Korean Analyst Report Discovery

Find sell-side broker reports for a given KRX ticker, extract structured metadata per report (broker, analyst, date, rating, target price, PDF URL), and write a dated JSON file for downstream kr-analyst-report-fetch to consume.

When To Use

• User asks about street consensus / target prices / broker coverage for a KRX stock (증권사 리포트, 컨센서스, 애널리스트 TP).
• Before writing a full memo or pre-earnings note where sell-side framing is relevant.
• Before a post-earnings note where consensus-vs-print matters.

Do not use this skill to:

• Fetch generic financial news or company IR decks.
• Discover non-Korean broker coverage.
• Produce a thesis — that belongs downstream in kr-stock-analysis.

Source Policy

1. Hankyung Consensus (consensus.hankyung.com) is the primary source. Richer metadata (analyst name, target-price deltas, broker list) and the direct PDF links are usually unauthenticated.
2. Naver Pay Research (finance.naver.com/research/company_list.naver) is the fallback when Hankyung returns an empty list or fails.
3. Login-gated reports stay in the index with pdfUrl: null, requiresAuth: true, and are counted but not downloaded downstream.

Workflow

1. Parse inputs. Require --company and --ticker. Default --lookback-days 365, --source hankyung (other values: naver, both), --max-reports 0 (unlimited).
2. Check cache. If .tmp/analyst-report-cache/index/<ticker>/<YYYY-MM-DD>.json exists for today, prefer it unless --no-cache.
3. Compute window. endDate = today, startDate = today - lookbackDays.
4. Scrape primary source via lib/hankyung.js:
   • Build list URL with the date window and ticker/company as filters.
   • Paginate by now_page=1..N until no new rows appear or the earliest row crosses startDate.
   • Parse each row into {reportId, broker, analyst, publishedDate, title, rating, targetPrice, currency, pdfUrl, landingUrl}.
   • Probe PDF URLs with a HEAD-like downloadFile trial (or a cheap browseText) to detect login walls; set requiresAuth: true where appropriate.
5. Fallback to Naver Pay Research via lib/naver-research.js if the primary returned zero usable rows OR the caller passed --source naver. When --source both, run both and dedupe by (broker, publishedDate, title).
6. Deduplicate by (broker, publishedDate, title). When two sources disagree on target price, prefer Hankyung (closer to the original filing).
7. Filter by window. Drop any row older than startDate.
8. Sort by publishedDate descending, then by broker ascending.
9. Write a dated JSON file to the --output path.

Read references/workflow.md for source-specific URL shapes and row-parsing rules. Read references/output-format.md for the JSON schema.

Operating Rules

1. Never invent a target price, rating, or analyst name. If the row didn't yield it, the field is null.
2. Currencies: assume KRW for KOSPI/KOSDAQ tickers. If the row explicitly shows USD, set currency: "USD" and keep the numeric TP as-is.
3. Target prices are normalized to integer KRW when the source uses 원 units. Strip commas and currency symbols before parsing.
4. Rating normalization: map Korean ratings to canonical English tokens: 매수/BUY → BUY, 시장수익률/Market Perform → HOLD, 비중확대/OVERWEIGHT → BUY, 중립/Neutral → HOLD, 매도/SELL → SELL. Preserve the original string under ratingRaw.
5. Respect a 1-second inter-request delay (enforced by kr-web-browse).
6. Never scrape analyst emails, phone numbers, or affiliated-firm internal IDs — just the public metadata.
7. Keep the output deterministic: same inputs, same list. Rank by (publishedDate desc, broker asc, reportId asc).
8. Fail loud if neither source returns any rows and no cache exists. Do not silently produce an empty index.

Scripts

discover-reports.js

node skills/kr-analyst-report-discover/scripts/discover-reports.js \
  --company "엘앤에프" --ticker 066970 \
  --lookback-days 365 \
  --output .tmp/analyst-report-cache/index/066970/2026-04-18.json

Optional flags: --source hankyung|naver|both (default hankyung), --max-reports N (default 0 = unlimited), --cache-dir PATH, --no-cache, --verbose, --no-probe-auth (skip the per-PDF auth probe for speed).