valuation-comps

English · 简体中文 · Site

Trading comps that survive scrutiny. Hand a coding agent a ticker and it runs the full 8-step comparable-company analysis — peer screening, multiple alignment, quartiles, implied valuation range — and writes up what the market is actually paying for this kind of business.

The catch most LLM finance workflows hit: the model does the arithmetic in its head and the numbers drift. This skill refuses that. Every median, quartile and implied share price is computed by a unit-tested script. The agent fetches data and interprets; it never multiplies.

---

What you get

**样本数**：4 家 peer

| 倍数 | 可用样本 | 中位数 | 均值 | 25 分位 | 75 分位 | 隐含每股(25/中/75) |
|---|---|---|---|---|---|---|
| EV/Revenue | 4 | 4.05x | 4.2x | 3.83x | 4.42x | 22.41 / 23.81 / 26.16 |
| EV/EBITDA  | 4 | 13.8x | 14.18x | 13.32x | 14.65x | 13.49 / 14.03 / 14.98 |
| EV/GrossProfit | 0 | — | — | — | — | 无有效样本 |
| P/E | 4 | 21.6x | 22.03x | 20.77x | 22.85x | 51.94 / 54.0 / 57.12 |

Missing data shows up as N/A and gets skipped, never filled with a plausible-looking guess. That one rule is the difference between a comps table you can defend and one that quietly lies.

Quick start

python3 scripts/comps_calc_test.py                       # verify the math (CI runs this too)
python3 scripts/comps_calc.py --input examples/input.json # run the sample

Input is plain JSON — peer multiples plus the target's forecast metrics:

{
  "net_debt": 1200,
  "shares": 800,
  "target": { "revenue": 5000, "ebitda": 900, "eps": 2.5 },
  "peers": [
    {"name": "Peer A", "EV/Revenue": 4.2, "EV/EBITDA": 14.1, "P/E": 22.0},
    {"name": "Peer B", "EV/Revenue": 3.6, "EV/EBITDA": 12.8, "P/E": 19.5}
  ]
}

Point your agent at SKILL.md and it will gather the peers, fill this in, and run it. Zero third-party dependencies — Python stdlib only.

The 8 steps

1. Frame the target and the viewpoint (control vs minority, equity vs EV)
2. Screen peers on business model, size and geography — see references/peer_selection.md
3. Pull multiples from a real data source
4. Align the calendar (LTM / NTM / FY+1) and the currency — see references/multiples.md
5. Compute central tendency — median and quartiles, not the mean
6. Build the sensitivity range
7. Cross-check against DCF
8. Conclude, and name the variable that would make you re-underwrite

Guardrails

• EV multiples include net debt; P/E does not. The script keeps the two lanes separate so they never get mixed in the same sentence.
• Negative or missing denominators are dropped, not patched.
• A >20% gap vs. DCF blocks the conclusion. You must reconcile methodology or peer set first.
• Cheap is not cheap. references/value_trap.md lists the structural reasons a low multiple is correct.

Data sources — no fabrication

1. westock-data — structured A-share / HK / US data with forward consensus
2. akshare-stock, neodata-financial-search
3. WebSearch, company IR, annual reports as fallback

Anything unavailable is written as N/A. The skill would rather hand you a hole than a number.

Part of a three-skill loop

Skill	Question it answers
macro-dashboard	Should I be deploying capital at all right now?
valuation-comps (here)	Is this cheap or expensive relative to its peers?
dcf-quick	What is it worth, what is my downside, how big a position?

Macro switch → screen → price. Each one is independently useful; together they keep you from sizing a position off a valuation you ran in isolation.

License

MIT. Analysis is a tool for thinking, not a substitute for it — the decision stays yours.