How the News-Pulse Skill Attributes Stock Price Movements to Specific Events in AI-Berkshire

The news-pulse skill uses a deterministic multi-agent workflow that aggregates evidence from four orthogonal research dimensions—corporate events, regulatory changes, industry dynamics, and market sentiment—to produce a weighted, confidence-rated attribution of observed price movements.

The news-pulse skill in the xbtlin/ai-berkshire repository implements a structured approach to financial forensics. By deploying specialized sub-agents that investigate distinct causal factors, the system transforms raw price volatility into actionable, evidence-based explanations grounded in source code logic defined in skills/news-pulse.md.

Workflow Architecture: From Price Drop to Attribution Report

The attribution process follows a strict three-stage pipeline that ensures comprehensive coverage while maintaining analytical rigor.

Parameter Intake and Information Grading

The skill receives the company name, optional price-movement description, time window, and focus dimension upon invocation (source L21-L28). Before launching investigations, the system computes an information-availability grading (A/B/C rating) that modulates search aggressiveness for each sub-agent (source L34-L41).

This grading ensures that agents adjust their search depth based on data richness, preventing over-confidence in sparse information environments.

Team Creation and Parallel Processing

A temporary team ({公司名}-newspulse) is instantiated to manage the investigation (source L46-L53). Four parallel tasks are spawned simultaneously, with each agent operating under a strict output schema requiring core findings, timeline tables, attribution confidence scores, and data-gap statements (source L35-L42).

The Four Research Dimensions

Each sub-agent investigates a distinct causal layer, ensuring no single narrative dominates the attribution.

Company-Level Corporate Events (company-event-scout)

The company-event-scout agent examines official announcements, earnings releases, M&A activity, and litigation events (source L54-L66). Output includes dates, event descriptions, sources, and relevance ratings (high/medium/low).

Regulatory and Policy Changes (regulatory-watcher)

The regulatory-watcher agent monitors industry-specific rules, cross-border policy shifts, tax modifications, and antitrust actions (source L67-L78). It categorizes impacts as direct or indirect and assigns temporal relevance scores.

Industry and Competitor Dynamics (industry-peer-analyst)

The industry-peer-analyst evaluates peer earnings, supply-chain disruptions, and beta-wave assessments across the competitive landscape (source L80-L90). This agent captures sector-wide movements that might explain systematic price drift.

Market Sentiment and Capital Flow (sentiment-tracker)

The sentiment-tracker aggregates analyst rating changes, 13F institutional holdings updates, short-interest spikes, and key opinion leader commentary (e.g., posts from influential investors like 段永ping) (source L92-L104). It outputs sentiment polarity, source attribution, and confidence metrics.

All four agents execute in parallel background processes, typically completing within approximately ten minutes (source L107-L113).

Aggregation Logic: How Attribution Is Calculated

When sub-agents return their findings, the team-lead agent executes a deterministic merging process defined in skills/news-pulse.md (source L154-L176).

Weighted Timeline Construction

The lead agent synthesizes a merged chronological table that includes a weight column (🔴 high, 🟡 medium, ⚪ low) indicating how strongly each event explains the price swing (source L166-L174). These weights derive directly from the relevance annotations supplied by each sub-agent (e.g., “与股价异动相关性：高/中/低” in the company-event task (source L64-L66)).

Evidence Matrix and Confidence Scoring

The final report contains an evidence matrix with rows for candidate explanations, supporting evidence, counter-evidence, confidence scores, and persistence estimates (source L175-L181). This structure forces explicit consideration of alternative hypotheses.

Nature Classification

The team-lead explicitly classifies the movement nature by selecting one of four categories: 价值事件 (value event), 情绪/技术波动 (sentiment/technical fluctuation), 真因不明 (cause unknown), or 混合 (mixed) (source L184-L188). If no event reaches a high relevance threshold, the skill automatically flags the case as “真因不明” and recommends continued monitoring.

Practical Usage and Output Format

Invoking the Skill

Users trigger the attribution workflow via slash command or API call:

# Invoke the skill with specific parameters

/pulse-news 公司名="拼多多" 时间窗口=14 股价异动="跌 12%/3天"

The price-movement description is injected into each agent’s prompt to contextualize relevance scoring (source L121-L122).

Understanding the Attribution Report

Sub-agents return structured JSON-style outputs:

{
  "timeline": [
    {"date":"2026-04-28","event":"公司发布 2025 Q4 财报，净利润下滑 8%","source":"https://example.com/notice","relevance":"高"},
    {"date":"2026-04-27","event":"监管发布新政策限制电子商务平台促销","source":"https://gov.cn/policy","relevance":"中"}
  ],
  "confidence":"0.85",
  "data_gap":"未找到关于供应链成本的公开信息"
}

The final aggregated attribution produces a weighted timeline:

| 日期       | 维度       | 事件                                 | 来源                         | 异动归因权重 |
|------------|------------|--------------------------------------|------------------------------|--------------|
| 2026-04-28 | 公司       | 财报利润下滑 8%                      | https://example.com/notice   | 🔴 高        |
| 2026-04-27 | 监管       | 新政策限制促销活动                  | https://gov.cn/policy        | 🟡 中        |
| 2026-04-26 | 情绪       | 段永平在雪球发帖质疑业绩前景       | https://xueqiu.com/post/123  | ⚪ 低        |

Persistence and Cleanup

Final reports are persisted under reports/{公司名}/{公司名}-news-{YYYYMMDD}.md (source L177-L180). After delivery, the temporary team is deleted to free computational resources (source L221-L223).

Summary

• The news-pulse skill implements a multi-agent workflow that investigates four orthogonal dimensions: corporate events, regulatory changes, industry dynamics, and market sentiment.
• Each sub-agent assigns relevance ratings (high/medium/low) to discovered events, which the team-lead converts into weighted attribution scores (🔴🟡⚪).
• The system outputs a merged timeline, evidence matrix, and explicit nature classification (value event, sentiment/technical, unknown, or mixed).
• If evidence is insufficient, the skill explicitly flags "真因不明" (cause unknown) rather than forcing false attribution.
• All analysis is logged to timestamped markdown files in the reports/ directory before the temporary agent team is cleaned up.

Frequently Asked Questions

How does the news-pulse skill determine which events matter most?

The skill weights events based on relevance annotations provided by each sub-agent during the parallel investigation phase. Each agent tags findings with high, medium, or low relevance relative to the specific price movement described. The team-lead agent tallies these tags across all four dimensions and assigns final weights (🔴 high, 🟡 medium, ⚪ low) in the merged timeline, ensuring cross-dimensional validation before prioritizing any single explanation.

What happens when the skill cannot find a clear cause for a price movement?

If no event reaches the high relevance threshold after aggregating all four research dimensions, the team-lead automatically classifies the case as "真因不明" (cause unknown) according to the nature checklist logic (source L186-L188). The final report will explicitly state this uncertainty and recommend continued monitoring rather than providing a speculative attribution.

How long does the news-pulse analysis take to complete?

The four sub-agents execute in parallel background processes, with the complete investigation typically finishing within approximately 10 minutes (source L107-L113). This parallel architecture ensures comprehensive coverage across corporate, regulatory, industry, and sentiment dimensions without sequential delay.

Where are the final attribution reports stored?

Completed reports are persisted as markdown files in the repository's file system under the path reports/{company_name}/{company_name}-news-{YYYYMMDD}.md (source L177-L180). After storage, the temporary agent team created for that specific analysis is deleted to free resources (source L221-L223).

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}