SwingGuard: Risk-Conditioned Multi-Agent Swing Trading with Dual-Timescale Governance

Introduction

• Risk can change within minutes while a strategy formed from a completed daily research window remains active, making position direction and executable exposure distinct concerns.Systems must coordinate what position to seek, how much remains executable, and when exposure may be restored.Source: Introduction · p. 1
• SwingGuard lets risk alter direction, exposure, and parameters during deliberation rather than modifying only a completed proposal.The paper frames this as a test of whether conditioning improves the return–risk trade-off while preserving an auditable execution path.Source: Introduction · p. 1
• SwingGuard pairs bullish–bearish multi-agent analysis and a long-only Portfolio Architect with a five-minute Risk Controller that maintains persistent states and constrains execution.Risk context enters before proposal finalization, and a deterministic audit enforces controller boundaries afterward.Source: Introduction · pp. 1–2
• The evaluation covers BTC and ETH against passive, rule-based, machine-learning, and single-agent references using cost-adjusted performance and trace-level evidence.The stated evaluation period is August 2024 through January 2025.Source: Introduction · pp. 1–2

Related Work

• Forecasting and reinforcement-learning methods estimate returns or map signals to positions, while volatility-managed and regime-switching methods treat exposure and persistent market state as control objects.SwingGuard builds on the distinction between a directional thesis and executable exposure.Source: Related Work · p. 2
• Financial LLM and multi-agent systems use domain resources, specialized roles, deliberation, and collaboration, while newer general-reasoning methods structure planning, execution, or information acquisition.The cited structured-reasoning approaches address general reasoning or retrieval rather than financial risk governance.Source: Related Work · p. 2
• SwingGuard addresses asynchronous trading by using faster risk updates to condition strategy generation and govern the active swing position, linking evidence, risk states, constraints, and realized positions.This extends beyond post-hoc attribution through an auditable strategy–risk interface.Source: Related Work · p. 2

• Risk-conditioned generation admits proposals within F(Rd,0), whereas post-hoc projection enforces bounds but cannot reconstruct risk-shaped rationale or coupled parameters.Source: Problem Formulation · p. 2
• Daily specialists analyze technical, news, and fundamentals evidence, opposing managers debate long-only exposure, and the Portfolio Architect uses risk context to set the target and executable parameters.Source: Dual-Timescale SwingGuard Framework · p. 3, Daily Multi-Agent Strategy Generation · p. 3
• Five-minute rule screening and six-hour persistence trigger retrieval-augmented diagnosis only for accumulated or severe events, producing a classified condition with confidence and supporting evidence.Source: Five-Minute Hierarchical Risk Diagnosis · p. 3, Figure 2 · p. 4
• The controller fuses normalized event fields with drawdown, volatility, trend, and exposure; persistent states use holds and recovery predicates to compress, eliminate, or stage exposure.NORMAL permits the standard cap, ALERT compresses exposure, SURVIVAL forces zero, and RECOVERY restores it in stages; renewed risk can re-escalate recovery.Source: Problem Formulation · p. 2, Persistent Risk State and Recovery · p. 4, Five-Minute Hierarchical Risk Diagnosis · p. 4
• A deterministic auditor validates the proposal, arbitration applies the current state action, and the execution manager maps position changes to exchange actions while recording the risk-to-position trace.The preceding position remains active until the next five-minute open, and risk can reduce, liquidate, or stage re-entry without regenerating the daily strategy.Source: Problem Formulation · p. 2, Risk-Conditioned Strategy Generation,… · p. 4

Experimental Setup

• The evaluation targets cost-adjusted value, cross-asset consistency, and risk-to-execution auditability.External baselines measure comparative performance, while controlled variants test architectural effects.Source: Experimental Setup · p. 5
• BTC and ETH are evaluated long-only from August 1, 2024 through January 31, 2025, with 52,992 five-minute observations per asset.Earlier data support indicator warm-up and Random Forest training; the joint portfolio synchronizes executed BTC and ETH positions without another LLM decision.Source: Experimental Setup · p. 5
• Baselines span buy-and-hold, technical and trend rules, Random Forest, volatility-managed and regime-switching rules, and a Daily Single-Agent LLM.The Daily Single-Agent LLM uses the same model, decoding settings, token limit, point-in-time modalities, and daily cadence, but directly outputs a long-only target exposure.Source: Experimental Setup · p. 5
• Full versus Post-hoc Risk isolates pre-generation conditioning, while Post-hoc Risk versus No Risk Controller isolates downstream governance; analyst reports and debate are separately removed.Unaffected upstream outputs are held fixed, and affected downstream decisions are regenerated.Source: Experimental Setup · p. 5, Table 1 · p. 5
• Signals execute at the next five-minute open, with a shared 15-basis-point one-way cost comprising 10-basis-point commission and 5-basis-point slippage.Primary metrics are cumulative net return, annualized Sharpe, and maximum drawdown; additional metrics include volatility, Sortino, Calmar, turnover, and average exposure.Source: Experimental Setup · p. 5
• Decision inputs obey point-in-time availability, Random Forest training ends before evaluation, and ex-post regimes and future troughs are reserved for retrospective analysis.Position changes execute no earlier than the next five-minute open.Source: Experimental Setup · p. 5

Cross-Asset Economic Performance

• SwingGuard is the only reported method with double-digit positive returns on both BTC and ETH, Sharpe ratios above 2, and MDD below 6%.It also has the highest Sharpe and smallest drawdown point estimates, and the lowest reported annualized volatility, in both single-asset comparisons.Source: Cross-Asset Economic Performance · p. 5, Table 2 · p. 6

Capital Efficiency and Temporal Concentration

• With 31.74% average exposure, the capped portfolio returns 37.16% with 2.98 Sharpe and 6.44% drawdown, versus equal-weight buy-and-hold at 30.11%, 1.19, and 28.47%.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%, with Sharpe ratios of 3.21, 2.62, and 2.15, respectively.Source: Capital Efficiency and Temporal… · p. 6
• The capped portfolio reaches Sortino and Calmar ratios of 4.26 and 13.54, respectively, while average exposure indicates selective rather than continuous capital deployment.BTC and ETH Sortino ratios are 3.89 and 3.36, and Calmar ratios are 11.60 and 5.66, respectively.Source: Capital Efficiency and Temporal… · p. 6

Cross-Asset Economic Performance

• BTC, ETH, and the portfolio are profitable in five of six months; the largest single-asset monthly loss is 0.94%, and the portfolio’s sole monthly loss is 0.02%.November provides the largest common opportunity, while divergent December and January exposures offset losses across assets.Source: Capital Efficiency and Temporal… · p. 6

Risk-to-Execution Auditability

• At the December 18 survival intervention, the controller closed a 30% BTC position before an 8.82% decline to the subsequent local trough, with exposure remaining zero during the decline.The trace links detected risk, state transition, controller action, and realized position; aggregate paths show shallow drawdowns throughout.Source: Figure 3 · p. 6, Risk-to-Execution Auditability · pp. 6–7, Figure 4 · p. 7

Controlled Ablation Results

• Sharpe rises from 2.79 to 2.98, return from 35.03% to 37.16%, and MDD improves from -8.62% to -6.44% for Full SwingGuard versus Post-hoc Risk.The controlled comparison changes when risk enters strategy generation while retaining the downstream safety mechanism.Source: Table 3 · p. 7, Controlled Ablation Results · p. 7

• Daily multi-agent reasoning forms a risk-conditioned target while the five-minute Risk Controller maintains context, constrains exposure, and records the path from evidence to execution.Source: Conclusion and Future Work · p. 7
• Across BTC and ETH, SwingGuard remains profitable with the highest Sharpe and lowest drawdown point estimates among reported methods.Source: Conclusion and Future Work · p. 7
• The capped portfolio earns a higher return than equal-weight buy-and-hold at lower exposure, and event traces link risk states to positions.Source: Conclusion and Future Work · p. 8
• Ablations show the largest deterioration when downstream governance is removed, while specialized analysis and deliberation have complementary effects.Source: Conclusion and Future Work · p. 8
• Future work will test broader regimes and recovery efficiency.Source: Conclusion and Future Work · p. 8