longbridge-graham-screener

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.

Response language: match the user's input language — Simplified Chinese / Traditional Chinese / English.

When to use

• "港股里现在有没有 NCAV<1 的股票" / "港股裡現在有沒有 NCAV<1 的股票" / "any HK stocks with NCAV < 1"
• "给我 A 股 PE 最低的 20 只股票，格雷厄姆标准" / "給我 A 股 PE 最低的 20 隻股票，格雷厄姆標準" / "top 20 lowest-PE A-shares by Graham standard"
• "帮我筛一下符合格雷厄姆防御型投资者条件的美股" / "幫我篩一下符合格雷厄姆防禦型投資者條件的美股" / "screen US stocks meeting Graham defensive-investor rules"
• "沪深300里的烟蒂股" / "滬深300裡的煙蒂股" / "cigar-butt names in CSI 300"
• "每周更新港股 NCAV 排行榜" / "每週更新港股 NCAV 排行榜" / "weekly HK NCAV leaderboard"

For single-stock deep diagnostic use longbridge-graham-stock-analysis. For broader low-PE / low-PB / high-ROE value (not NCAV-centric) use longbridge-value-screen. For high-dividend-only screens use longbridge-dividend-screen.

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.

# 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:

``