Multi-Agent Forecast Dashboard is an open-source LangGraph project for multi-agent XAU/USD research. It combines real-time gold prices, historical daily bars, news, macro data, alternative signals, and risk analysis into a coordinated workflow that produces structured entries, take-profit / stop-loss levels, time-window outlooks, and a final research conclusion.
It now also includes a Two-Round Adversarial Trading Committee layer and a Prompt Registry / Prompt Template Management Layer for the committee agents. The committee layer turns specialist analysis into an evidence package, runs a two-round bull vs. bear debate, and lets a chair agent arbitrate the final research posture. The prompt registry stores versioned English runtime prompts alongside Chinese maintenance translations so the committee prompts stay auditable, reviewable, and easy to evolve.
This project is for research and decision support only. It is not investment advice and does not provide automated trading.
黄金日线结构
- Provide daily entry, take-profit, and stop-loss levels, plus outlooks for the next 3 days, 3-5 days, 6-15 days, and 15+ days
- Fetch TradingView XAU/USD realtime quotes and daily snapshots
- Ingest historical daily bars as the foundation for indicator computation and evaluation
- Run specialized agents from technical, macro, news, sentiment, alternative-data, and risk perspectives
- Aggregate their outputs into a single structured forecast
- Display current price, daily direction, confidence, entry / take-profit / stop-loss, time-window outlooks, and research summaries in the frontend dashboard
- Persist research runs, forecasts, agent states, and feedback loops
| Agent | Role | Focus |
|---|---|---|
technical |
Technical analysis | Price structure, trend, patterns, moving averages, RSI, ATR, smart-money logic |
macro |
Macro analysis | Dollar index, rates, inflation, liquidity, and broader macro conditions |
news |
News analysis | News flow, sentiment, event impact, and representative Chinese headlines |
market_sentiment |
Market sentiment analysis | Combined sentiment from RSI, CFTC positioning, news flow, Polymarket, and historical feedback |
alt_data |
Alternative data analysis | Pizza Index, macro signals, and other supplemental indicators |
risk |
Risk analysis | Volatility boundaries, drawdown / rebound risk, key price levels, and defensive posture |
forecast |
Final aggregation | Merge all agent outputs into the final direction, trade levels, and time-window outlook |
bull / bear |
Trading committee debate | Build evidence-backed opening, rebuttal, and final positions from the evidence package |
chair |
Trading committee chair | Arbitrate the final bias, actionability, and trade plan |
repair |
Committee repair | Apply conservative fixes when validation fails |
The committee layer is intentionally structured and bounded:
- Specialist agents produce independent analysis first.
node_build_evidence_packageconsolidates the specialist output into a normalized evidence package.- Round 1 opening cases are produced by the bull and bear agents in parallel.
- Round 2 rebuttals respond point-by-point to the opposite opening case.
- Final positions state whether each side still stands by its view or downgrades to observe-only / no-trade.
- The chair agent arbitrates the final bias, actionability, and trade plan.
- A rule-based validator checks the chair decision, and a repair agent can make one or two conservative corrections if needed.
All committee reasoning must stay inside the evidence package. Committee prompts are loaded from the Prompt Registry and are not hard-coded in Python.
The Prompt Registry stores committee prompts in the database with:
- versioned prompt keys
- English runtime prompt text for the LLM
- Chinese maintenance translations for review and auditing
- active-version lookup
- prompt metadata such as prompt type, node name, model family, and output schema reference
This makes committee prompts easier to review, safer to update, and simpler to trace back to the exact version used for a given research run.
| Data source / tool | Purpose |
|---|---|
| TradingView | Realtime XAU/USD quote, daily snapshot, and price reference |
| Reuters / NewsFlow | Mainstream news flow, sentiment, and representative headlines |
| CFTC | Positioning / COT-related inputs |
| Polymarket | Event probabilities and market expectation signals |
| Pizza Index | Alternative market sentiment indicator |
| Macro data | Treasury yields, CPI, Fed funds, and other macro inputs |
| Historical feedback | Previous forecast outcomes and evaluation feedback |
| Local daily bars | Completed daily-bar inputs for evaluation and indicator foundations |
| Technical indicator tools | SMA, EMA, RSI, ATR, and volatility-structure calculations |
flowchart LR
A[TradingView / News / CFTC / Polymarket / Macro data] --> B[Load market data and latest quote]
B --> C[Compute technical indicators]
C --> D[Run specialist agents]
D --> E[Build evidence package]
P[Prompt registry / prompt versions] -.-> G
P -.-> H1
P -.-> I1
P -.-> H2
P -.-> I2
P -.-> H3
P -.-> I3
subgraph S[Two-round committee debate]
direction LR
E --> G{Branch into bull and bear tracks}
G --> H1[Bull opening]
G --> I1[Bear opening]
H1 --> H2[Bull rebuttal]
I1 --> I2[Bear rebuttal]
H2 --> H3[Bull final position]
I2 --> I3[Bear final position]
H3 --> J[Chair arbitration]
I3 --> J
end
J --> K[Validate and repair if needed]
K --> L[Persist forecast and committee trace]
L --> M[Display in dashboard]
L --> N[Historical feedback loop]
- Load local market data and the latest quote snapshot
- Compute technical indicators and structural signals
- Run the specialist agents by domain
- Consolidate specialist outputs into an evidence package
- Branch into the committee debate after the evidence package is ready
- Run bull and bear opening, rebuttal, and final-position tracks
- Let the chair agent arbitrate the final research posture
- Validate the decision and apply bounded repair if required
- Persist the structured forecast and committee trace for later review
- Render the latest result in the frontend dashboard
- Feed the result into the historical feedback loop
- LangGraph provides explicit orchestration for the multi-agent workflow
- Structured outputs make each agent result safe to render in the frontend
- Separation of tools and data sources keeps inputs, reasoning, and outputs decoupled
- Historical feedback loops turn the forecast into a repeatable research process, not a one-off text output
- Add historical evaluation analysis so past forecasts can be reviewed from a richer time-series perspective
- Expand the research view with longer-horizon retrospective comparisons and performance summaries
- Use historical context to compare how agent conclusions evolve across different market regimes
The project uses the following tools and workflow helpers:
- Codex: code authoring, debugging, refactoring, and frontend iteration
- OpenSpec: change definitions, design boundaries, and implementation scope
- Superpowers: brainstorming, planning, execution discipline, and quality checks
- ui-ux-pro-max: dashboard and README visual review, layout guidance, and style consistency
The local setup reads from dev.env by default. The Chinese README and this English README share the same project structure and scope.
If you prefer Chinese documentation, switch to 中文 README.
- Start local infrastructure with Docker Compose:
docker compose up -dThis uses the repository-level docker-compose.yml to start PostgreSQL for local persistence.
- Start the backend API:
uv run uvicorn goldfxgraph.api.app:create_app --factory --host 0.0.0.0 --port 8000 --reload- Start the frontend dashboard:
cd apps/web
npm install
npm run devThe frontend runs on http://localhost:5173 and talks to the backend through VITE_API_BASE_URL=http://localhost:8000.
Multi-Agent Forecast Dashboard is not a trading system.
This project does not provide financial advice, investment recommendations, trading signals, or automated execution. All outputs are for research, learning, and workflow exploration only.
MIT