dsh-valuation-service

A deterministic financial-data and valuation plugin for DeepSeek Harness. It reads a workpaper layout of 01.json / 02.json / 04.json, computes six ratio groups, forecast ratios, DCF input slots, comparable PS / PE and IRR, and writes one unified schema per project set.

It registers one agent tool, valuation_service. No AI inference, no invented numbers: anything missing is left blank and marked missing for a human or a downstream research step to fill.

What it computes

Block	Content
ratios.historical	profitability, solvency, efficiency, cash flow, valuation, growth
ratios.forecast	forecast-period profitability / growth / expense; BS- and CF-dependent ratios marked missing
dcf	full input slots (WACC, g, tax, EBIT, D&A, capex, ΔWC, net debt), outputs, missing list, hand-off anchors
comparable	PS and PE in parallel: trimmed mean, all-sample mean, 1-year mean, pre-money multiple, premium
irr	target / conservative / optimistic, closed-vs-open judgement, deviation
assumption_diff	material assumptions vs assumptions computed from actuals, field by field
pending / gate	unresolved conflicts and C2–C7 gate status when the workpapers carry them

Input layout

<base>/
├── 项目A/
│   └── json/
│       ├── 01.json   # financial.raw (BS/IS/CF) + financial.ratios
│       ├── 02.json   # assumptions.forecast + assumptions.comparables + assumptions.valuation
│       └── 04.json   # validation.irr_recompute
└── 项目B/
    └── json/...

--base defaults to $DSH_VALUATION_BASE or ./数据底表. Projects are discovered dynamically: every directory with all three JSON files is included. No project list is hard-coded.

01.json may use either raw = {bs, is, cf} or report-title sections (e.g. "一、资产负债表 · 资产"); both are normalized.

Install

dsh plugin add https://github.com/zirscard/dsh-valuation-service

Or from a local directory:

dsh plugin add link:/absolute/path/to/dsh-valuation-service

Restart DSH; a new session gets the valuation_service tool.

Tool usage

{
  "base": "/path/to/数据底表",
  "projects": "项目A,项目B",
  "write": true
}

• base (string, optional): workpaper root; relative paths resolve against the session working directory.
• projects (string or array, optional): restrict to these project directory names.
• write (boolean, optional, default false): write <base>/valuations-service.json.
• out (string, optional): override the output path.

The tool returns a per-project summary (DCF status and missing list, PS/PE status, IRR closed flag, assumption-diff count, gate) plus the output path.

CLI usage

The Python entry point is usable without DSH (standard library only):

python3 core/valuation_service.py --base /path/to/数据底表 --dry-run
python3 core/valuation_service.py --base /path/to/数据底表 --projects 项目A
python3 core/valuation_service.py --base /path/to/数据底表 --write
python3 core/valuation_service.py --base /path/to/数据底表 --write --json

--json prints one machine-readable summary object; --write writes the aggregate file; without --write it is a dry run.

Unified schema

{
  "schema": "m1.valuation-service/v2",
  "project_id": "...",
  "project_name": "...",
  "currency": "CNY",
  "method": "...",
  "financial_inputs": {"period": "...", "is": {}, "bs": {}, "cf": {}},
  "ratios": {"historical": {}, "forecast": {}},
  "dcf": {"status": "computed|partial|unavailable", "inputs": {}, "outputs": {}, "missing": [], "m3_handoff": {}},
  "comparable": {"ps": {}, "pe": {}},
  "irr": {"target": null, "conservative": null, "optimistic": null, "closed": false, "judge": "", "dev": null},
  "assumption_diff": [],
  "pending": [],
  "gate": {}
}

The aggregate file wraps records as {"schema": "m1.valuation-service/aggregate/v2", "generated_at": "...", "base": "...", "projects": [...], "errors": [...]}.

Formulas

• Ratios: gross / operating / net margin, ROE, ROA; current / quick / cash ratio, debt-to-equity, debt-to-assets, interest coverage; asset / inventory / receivables turnover and days; OCF ratios; PS / PE / EV / EV-EBITDA; revenue and net-income growth.
• DCF: FCFF = EBIT × (1 − t) + D&A − capex − ΔWC; TV = FCFF_last × (1 + g) / (WACC − g); EV = Σ PV(FCFF) + PV(TV); equity = EV − net debt.
• Comparable: trimmed mean (10% each side when n ≥ 5), all-sample mean, 1-year mean, pre-money multiple, premium.
• IRR: (exit value / invested capital) ^ (1 / years) − 1.
• Assumption diff: material assumption vs module-computed actual, per field.

Ratio formulas are adapted from anthropics/claude-cookbooks · analyzing-financial-statements/calculate_ratios.py (MIT), mapped onto Chinese three-statement labels.

Adaptation layer

Data variants are handled in one place instead of per project:

Variant	Function	Rule
raw layout ({bs,is,cf} vs report-title sections)	_normalize_raw	normalize by 资产负债表 / 利润表 / 现金流量表 keywords
units (thousand / ten-thousand / million CNY)	_unit_for	field-name suffix first, then forecast.unit
field-name variants	_find_param	keyword match + exclusion words
total vs line item (资产总计 vs 流动资产合计)	_find_val	totals require an exact match, no substring fallback
self-comparison in the comp set	_included	drop rows whose treatment says 自比 / 剔除
no market PE multiple	comparable.pe.status=assumption	use the assumption range explicitly, never present it as a market multiple

If a new variant is structural (layout / unit / naming), add it to the adapter. If it is a measurement convention, add it to the skill formulas. If data is missing, leave it blank and mark missing.

DCF hand-off

Missing DCF inputs stay blank; dcf.m3_handoff.anchors lists what each one needs:

Missing input	Anchor to fill
WACC	risk-free rate / beta / equity risk premium / capital structure
g	long-run GDP growth / industry terminal growth
capex	capex plan / construction in progress / capacity expansion
D&A	depreciation and amortization policy
ΔWC	receivables / inventory / payables turnover forecast
tax rate	effective tax rate
net debt	interest-bearing debt − cash

Test

python3 test/run_tests.py

The test builds a synthetic three-project workpaper set in a temporary directory (no client data), runs the service, and checks ratio values, DCF missing slots, comparable status, IRR and the written aggregate file.

Boundaries

• No AI inference, no model calls, no network requests.
• No client data or project names are shipped in this repository.
• Missing inputs are marked, not fabricated.
• It does not replace a valuation model: it standardizes inputs and keeps the slots.

License

MIT. Ratio formulas adapted from anthropics/claude-cookbooks (MIT); see THIRD_PARTY_NOTICES.md.