🤖 How AI Analysis Works

JSIGNAL combines a general-purpose AI with financial data to deliver investor-oriented summaries of Japanese corporate disclosures.

1. Overview

JSIGNAL ingests publicly disclosed corporate information from EDINET, processes the documents through AI to generate:

📝 Bilingual summaries (Japanese + English) for investors
📊 Sentiment classification (positive / neutral / negative)
⚠️ Risk factor extraction (categorized with severity scores)
💹 Revenue and profit changes (YoY %)
🔮 Forward guidance summary
📈 Growth score (0-10) with classified growth drivers
🏷 Signal tags (earnings, buyback, M&A, guidance up/down, etc.)

📈 Growth Score (paired with Risk)

For earnings-related filings (results, guidance revisions, annual / semi-annual reports), the AI also estimates a growth score (0-10), a short growth outlook, and classified growth drivers. It reads the document (for long reports, the MD&A / business sections) and judges qualitatively and quantitatively:

Score	Meaning
8-10	High growth — double-digit revenue/profit growth, clear upward revision, strong strategy
5-7	Moderate — steady / mild growth
2-4	Low / flat
0-1	Declining / shrinking

Growth drivers are classified from a fixed set: Volume, Pricing, New Product, New Market, Capacity, M&A/Alliance, Margin, Demand, Backlog, R&D. When growth cannot be assessed from the document, no score is assigned (the AI does not guess).

2. AI Models Used

Model	Provider	Use Case
Claude Sonnet 4.6	Anthropic	Main summarization, risk extraction, bilingual generation
Claude Haiku 4.5	Anthropic	Title translation, lighter documents (large shareholders, etc.)

Note: Claude is a general-purpose large language model, not a financial-specialized AI. We configure it to behave as an equity analyst through system prompts and company context injection.

3. Data Sources

EDINET — Financial Services Agency formal filings (annual securities reports, quarterly / semi-annual reports, large shareholding reports, share buyback reports, etc.)
EDINET DB — Third-party API for company financial data enrichment

4. Processing Flow

1Disclosure Ingestion: New disclosures are fetched from EDINET (every 15 min)

2PDF Text Extraction: PDF documents are parsed into plain text

3Quality Check: Text is checked for garbled characters or low Japanese ratio (likely scan/image PDFs are skipped)

4Company Context Loading: Financial data (revenue, ROE, PER, etc.) is loaded from companies database

5AI Summarization: First 3,000-5,000 characters + context sent to Claude with system prompt

6Structured Output: AI returns JSON with summary, sentiment, key points, revenue/profit changes, guidance

7Risk Extraction: For risk-relevant disclosures, separate AI call extracts categorized risk factors

8Storage & Display: Results saved to database, available in dashboard and alerts

5. AI Filter Criteria

To control costs and ensure relevance, AI processing is limited to disclosures matching the following keywords (titles or document types):

Earnings: 決算 / 四半期 / 業績 / 有価証券 / 半期
Positive: 上方修正 (upward revision) / 増配 (dividend increase) / 自己株式取得 (share buyback) / 業務提携 / 株式分割
Negative: 下方修正 (downward revision)
Litigation: 訴訟 / 損害賠償 / 課徴金 / 業務改善命令
Irregularities: 不正 / 不適切 / 不祥事 / 第三者委員会
Loss: 特別損失 / 減損 / 継続企業 (going concern)
Security: 不正アクセス / 情報漏洩
Shareholders: 大量保有 / 変更報告

6. System Prompt (Excerpt)

We configure the AI as an equity analyst with the following persona:

"You are an experienced equity analyst. Generate bilingual (Japanese + English) summaries of Japanese listed company disclosures for retail investors. Use specific currency formats (¥M / ¥B / ¥T for English, 百万円 / 億円 / 兆円 for Japanese). Be concise (max ~120 chars Japanese / 200 chars English). If a number is not in the source, mark it as N/A."

7. Company Context Injection

When available, the AI receives the company's financial profile alongside the disclosure text. Example:

【Company Context】
- Company: TOYOTA MOTOR CORPORATION [7203]
- Industry: Transportation Equipment
- Credit Rating: A (78/100)
- Latest FY 2025: Revenue ¥48.04T / OP ¥4.80T / Net ¥4.77T
- BS: Total Assets ¥93.60T / Equity ¥35.92T / Debt ¥38.79T
- Ratios: ROE 13.6% / Equity Ratio 38% / PER 7.3
- Forecast FY 2027: Revenue +0.6% / OP -20.3% / Net -22.0%

This allows the AI to interpret disclosures in the context of the company's financial health.

8. Limitations

AI output has limitations and may contain errors:

The AI is general-purpose, not specifically trained on Japanese financial data
Numerical values may occasionally be misread or hallucinated
The AI does not have access to real-time market data or share prices
The AI does not perform quantitative modeling (DCF, sensitivity analysis, etc.)
Sentiment classification reflects the AI's interpretation, not market consensus
Risk severity scores are AI estimates, not authoritative ratings
Growth scores are AI estimates and do not guarantee future growth or performance
Translation quality may vary, especially for technical / industry-specific terms

9. Investment Decision Disclaimer

AI analysis is provided for reference only.

All output from JSIGNAL is informational. It is not investment advice, recommendation, or solicitation. Investment decisions are the user's sole responsibility. JSIGNAL does not guarantee accuracy or completeness.

For investment advice, please consult a licensed financial advisor.

10. Cost & Sustainability

To deliver service sustainably, JSIGNAL applies cost-saving measures:

Only EDINET filings matching specific keywords are AI-processed; out-of-scope filings are linked to the EDINET public viewer instead
Document text is truncated to first 3,000-5,000 chars for summarization
Low-priority documents use the lighter Haiku model
Existing risks are translated using Haiku to reduce cost

For technical questions: contact@logic-magic.net