk Controller No No Yes Yes No Analyst Team Yes Yes No Yes No Bull/Bear Debate Yes Yes Yes No Table 1: Controlled SwingGuard configurations. Pre and Post denote risk conditioning before strategy generation and risk governance afterward. Execution and metrics. Signals execute at the next five-minute open. Every method uses the same prices, long-only constraint, 10-basis-point commission, and 5-basis-point slippage, for a 15-basis-point one-way cost; turnover is charged whenever executed exposure changes. Primary comparisons use cumulative net return, annualized Sharpe (Sharpe 1994), and maximum drawdown. We additionally compute annualized volatility, Sortino (Sortino and Price 1994), Calmar, turnover, and average exposure from the five-minute marked-to-market series, using a 365-day cryptocurrency calendar and zero risk-free rate. Monthly net returns measure temporal concentration, while single-asset and exposure-capped portfolio results measure cross-asset consistency. Look-ahead bias control. All decision inputs obey a point-in-time availability boundary. For a decision timestamp t, technical features use only bars closed by t; the news window ends at t; on-chain observations must already be observable; and WM2NS contributes only the latest observation whose release date has occurred by t. The Random Forest is trained exclusively on data before August 1, 2024 and is frozen throughout evaluation. SwingGuard and every baseline form signals from the same historical boundary, and any resulting position change is executed no earlier than the next five-minute open. Finally, ex-post regime labels and future local troughs are used only for retrospective analysis and figure annotation; neither enters agent prompts, baseline features, portfolio construction, nor ablation regeneration. Experimental Results Cross-Asset Economic Performance Table 2 compares cost-adjusted BTC and ETH performance. SwingGuard is the only reported method that produces double-digit positive returns on both assets while maintaining Sharpe ratios above 2 and maximum drawdown (MDD) below 6%. It also has the highest Sharpe and smallest drawdown point estimates, together with the lowest reported annualized volatility, in both single-asset comparisons. BTC ETH Method Sharpe ↑ Return (%) ↑ MDD (%) ↑ Vol. (%) ↓ Sharpe ↑ Return (%) ↑ MDD (%) ↑ Vol. (%) ↓ Buy-and-hold 1.94 58.25 -24.70 54.45 0.40 1.96 -33.06 68.97 MA (20d) 1.91 41.69 -15.32 40.52 -0.52 -14.53 -31.10 42.34 Donchian–ATR 4h 1.95 36.94 -12.22 35.22 0.28 1.78 -23.64 38.58 ATR trend exposure 4h 1.34 17.29 -10.91 26.25 0.07 -0.67 -16.74 25.06 Random Forest 4h 0.50 6.00 -21.98 36.75 -0.54 -12.94 -28.76 37.91 Volatility managed 1.93 42.13 -19.70 40.23 0.22 0.60 -22.31 38.19 Observable regime switching 1.52 27.54 -18.50 36.00 0.68 9.23 -22.48 34.75 Daily Single-Agent LLM 0.56 7.02 -12.41 36.79 0.23 1.17 -16.45 35.64 Full SwingGuard 2.70 19.91 -3.74 13.67 2.36 15.35 -5.78 12.32 Table 2: Cost-adjusted performance from August 2024 to January 2025. MDD is reported as a negative return; Vol. denotes annualized volatility. Arrows show the preferred direction. Blue, orange, green, and gray-blue identify the top three results per column for Sharpe, return, MDD, and volatility, respectively; the best result is boldfaced. Volatility management preserves much of the BTC opportunity (42.13% return and 1.93 Sharpe) but transfers weakly to ETH. Observable regime switching improves ETH return to 9.23% and Sharpe to 0.68, yet retains 22.48% drawdown. SwingGuard remains profitable on both assets, exceeds passive ETH return by 13.39 percentage points, and reduces ETH drawdown by 82.5%. Its higher Sharpe and lower drawdown on both assets are not reproduced by either simple exposure scaling or an observable trend–volatility filter. Its BTC and ETH annualized volatilities, 13.67% and 12.32%, are approximately half the next-lowest reported values of 26.25% and 25.06%, respectively. The Daily Single-Agent LLM aligns the daily cadence and continuous long-only position space, providing a direct single-agent reference. It is profitable on both assets but transfers weakly to ETH, whereas SwingGuard preserves double-digit ETH return and improves BTC Sharpe and drawdown. The remaining system-level distinction is SwingGuard’s specialized roles, deliberation, and five-minute risk governance. Capital Efficiency and Temporal Concentration SwingGuard attains this risk-adjusted performance with limited capital deployment: average exposure is 19.42% for BTC, 12.56% for ETH, and 31.74% after the joint cap. The synchronized, cost-adjusted portfolio returns 37.16% with 2.98 Sharpe and 6.44% drawdown, versus 30.11%, 1.19, and 28.47% for equal-weight buy-and-hold. Because the joint result uses capped positions rather than summed returns, its comparison with buy-and-hold reflects active timing, cash allocation, proportional exposure scaling, and associated trading costs under a shared capital budget. BTC, ETH, and joint turnover are 14.54, 11.50, and 25.80. Repricing the same executed positions at 5, 30, and 50 basis points one-way cost leaves joint returns of 40.75%, 31.95%, and 25.31% and Sharpe ratios of 3.21, 2.62, and 2.15, respectively. With 31.74% average exposure, the capped portfolio exceeds equal-weight buy-and-hold return by 7.05 percentage points, raises Sharpe from 1.19 to 2.98, and reduces drawdown magnitude by 77.4%, linking selective deployment to capital efficiency and downside control. Downside-sensitive metrics reinforce the shallow-drawdown result. BTC and ETH obtain Sortino ratios of 3.89 and 3.36 and Calmar ratios of 11.60 and 5.66; the capped portfolio reaches 4.26 and 13.54, respectively. Sortino ratios above the corresponding Sharpe ratios indicate that downside variation is smaller than total return variation. Separately, low average exposure shows that this risk-adjusted performance is achieved through selective capital deployment rather than continuous market participation. BTC, ETH, and the portfolio are profitable in five of six months. November provides the largest common opportunity, while divergent December and January exposures offset losses across assets. Returns are therefore concentrated but not confined to a single month: the largest single-asset monthly loss is 0.94%, and the portfolio’s sole monthly loss is 0.02%. SwingGuard captures the principal opportunity while preserving gains as the assets later weaken or diverge. Figure 3: BTC risk-to-execution trace around the December 18 survival intervention. The controller closed a 30% position before an 8.82% decline to the subsequent local trough. Figure 4: Cost-adjusted five-minute net value and drawdown paths for SwingGuard on BTC, ETH, and the exposure-capped joint portfolio. Risk-to-Execution Auditability Figure 3 traces a representative intervention while the system held an active BTC position. At 21:15 on December 18, the controller entered the survival response and closed the 30% exposure. BTC subsequently declined 8.82% to the local trough on December 20 while executed exposure remained zero. The trace links detected risk, state transition, controller action, and realized position over the subsequent decline. Figure 4 shows shallow drawdowns throughout the five-minute paths. The BTC and ETH equity curves rise most during the shared November opportunity and then consolidate without surrendering the accumulated gain. The December trace complements this aggregate pattern: nonzero exposure precedes escalation, the state changes before the trough, and exposure remains zero during the decline. Together they connect state transitions to trading actions; the controlled comparison below then tests the incremental effect of pre-generation conditioning. Controlled Ablation Results Table 3 reports the configurations in Table 1 on the exposure-capped BTC–ETH portfolio. Full versus Post-hoc Risk tests when risk enters strategy generation; Post-hoc Risk versus No Risk Controller tests downstream governance; the final two variants measure the Analyst Team and debate components. Mode Sharpe ↑ Return (%) ↑ MDD (%) ↑ Full SwingGuard 2.98 37.16 -6.44 Post-hoc Risk 2.79 35.03 -8.62 No Risk Controller 1.87 23.02 -14.19 No Analyst Team 2.88 35.96 -6.71 No Bull/Bear Debate 2.82 34.90 -7.03 Table 3: Controlled ablation results for the exposure-capped BTC–ETH portfolio. Blue, orange, and green identify the top three results for Sharpe, return, and MDD, respectively; the best result is boldfaced. MDD is reported as a negative return, so higher is better. Full SwingGuard ranks first on all three metrics. Against Post-hoc Risk, it increases Sharpe from 2.79 to 2.98 and return from 35.03% to 37.16%, while improving MDD from -8.62% to -6.44%. Because these configurations share evidence, strategy machinery, audit, and controller, the observed difference is associated with placing risk before strategy finalization. Sharpe and return improve by 6.8% and 6.1%, while drawdown magnitude falls by 25.3%, indicating that conditioning has its strongest relative effect on downside quality. With pre-generation conditioning absent, removing downstream governance produces the largest adjacent deterioration: Sharpe falls from 2.79 to 1.87, return from 35.03% to 23.02%, and MDD deteriorates from -8.62% to -14.19%. Removing the Analyst Team or debate yields smaller but directionally consistent losses: Sharpe declines to 2.88 and 2.82, respectively, with lower returns and higher drawdowns. This ordering associates the largest increment with five-minute governance, a further improvement to upstream conditioning, and complementary gains to specialization and deliberation. Together, these comparisons separate two sources of stability: downstream governance limits realized exposure between revisions, while pre-generation conditioning changes the proposal entering that layer. Their combined advantage reflects coordinated intervention timing and proposal formation. Conclusion and Future Work SwingGuard formulates cryptocurrency swing trading as dual-timescale control: daily multi-agent reasoning forms a risk-conditioned target, while the five-minute Risk Controller maintains the System Risk Context, constrains exposure, and records the path from evidence to execution. This coordination lets risk shape proposal formation and maintain exposure compatibility between daily revisions. Across BTC and ETH, SwingGuard remains profitable with the highest Sharpe and lowest drawdown point estimates among reported methods. The capped portfolio also earns a higher return than equal-weight buy-and-hold at lower exposure. Monthly results support cross-asset consistency, while the event trace establishes auditable state-to-position linkage. Ablations support both risk pathways, with the largest deterioration after removing downstream governance. Supporting ablations distinguish this primary risk-governance contribution from the complementary effects of specialized analysis and deliberation. Future work will test broader regimes and recovery efficiency. Acknowledgments This work was supported in part by Ripple under University Blockchain Research Initiative (UBRI) (Feng et al. 2022). References Anamika and Subramaniam (2022) A. Anamika and S. Subramaniam Do news headlines matter in the cryptocurrency market?. Applied Economics 54 (54), p. 6322–6338. Cited by: Related Work. Asness et al. (2013) C. S. Asness, T. J. Moskowitz, and L. H. Pedersen Value and momentum everywhere. The journal of finance 68 (3), p. 929–985. Cited by: Related Work, Models and baselines.. Breiman (2001) L. Breiman Random forests. Machine learning 45 (1), p. 5–32. Cited by: Models and baselines.. Cakici et al. (2024) N. Cakici, S. J. H. Shahzad, B. Będowska-Sójka, and A. Zaremba Machine learning and the cross-section of cryptocurrency returns. International Review of Financial Analysis 94, p. 103244. Cited by: Introduction, Related Work. Corbet et al. (2019) S. Corbet, B. Lucey, A. Urquhart, and L. Yarovaya Cryptocurrencies as a financial asset: a systematic analysis. International Review of Financial Analysis 62, p. 182–199. Cited by: Introduction. Ding et al. (2024) Q. Ding, H. Shi, J. Guo, and B. Liu TradExpert: revolutionizing trading with mixture of expert LLMs. External Links: 2411.00782 Cited by: Related Work. Du et al. (2023) Y. Du, S. Li, A. Torralba, J. B. Tenenbaum, and I. Mordatch Improving factuality and reasoning in language models through multiagent debate. External Links: 2305.14325 Cited by: Related Work, Daily Multi-Agent Strategy Generation. Fang et al. (2025) F. Fang, C. Ventre, M. Basios, L. Kanthan, D. Martinez-Rego, F. Wu, and L. Li Cryptocurrency trading: a comprehensive survey. In Blockchain, Crypto Assets, and Financial Innovation, p. 55–127. External Links: Document Cited by: Introduction. Feng et al. (2022) Y. Feng, J. Xu, and L. Weymouth University blockchain research initiative (ubri): boosting blockchain education and research. IEEE Potentials 41 (6), p. 19–25. Cited by: Acknowledgments. Gu et al. (2020) S. Gu, B. Kelly, and D. Xiu Empirical asset pricing via machine learning. The Review of Financial Studies 33 (5), p. 2223–2273. Cited by: Models and baselines.. Hamilton (1989) J. D. Hamilton A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica: Journal of the econometric society 57 (2), p. 357–384. Cited by: Related Work, Models and baselines.. Harvey et al. (2018) C. R. Harvey, E. Hoyle, R. Korgaonkar, S. Rattray, M. Sargaison, and O. Van Hemert The impact of volatility targeting. Note: SSRN Working Paper 3175538 Cited by: Related Work, Models and baselines.. Hong et al. (2024) S. Hong, M. Zhuge, J. Chen, X. Zheng, Y. Cheng, J. Wang, C. Zhang, Z. Wang, S. K. S. Yau, Z. Lin, L. Zhou, C. Ran, L. Xiao, C. Wu, and J. Schmidhuber MetaGPT: meta programming for a multi-agent collaborative framework. In International Conference on Learning Representations, Cited by: Related Work, Dual-Timescale SwingGuard Framework. Hudson and Urquhart (2021) R. Hudson and A. Urquhart Technical trading and cryptocurrencies. Annals of Operations Research 297 (1), p. 191–220. Cited by: Related Work, Models and baselines.. Jiang and Liang (2017) Z. Jiang and J. Liang Cryptocurrency portfolio management with deep reinforcement learning. In 2017 Intelligent systems conference (IntelliSys), p. 905–913. Cited by: Introduction, Related Work. Jing and Kang (2024) L. Jing and Y. Kang Automated cryptocurrency trading approach using ensemble deep reinforcement learning: learn to understand candlesticks. Expert Systems with Applications 237, p. 121373. Cited by: Introduction, Related Work. John et al. (2024) D. L. John, S. Binnewies, and B. Stantic Cryptocurrency price prediction algorithms: a survey and future directions. Forecasting 6 (3), p. 637–671. Cited by: Related Work. Koa et al. (2024) K. J. Koa, Y. Ma, R. Ng, and T. Chua Learning to generate explainable stock predictions using self-reflective large language models. In Proceedings of the ACM web conference 2024, p. 4304–4315. Cited by: Related Work. Lahmiri and Bekiros (2019) S. Lahmiri and S. Bekiros Cryptocurrency forecasting with deep learning chaotic neural networks. Chaos, Solitons & Fractals 118, p. 35–40. Cited by: Introduction. Lewis et al. (2020) P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. Yih, T. Rocktäschel, S. Riedel, and D. Kiela Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, Vol. 33, p. 9459–9474. Cited by: Five-Minute Hierarchical Risk Diagnosis. Li et al. (2023) Y. Li, S. Wang, H. Ding, and H. Chen Large language models in finance: a survey. In Proceedings of the fourth ACM international conference on AI in finance, p. 374–382. Cited by: Related Work. Li et al. (2024) Y. Li, B. Luo, Q. Wang, N. Chen, X. Liu, and B. He CryptoTrade: a reflective LLM-based agent to guide zero-shot cryptocurrency trading. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, p. 1094–1106. Cited by: Related Work. Liang et al. (2024) T. Liang, Z. He, W. Jiao, X. Wang, Y. Wang, R. Wang, Y. Yang, S. Shi, and Z. Tu Encouraging divergent thinking in large language models through multi-agent debate. In Proceedings of the 2024 conference on empirical methods in natural language processing, p. 17889–17904. Cited by: Related Work, Daily Multi-Agent Strategy Generation. Liu et al. (2023a) X. Liu, G. Wang, H. Yang, and D. Zha FinGPT: democratizing internet-scale data for financial large language models. External Links: 2307.10485 Cited by: Introduction, Related Work. Liu et al. (2023b) Y. Liu, Z. Li, R. Nekhili, and J. Sultan Forecasting cryptocurrency returns with machine learning. Research in International Business and Finance 64, p. 101905. Cited by: Introduction, Related Work. Liu et al. (2022) Y. Liu, A. Tsyvinski, and X. Wu Common risk factors in cryptocurrency. The Journal of Finance 77 (2), p. 1133–1177. Cited by: Introduction. Liu and Tsyvinski (2021) Y. Liu and A. Tsyvinski Risks and returns of cryptocurrency. The Review of Financial Studies 34 (6), p. 2689–2727. Cited by: Introduction. Lundberg and Lee (2017) S. M. Lundberg and S. Lee A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, Vol. 30. Cited by: Related Work. Luo et al. (2026) Y. Luo, Y. Feng, J. Xu, and Y. Liu Resisting manipulative bots in meme coin copy trading: a multi-agent approach with chain-of-thought reasoning. In Proceedings of the ACM Web Conference 2026, p. 3414–3425. Cited by: Related Work. Luo et al. (2025) Y. Luo, Y. Feng, J. Xu, P. Tasca, and Y. Liu LLM-powered multi-agent system for automated crypto portfolio management. External Links: 2501.00826 Cited by: Introduction, Related Work. Makarov and Schoar (2020) I. Makarov and A. Schoar Trading and arbitrage in cryptocurrency markets. Journal of Financial Economics 135 (2), p. 293–319. Cited by: Introduction. Moreira and Muir (2017) A. Moreira and T. Muir Volatility-managed portfolios. The Journal of Finance 72 (4), p. 1611–1644. Cited by: Related Work, Models and baselines.. Nie et al. (2023) Y. Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam A time series is worth 64 words: long-term forecasting with transformers. In International Conference on Learning Representations, Cited by: Introduction. Omrane et al. (2025) W. B. Omrane, H. Dabbou, S. Saadi, T. Savaser, and S. Sebai Exploring volatility reactions in cryptocurrency markets using intraday macroeconomic news analysis. International Review of Economics & Finance 103, p. 104509. Cited by: Introduction. Qian et al. (2025) C. Qian, Z. Xie, Y. Wang, W. Liu, K. Zhu, H. Xia, Y. Dang, Z. Du, W. Chen, C. Yang, et al. Scaling large language model-based multi-agent collaboration. In International Conference on Learning Representations, Cited by: Related Work, Dual-Timescale SwingGuard Framework. Rudin (2019) C. Rudin Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence 1 (5), p. 206–215. Cited by: Related Work. Sharpe (1994) W. F. Sharpe The sharpe ratio. The Journal of Portfolio Management 21 (1), p. 49–58. Cited by: Execution and metrics.. Shinn et al. (2023) N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao Reflexion: language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems, Vol. 36, p. 8634–8652. Cited by: Five-Minute Hierarchical Risk Diagnosis. Sortino and Price (1994) F. A. Sortino and L. N. Price Performance measurement in a downside risk framework. the Journal of Investing 3 (3), p. 59–64. Cited by: Execution and metrics.. Tharun et al. (2023) S. Tharun, G. Saranya, T. Tamilvizhi, and R. Surendran Cryptocurrency price prediction using deep learning. In International Conference on Mining Intelligence and Knowledge Exploration, p. 283–300. Cited by: Introduction. Wang et al. (2024) Q. Wang, Y. Gao, Z. Tang, B. Luo, N. Chen, and B. He Exploring LLM cryptocurrency trading through fact-subjectivity aware reasoning. External Links: 2410.12464 Cited by: Related Work. Wei et al. (2022) J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V. Le, and D. Zhou Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, Vol. 35, p. 24824–24837. Cited by: Introduction. Wilder (1978) J. W. Wilder New concepts in technical trading systems. Trend Research, Greensboro, NC. Cited by: Models and baselines.. Wu et al. (2023a) H. Wu, T. Hu, Y. Liu, H. Zhou, J. Wang, and M. Long TimesNet: temporal 2D-variation modeling for general time series analysis. In International Conference on Learning Representations, Cited by: Introduction. Wu et al. (2021) H. Wu, J. Xu, J. Wang, and M. Long Autoformer: decomposition transformers with auto-correlation for long-term series forecasting. In Advances in Neural Information Processing Systems, Vol. 34, p. 22419–22430. Cited by: Introduction. Wu et al. (2023b) Q. Wu, G. Bansal, J. Zhang, Y. Wu, B. Li, E. Zhu, L. Jiang, X. Zhang, S. Zhang, J. Liu, A. H. Awadallah, R. W. White, D. Burger, and C. Wang AutoGen: enabling next-gen LLM applications via multi-agent conversation. External Links: 2308.08155 Cited by: Related Work, Dual-Timescale SwingGuard Framework. Xiao et al. (2024) Y. Xiao, E. Sun, D. Luo, and W. Wang TradingAgents: multi-agents LLM financial trading framework. External Links: 2412.20138 Cited by: Introduction, Related Work. Xie et al. (2023) Q. Xie, W. Han, X. Zhang, Y. Lai, M. Peng, A. Lopez-Lira, and J. Huang PIXIU: a large language model, instruction data and evaluation benchmark for finance. In Advances in Neural Information Processing Systems, Vol. 36. Cited by: Introduction, Related Work. Xiong et al. (2026a) S. Xiong, O. Gungordu, J. C. Kerce, and F. Fekri Adaptive information control for search-augmented llm reasoning. arXiv preprint arXiv:2602.01672. Cited by: Related Work. Xiong et al. (2026b) S. Xiong, A. Payani, and F. Fekri Enhancing language model reasoning with structured multi-level modeling. In International Conference on Learning Representations, Vol. 2026, p. 36557–36610. Cited by: Introduction, Related Work. Xiong et al. (2025) S. Xiong, A. Payani, Y. Yang, and F. Fekri Deliberate reasoning in language models as structure-aware planning with an accurate world model. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 31900–31931. External Links: Document Cited by: Introduction, Related Work. Yeo et al. (2023) W. J. Yeo, W. Van Der Heever, R. Mao, E. Cambria, R. Satapathy, and G. Mengaldo A comprehensive review on financial explainable AI. External Links: 2309.11960 Cited by: Related Work. Yu et al. (2024) Y. Yu, Z. Yao, H. Li, Z. Deng, Y. Jiang, Y. Cao, Z. Chen, J. W. Suchow, Z. Cui, R. Liu, Z. Xu, D. Zhang, K. Subbalakshmi, G. Xiong, Y. He, J. Huang, D. Li, and Q. Xie FinCon: a synthesized LLM multi-agent system with conceptual verbal reinforcement for enhanced financial decision making. In Advances in Neural Information Processing Systems, Vol. 37, p. 137010–137045. Cited by: Introduction, Related Work. Zhang et al. (2024) J. Zhang, K. Cai, and J. Wen A survey of deep learning applications in cryptocurrency. iScience 27 (1), p. 108509. External Links: Document Cited by: Related Work. Zhou et al. (2021) H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang Informer: beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, Vol. 35, p. 11106–11115. Cited by: Introduction.