News-Driven Multi-Agent Stock Selection: Beyond "News → LLM → Guess"

Published on ClawList.io | AI Automation & OpenClaw Skills

---

Most AI stock-picking demos follow the same tired pattern: feed a news headline into an LLM, receive a ticker recommendation, pretend that counts as analysis. It does not. The model has no idea whether the sector actually moved, whether institutional money is flowing in, or whether the chart confirms the narrative. It is pattern-matching on text and calling it finance.

A more rigorous architecture has been circulating in AI engineering circles, originally outlined by @WuChuanIJ, and it solves this problem cleanly. The idea is a news-driven multi-agent stock selection pipeline where each agent handles a discrete, verifiable task — and no recommendation surfaces without passing through every stage. The result is a system that is explainable, auditable, and actually useful for post-hoc review.

---

The Architecture: Four Agents, One Chain

The pipeline is a sequential multi-agent chain. Each agent produces structured output that feeds the next. No step is skipped, and every decision carries traceable evidence.

Hot Topic Agent
     ↓
Market Validation Agent  (sector momentum / capital flow / moving averages / sentiment)
     ↓
Stock Selection Agent    (individual stock evidence: sentiment + fund flow)
     ↓
Risk Management Agent    (final aggregation and go/no-go decision)

This is not a single prompt with instructions to "be careful." Each agent is scoped to a specific domain of knowledge, reducing hallucination surface and making failures easier to diagnose.

---

Agent Breakdown

Agent 1 — Hot Topic Detector

The entry point ingests news feeds, social media signals, earnings transcripts, or regulatory filings and extracts thematic clusters: which industries or narratives are generating outsized attention right now. The output is not a list of stocks. It is a list of themes — "domestic semiconductor substitution," "AI inference hardware demand," "grid modernization spending" — with source citations attached.

This separation matters. Mixing topic extraction with stock selection in a single prompt produces recommendations that are hard to interrogate. Keeping them separate means you can audit why a theme was flagged independently of why a stock was ultimately chosen.

# Simplified theme extraction output
{
  "themes": [
    {
      "label": "AI inference hardware demand",
      "confidence": 0.87,
      "sources": ["Reuters 2026-03-03", "Bloomberg 2026-03-03"],
      "related_sectors": ["semiconductors", "data center cooling"]
    }
  ]
}

Agent 2 — Market Validation Agent

This is the layer most demo systems omit entirely. Before a theme touches a stock recommendation, the Market Validation Agent checks whether the market is actually confirming the narrative with price and capital behavior. It queries:

• Sector-level momentum — is the relevant ETF or index trending?
• Capital flow data — is institutional or smart money flowing into the sector?
• Moving average alignment — is price structure constructive (e.g., above 20/50 MA)?
• Sentiment heat — is retail attention rising or already exhausted?

A hot headline about EV batteries means nothing if the sector index has been in distribution for two weeks and fund outflows are accelerating. This agent gates the pipeline: themes that fail market validation are dropped. Only confirmed themes proceed.

# Validation gate logic (pseudocode)
def validate_theme(theme, market_data):
    sector = map_theme_to_sector(theme)
    checks = {
        "momentum": sector_momentum_score(sector) > THRESHOLD,
        "capital_flow": net_fund_flow(sector, days=5) > 0,
        "ma_alignment": price_above_ma(sector, period=20),
        "sentiment": sentiment_trend(sector) in ["rising", "neutral"]
    }
    return all(checks.values()), checks

The structured output of failed checks is preserved. In a post-trade review, you can see exactly which signal caused a theme to be dropped — valuable for refining thresholds over time.

Agent 3 — Stock Selection Agent

With a validated theme in hand, the Stock Selection Agent narrows to individual tickers. Crucially, it does not stop at "this stock is in the sector." It must produce per-stock evidence across two dimensions:

• Sentiment signal — is this specific ticker generating rising discussion volume, not just the sector broadly?
• Capital flow signal — is fund flow data at the individual stock level confirming accumulation?

Both signals must be present. A stock with strong fund inflow but no sentiment traction may be a quiet accumulation play — worth flagging but not a primary pick. A stock trending on social media with no fund flow confirmation is noise. The agent ranks candidates by the strength of both signals combined and passes the top candidates forward with their evidence attached.

# Stock candidate output schema
{
  "ticker": "NVDA",
  "theme": "AI inference hardware demand",
  "sentiment_score": 0.91,
  "fund_flow_7d": "+$420M",
  "evidence_summary": "Mentioned in 3 analyst upgrades, trending in AI developer communities, consistent net buying over 7 sessions",
  "rank": 1
}

Agent 4 — Risk Management Agent

The final agent does not generate new signals. Its job is aggregation and risk-adjusted decision making. It takes all upstream outputs — theme confidence, validation checks, stock evidence — and applies portfolio-level constraints:

• Concentration limits (how many picks share the same sector or theme?)
• Volatility filtering (is implied volatility unusually elevated, suggesting event risk?)
• Contradiction detection (does the risk agent see any upstream signals that conflict with each other?)

The output is a final, structured recommendation set with a complete audit trail. Every recommended ticker links back through all four agents to its originating news source. This is what "explainable" means in practice — not a sentence saying "we considered risk," but a machine-readable chain of evidence.

---

Why This Architecture Wins in Production

Three properties make this system genuinely useful beyond a demo:

Explainability. Every recommendation has a traceable evidence chain. You can answer "why did the system recommend this ticker?" with specific, structured data rather than vibes.

Auditability. The pipeline state at any timestamp can be replayed. If a recommended stock moved against expectations, you can re-examine exactly what data each agent saw, which checks passed, and where the reasoning broke down.

Reviewability. Post-session review is built into the design. Teams can inspect dropped themes, failed validation gates, and downranked stocks to understand what the system almost recommended — as informative as what it did recommend.

---

Building This With OpenClaw

This pipeline maps naturally onto an OpenClaw multi-agent skill chain. Each agent becomes a discrete skill node with defined input/output schemas. The validation agent can call external market data tools directly. The risk agent can write structured summaries to a shared memory context that downstream systems or human reviewers access.

The key implementation decision is state persistence between agents — each agent should receive not just the final output of its predecessor but the full structured evidence object. This is what enables the audit trail. If you collapse intermediate state into a summary string, you lose auditability.

---

Conclusion

The difference between a toy and a tool is whether you can trust it when it matters. A single LLM prompt cannot tell you why it picked a stock, cannot be replayed after a bad outcome, and has no mechanism for checking whether the market agrees with the news. A properly designed multi-agent chain addresses all three problems.

The architecture described here — theme detection, market validation, evidence-based stock selection, risk aggregation — is a practical template for any team building AI-assisted research tools. The individual components are not exotic. The power is in the sequencing, the structured handoffs, and the discipline of keeping each agent focused on one verifiable question.

That is what separates signal from noise.

---

Original concept by @WuChuanIJ. Explore more AI automation architectures and OpenClaw skill patterns at ClawList.io.