← Files LongbridgeARCHIVED FILE
references/graham-screener.md
14.2 KB · Oct 3, 2026 · 06:02 UTC
# longbridge-graham-screener Prompt-only batch screener. Given a market or index universe, applies Graham's six hard filters (NCAV, PE, PB, dividend yield, debt coverage, earnings stability), scores each name on a 100-point static + dynamic composite, and returns a ranked list of stocks meeting Graham's quantitative criteria, with Graham buy lines, holding-period expectation, and value-trap warnings. ## Cognitive frame (do not skip) Graham screening is **patient arbitrage**, not a dip-buy signal. Hold periods: 6–18 months (explicit catalyst), 1–3 years (sector re-rating), 3–5 years (organic accrual), or never (value trap — exit). Every ranking output **must** carry the waiting-cost (dividend yield) and the holding-period framing; never display a score number on its own. Two failure modes the user must be able to distinguish — and the screener must signal: - **"Cheap but time hasn't come"** → NCAV stable QoQ, keep on the list. - **"Value trap"** → NCAV shrinking ≥3 quarters, persistent insider selling, negative OCF, PMI persistently <50, AR growth outpacing revenue. Override to ⚠️ regardless of static score. ## Workflow 1. **Confirm market + universe**. Ask the user for: (a) market — HK / US / A-share, (b) screening pool — single index (HSI, S&P 500, CSI 300, NDX, etc.) or a watchlist. Default batch ceiling: 300 names per run. 2. **Fetch constituent list** for the chosen index (or watchlist) — see [§CLI](#cli). 3. **Sector pre-classification**: **exclude** banks / insurance / REITs / pure financials / negative-equity / pure-holding shells from the universe before scoring — NCAV does not apply to balance sheets dominated by financial assets, so these are removed, not run through a substitute model. Count the exclusion in the Market Summary "Excluded — NCAV inapplicable" line. Separately, flag IPOs < 2 years (pro-rata earnings-stability sub-score, noted in row) and suspended names (last-trade snapshot, prepend ⏸). 4. **Batch-fetch fundamentals in parallel** (≤20 symbols per wave): balance sheet (annual + last 4Q), income statement (5y annual + 4Q quarterly), cash flow (4Q quarterly), `calc-index` + `quote` snapshot, `dividend`, `ownership`. 5. **Apply hard filters** with user-overridable thresholds — see [§Filters](#filters). 6. **Score each candidate** — static (0–100) + dynamic adjustments + value-trap override. See `references/criteria.md`. 7. **Reconciliation gate per row**: if balance-sheet sum / current-assets sum / shares×price mismatches the reported total by >3%, drop the row from the leaderboard with a "data anomaly" note (do not silently average). 8. **Rank by adjusted score** (descending); default top 10–20 returned. Emit the table + market summary + waiting-cost note + suggested next step. 9. **Append a Data Source Appendix** (mandatory) listing every Longbridge endpoint hit, every WebSearch hit (publisher + URL + access date), fetch timestamps, and any field substitutions. 10. **End every output** with: «筛选结果仅反映量化指标的符合程度，不代表对上述标的的投资建议。投资决策请结合个人情况独立判断。/ Screening results only reflect quantitative criteria compliance and do not constitute investment advice. 本筛选器仅静态硬指标 + 通用动态因子，未考虑公司层面主观估值陷阱。命中标的请用 `longbridge-graham-stock-analysis` 做单股深度诊断。» ## CLI Run `longbridge <subcommand> --help` to verify exact flags before each call — the CLI is the source of truth, do not hard-code flag spellings from memory. ```bash # Universe longbridge constituent <INDEX> --format json # e.g. 000300.SH, HSI.HK, SPX.US # Per-symbol snapshot (run in parallel, batches of ≤20) longbridge calc-index <SYMBOL> --format json # PE / PB / market cap / ROE / dividend yield longbridge quote <SYMBOL> --format json # current price + suspended flag longbridge dividend <SYMBOL> --format json # TTM dividend (waiting-cost) # Per-symbol fundamentals (run in parallel) longbridge financial-report <SYMBOL> --kind BS --report af --format json # 5y annual BS → NCAV longbridge financial-statement <SYMBOL> --kind BS --report qf --format json # last 4Q BS → NCAV trajectory longbridge financial-report <SYMBOL> --kind IS --report af --format json # 5y annual IS → earnings stability longbridge financial-report <SYMBOL> --kind CF --report qf --format json # 4Q CF → OCF rule # Insider / institutional flow (value-trap rule 2) longbridge ownership <SYMBOL> --format json longbridge insresearch <SYMBOL> --format json ``` ### WebSearch fallback (use only when Longbridge has the gap) | Missing input | WebSearch query pattern | | ----------------------------------------------------- | -------------------------------------------------------------------------------------------- | | Industry PMI / inventory cycle | `"<industry> PMI <year>"`, `"中国制造业 PMI <month>"`, `"<industry> inventory cycle <year>"` | | Capacity utilisation | `"<industry> capacity utilization <year>"` | | Sector outlook (qualitative) | `"<sector> outlook <year> site:reuters.com OR site:bloomberg.com OR site:wsj.com"` | | Recent insider transactions (if `ownership` is stale) | `"<ticker> insider selling <year>"`, `"<公司> 大股东减持 <month>"` | Each WebSearch-sourced figure must be tagged with publisher + URL + access date in the **Data Source Appendix**; never silently mix it into a Longbridge column. ## Filters User-overridable. Defaults from Graham's _Intelligent Investor_ with light modernisation; threshold detail and the composite scoring weights live in `references/criteria.md`. | Filter | Graham threshold | Notes | | -------------------------------------- | ----------------------------- | ----------------------------------------------------------------------------------- | | NCAV ratio (market cap ÷ haircut NCAV) | < 1.0 ideal, < 1.5 acceptable | Core. Haircut: cash 100% / AR 75% / inventory 50% / other CA 25% / liabilities 100% | | PE (TTM) | < 10 (defensive) | Rolling TTM | | PB | < 1.5 | — | | Dividend yield (TTM) | > 3% | Waiting-cost compensation | | Current assets ÷ total liabilities | > 2.0 | Debt safety | | Consecutive no-loss years | ≥ 5 | Earnings stability | | Dynamic warning state | none triggered | Optional pre-filter; defaults to "show but flag" | Default rank key = adjusted cigar-butt score (static composite + dynamic adjustments + value-trap override). The user can override the rank key to: static-only, NCAV ratio asc, PE asc, dividend yield desc. ## Special handling | Cohort | Treatment | | ----------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | Banks / insurance / REITs / pure financials / negative-equity / pure-holding shells | **Excluded** — NCAV is not a valid valuation lens for these business models. No substitute model is run; the symbols are dropped from the universe before scoring and counted in the Market Summary "Excluded — NCAV inapplicable" line. If the user wants these analysed, point them to `longbridge-valuation-methodology` or `longbridge-valuation`. | | IPOs listed < 2 years | Earnings-stability sub-score pro-rated on available years. Mark row "数据局限". | | Suspended stocks | Show last-trade snapshot, prepend "⏸ 停牌" / "⏸ Suspended" to the name, exclude from default top-N (still listable on request). | | Reconciliation fail >3% | Drop from leaderboard; surface in a "数据异常待复核" footer list with the failing check named. | ## Output Default leaderboard (top 10–20 rows). Full template, market-summary block, waiting-cost line, subscription hook, and the mandatory **Data Source Appendix** structure live in `references/output.md`. Minimum columns: ``` 排名 代码 名称 静态评分 调整后评分 NCAV比率 PE PB 股息率 Graham买入价 动态预警 数据局限 ``` After the table, every output must include: 1. **Market summary** — universe size, hard-filter pass count, dynamic-clean count. 2. **Waiting-cost line** — average dividend yield of the top-N; the «持有等待期间每年可获约 X% 股息回报作为补偿» line. 3. **Next-step recommendation** — point users to `longbridge-graham-stock-analysis` for any name they want to research further (筛选器未考虑主观估值陷阱因素). 4. **Data Source Appendix** — mandatory; every field, every Longbridge endpoint, every WebSearch hit (publisher + URL + date). 5. **Disclaimer** — full trilingual disclaimer from `references/output.md`, including: 「筛选结果仅反映量化指标的符合程度，不代表对上述标的的投资建议。投资决策请结合个人情况独立判断。/ Screening results only reflect quantitative criteria compliance and do not constitute investment advice.」 > ⚠️ 以上内容仅供参考，不构成投资建议。投资决策请结合自身风险承受能力独立判断。/ 以上內容僅供參考，不構成投資建議。投資決策請結合自身風險承受能力獨立判斷。/ For reference only. Not investment advice. Please make investment decisions independently based on your own risk tolerance. ## Error handling | Situation | 简体回复 | 繁體回覆 | English reply | | --------------------------------------------------------------- | ------------------------------------------------------------------ | ------------------------------------------------------------------ | ---------------------------------------------------------------------------------- | | `command not found: longbridge` | 回退到 MCP；若不可用，请安装 longbridge-terminal。 | 回退到 MCP；若不可用，請安裝 longbridge-terminal。 | Fall back to MCP; if unavailable install longbridge-terminal. | | stderr `not logged in` / `unauthorized` | 请运行 `longbridge auth login`。 | 請執行 `longbridge auth login`。 | Run `longbridge auth login`. | | `constituent` returns empty | 未能获取成分股，请确认指数代码（如 000300.SH / HSI.HK / SPX.US）。 | 未能獲取成分股，請確認指數代碼（如 000300.SH / HSI.HK / SPX.US）。 | Cannot fetch constituents; verify index symbol (e.g. 000300.SH / HSI.HK / SPX.US). | | BS / IS / CF partial fetch for a symbol | 该标的数据不完整，跳过并在「数据异常」脚注列出。 | 該標的數據不完整，跳過並於「數據異常」腳註列出。 | Symbol has incomplete fundamentals; skipped and listed in the data-anomaly footer. | | Industry-cycle data missing (Longbridge + WebSearch both empty) | 仅展示静态评分，动态层标注「数据不足」。 | 僅展示靜態評分，動態層標註「數據不足」。 | Static score only; dynamic layer marked "unavailable". | | Per-row reconciliation gap >3% | 从榜单剔除并在「数据异常」附录列出失败项及差异。 | 自榜單剔除並於「數據異常」附錄列出失敗項及差異。 | Drop from leaderboard; list failing check + gap in data-anomaly appendix. |
SHA-256: 94699f06d1ddff8066d22108168b34e4f42d9349475cf4e0faa31b738cd4a282