longbridge-buffett-moat-stock-screener

longbridge-buffett-moat-stock-screener

Prompt-only Buffett-style screener. Given a market / sector / preference query (no specific ticker), applies Buffett's two-layer model — hard quantitative filter then qualitative moat scoring — and returns a ranked candidate list (3–5 cards) with moat-type tags, Buffett-attitude verdicts, and one-click jumps into longbridge-buffett-moat-analyzer for deep diagnostics. Every figure traces to a row in the mandatory Data Source Appendix at the end of the output.

Response language: detect the user's input language (Simplified Chinese / Traditional Chinese / English) and render the entire report — every card, label, narrative paragraph, education block, appendix row, and disclaimer — in that one language. Do not mix languages within a single output. The output template in references/output.md is shown in English for reference; translate it as a whole into the user's language using the label-translation lookup in that file. The error/source tables inside this SKILL.md remain 3-column because they document what the skill says under each language — that 3-column form is for the skill's reference docs, not for the user-facing report.

When to use

• "帮我找几只巴菲特会买的股票" / "幫我找幾隻巴菲特會買的股票" / "find a few stocks Buffett would buy"
• "消费股里巴菲特风格的有哪些" / "消費股裡巴菲特風格的有哪些" / "which consumer names are Buffett-style"
• "我想找一只 10 年都不用动的股票" / "我想找一隻 10 年都不用動的股票" / "I want a 10-year-hold stock"
• "有没有类似茅台这种定价权强的公司" / "有沒有類似茅台這種定價權強的公司" / "any companies like Moutai with strong pricing power"
• "新能源里有没有符合巴菲特标准的" / "新能源裡有沒有符合巴菲特標準的" / "any new-energy names that meet Buffett's criteria"
• "ROE 连续 10 年超过 20%、PE 低于 25 的 A 股" / "ROE 連續 10 年超過 20%、PE 低於 25 的 A 股" / "A-shares with 10y ROE > 20% and PE < 25"
• "美股里巴菲特真实持仓过的股票" / "美股裡巴菲特真實持倉過的股票" / "US stocks Buffett has actually held"

For single-stock deep diagnostic on a name the user already picked, use longbridge-buffett-moat-analyzer. For Graham-style cigar-butt / NCAV screening, use longbridge-graham-screener. For broader low-PE / low-PB / high-ROE value (not moat-centric) use longbridge-value-screen. For high-dividend screens use longbridge-dividend-screen.

Cognitive frame (do not skip)

Screening is the upstream entry to the Buffett workflow: this skill generates candidates from zero; the downstream longbridge-buffett-moat-analyzer verifies a single name in depth. Both share the same moat lens — only the interaction shape differs.

Two principles must surface alongside every leaderboard:

1. Holding-period expectation — Buffett's reference is "ideally forever, at minimum 3 years." If the user's capital horizon is < 3 years, warn that this framework is the wrong lens regardless of how clean the screen looks.
2. Quality first, price second — "A great business at a fair price is far better than a fair business at a great price." A passing row tells the user the business is Buffett-grade; whether this price is sane is a Dimension-4 (valuation) check that may downgrade the verdict to "watch and wait for a better price". Never collapse quality and price into one number.

Failure modes the screener must flag honestly:

• "Great business, wait for price" → quality passes but valuation 偏贵 / 高估 → 🟡 watchlist with target-price band, not a buy.
• "Cheap but no moat" → low PE/PB but moat narrow or absent → not a Buffett candidate; redirect to longbridge-graham-screener / longbridge-graham-stock-analysis.
• Sector-disqualified industries (e.g. airlines — Buffett publicly called the industry "value-destroying") → return an honest empty result with redirect to a closer-to-Buffett sector, not a forced top-N.

Workflow

1. Clarify the query in at most one quick turn. Three input shapes are supported (see §3.2 of the design doc):
2. Preference-based (most common): "find me consumer stocks worth long-term holding".
3. Sector / theme-based: "Buffett-style names in new energy" / "Buffett's actual US holdings".
4. Condition-based (advanced): "A-shares with 10y ROE > 20% and PE < 25".

If the user just says "recommend stocks" with no constraint, ask 1–2 framing questions (preferred market: A / HK / US? prefer steady — consumer/healthcare/financials — or growth — tech/new-energy?) before screening.

1. Resolve the universe:
2. Sector / theme query → start from the index or sector universe via longbridge constituent <INDEX> or longbridge sector-screener candidate list.
3. Preference / condition query without a stated market → default to the market the user most often references (A-share if Mandarin / Cantonese trader signal, US for English, HK if the user mentioned 港股). Echo the chosen universe back in the Market Summary.
4. Cap batch size at 300 names per screen.
5. Sector triage — before any filter:
6. Excluded (no scoring, drop with a one-line explanation in the Market Summary): airlines (Buffett: "value-destroying"), pre-revenue / pure-loss biotech (no track record), ST / 退市风险 names, listed < 5 years, pure-shell / negative-equity. If the user explicitly asked for one of these (e.g. "Buffett-style airline"), return an honest empty result with a redirect to the most-similar non-excluded sector (e.g. "consumer staples with pricing power" or "regulated utilities with stable returns").
7. Sector-adjusted hard filter: banks / insurance / brokerage — replace FCF / leverage rules with ROA / NIM / NPL / CAR (note the substitution per row).
8. Listed 5–10 years: pro-rate the earnings-stability sub-score and flag "history-limited" in the row note.
9. Fetch raw data via Longbridge CLI first (parallel, ≤20 symbols per wave). See [§CLI](#cli). MCP fallback if longbridge is missing (see [§MCP fallback](#mcp-fallback)). WebSearch only for items genuinely outside Longbridge: industry outlook / disruption signals, brand-strength surveys, Buffett's own 13F holdings disclosure, qualitative management track record. Every WebSearch hit gets a publisher + URL + access-date row in the appendix.
10. Layer 1 — Hard quantitative filter (binary pass/fail). Symbols failing ≥ 2 filters drop out by default. See [§Filters](#filters) for the table; full thresholds in references/criteria.md.
11. Per-row reconciliation gate: if balance-sheet sum / current-assets sum / shares×price mismatches the reported total by >3%, drop the row from the leaderboard with a "数据异常" note in the data-anomaly footer. Do not silently smooth.
12. Layer 2 — Qualitative moat scoring on every name that passed Layer 1 plus reconciliation. Five-dimension weighted composite (0–100):
13. Moat type & width (35%) · Capital allocation (20%) · Earnings predictability (20%) · Valuation reasonableness (15%) · Long-term industry runway (10%).

Full rubric in references/criteria.md. Each dimension also gets a 1–5 star rating shown on the candidate card.

1. Verdict matrix — combine Layer 2 quality stars (moat + financials) with valuation tier. See references/criteria.md §Verdict matrix. Maps to one of three card verdicts:
2. 🟢 符合巴菲特选股标准 / Meets Buffett criteria — wide moat + clean financials + price 充足/一般.
3. 🟡 部分符合，关注估值变化 / Partially meets criteria — wide moat + clean financials + price 偏贵.
4. 🔴 当前不符合标准 / Does not meet criteria — wide moat but price 高估, OR moat narrow at any price (redirect to Graham).
5. Holding-period mapping — derive expected min hold from moat width (★★★★★ → 5y+, ★★★★ → 3–5y, ★★★ → 1–3y, ★★ or below → not a Buffett candidate). See references/criteria.md.
6. Rank and emit 3–5 candidate cards (not a giant leaderboard — the design doc's deliberate cap). Follow the candidate-card template in references/output.md.
7. Mandatory closing blocks (every output, no exceptions):
8. Selection rationale (2–3 sentences on why these names, current market caveat, deep-dive priority).
9. Holding-period & user-education block