fp-dcf（FP DCF 安全）

FP-DCF

Version: v0.4.0

Repository Workflow Notice

• For this repository, GitHub submissions do not use a separate branch workflow.
• Do not create a feature branch for routine updates; commit and publish on the designated branch directly unless the user explicitly asks for a different flow.
• If a default workflow suggests opening a branch or PR, treat that as incorrect for this repository.

Use this skill when the task is to estimate intrinsic value from structured fundamentals and assumption inputs using a disciplined, auditable DCF workflow.

Runtime Contract

This repository is executable when installed as a skill because it includes a concrete CLI entrypoint:

• Primary runner: {baseDir}/scripts/run_dcf.py
• Python module entrypoint: python3 -m fp_dcf.cli
• Sample input: {baseDir}/examples/sample_input.json
• Base Python dependencies for the default one-click path: akshare, baostock, numpy, pandas, yfinance, matplotlib

Preferred execution pattern:

1. Build a JSON payload that matches {baseDir}/examples/sample_input.json.
2. Write that payload to a temporary JSON file in the workspace.
3. Run one of:

- python3 {baseDir}/scripts/run_dcf.py --input /path/to/input.json --pretty - python {baseDir}/scripts/run_dcf.py --input /path/to/input.json --pretty

1. Read the JSON output and present the result to the user.

If the runtime supports stdin piping, this also works:

python3 {baseDir}/scripts/run_dcf.py --pretty < /path/to/input.json

Provider-backed normalization uses a local provider cache by default. To force a fresh pull for the current request, run:

python3 {baseDir}/scripts/run_dcf.py --input /path/to/input.json --pretty --refresh-provider

If the runtime needs an isolated cache location, pass:

python3 {baseDir}/scripts/run_dcf.py --input /path/to/input.json --pretty --cache-dir /path/to/cache

The main runner already returns a compact sensitivity summary and auto-renders both SVG and PNG chart artifacts by default. If the user explicitly asks for a WACC x Terminal Growth chart or sensitivity table, keep using the same main runner so the valuation JSON and artifact paths come back in a single output:

python3 {baseDir}/scripts/run_dcf.py \
  --input /path/to/input.json \
  --output /path/to/output.json \
  --pretty

That one command will default the chart artifact paths to /path/to/output.sensitivity.svg and /path/to/output.sensitivity.png.

Input Shape

The expected JSON object contains:

• ticker
• market
• valuation_model
• assumptions
• fundamentals

Minimum required values for a useful result:

• assumptions.effective_tax_rate
• assumptions.marginal_tax_rate
• assumptions.risk_free_rate
• assumptions.equity_risk_premium
• assumptions.beta
• assumptions.pre_tax_cost_of_debt
• fundamentals.fcff_anchor or fundamentals.ebit

Supported valuation_model values in v0.4.0:

• steady_state_single_stage
• two_stage
• three_stage

For two_stage, the engine continues to support the legacy assumptions.high_growth_rate / high_growth_years fields and also accepts assumptions.stage1_growth_rate / stage1_years as compatible aliases.

For three_stage, the valuation path requires these assumption fields:

• assumptions.terminal_growth_rate
• assumptions.stage1_growth_rate
• assumptions.stage1_years
• assumptions.stage2_end_growth_rate
• assumptions.stage2_years

For three_stage, this optional field is also supported:

• assumptions.stage2_decay_mode with default linear

three_stage applies to the main valuation path and to the separate payload.market_implied_stage1_growth backsolve. payload.implied_growth.model still supports only one_stage and two_stage.

If fundamentals.fcff_anchor is not supplied, the runner computes it from:

• ebit
• da
• capex
• delta_nwc or a fallback working-capital field

If those structured fields are mostly missing, the runner can auto-normalize them from a live provider when:

• provider is set to yahoo, or
• provider is set to akshare_baostock for market=CN, or
• the payload has a ticker but is missing core DCF inputs

The minimal provider-backed input shape is:

{
  "ticker": "AAPL",
  "market": "US",
  "provider": "yahoo",
  "statement_frequency": "A",
  "valuation_model": "steady_state_single_stage",
  "assumptions": {
    "terminal_growth_rate": 0.03
  }
}

For China A-shares, this explicit provider shape is also supported:

{
  "ticker": "600519.SH",
  "market": "CN",
  "provider": "akshare_baostock",
  "statement_frequency": "A",
  "valuation_model": "steady_state_single_stage",
  "assumptions": {
    "terminal_growth_rate": 0.025
  }
}

When market="CN" and Yahoo normalization fails, FP-DCF automatically falls back to akshare_baostock. In that path, AkShare provides statement data and BaoStock provides price history and the latest close.

The payload can also drive normalization behavior through an optional normalization object:

• normalization.provider
• normalization.use_cache
• normalization.refresh
• normalization.cache_dir

The payload can also request sensitivity analysis through an optional sensitivity object:

• sensitivity.enabled
• sensitivity.metric
• sensitivity.detail
• sensitivity.chart_path
• sensitivity.wacc_range_bps
• sensitivity.wacc_step_bps
• sensitivity.growth_range_bps
• sensitivity.growth_step_bps

The payload can also request market-implied stage-1 backsolves through an optional market_implied_stage1_growth object:

• market_implied_stage1_growth.enabled
• market_implied_stage1_growth.lower_bound
• market_implied_stage1_growth.upper_bound
• market_implied_stage1_growth.tolerance
• market_implied_stage1_growth.max_iterations

market_implied_stage1_growth supports only valuation_model=two_stage and valuation_model=three_stage. It solves for the stage1_growth_rate required to match the market price or market EV while keeping the other assumptions fixed. It is not a replacement for the existing implied_growth block.

Sensitivity output is enabled by default. To disable it for a specific run:

• pass --no-sensitivity, or
• set sensitivity.enabled=false

Live provider-backed runs are inherently date-sensitive. Do not hard-code expected valuation numbers when validating this path; validate the presence and plausibility of the returned fields instead.

Core Rules

Tax Policy

• Keep the operating tax estimate for FCFF separate from the marginal tax assumption used in WACC.
• If the statement-level tax rate is available, prefer it for EBIAT/NOPAT.
• If the marginal tax rate is manual or market-defaulted, expose that source in the output.
• Do not silently reuse one tax rate for both paths when the intended sources differ. If a fallback reuse happens, surface it in warnings.

Working-Capital Policy

Use this fallback order and report which path was used:

1. delta_nwc
2. op_nwc_delta
3. nwc_delta
4. cash-flow working-capital change fields such as change_in_working_capital

If all paths fail, flag the result as degraded rather than pretending the estimate is fully reliable.

FCFF Policy

• Prefer a normalized steady-state FCFF anchor for single-stage valuation.
• Prefer NOPAT + ROIC + reinvestment when the driver data is usable.
• Fall back to normalized historical FCFF only when the operating-driver path is incomplete.
• Do not discount historical realized FCFF as if it were a forward forecast.

WACC Policy

• Use explicit sources for risk-free rate, ERP, beta, pre-tax debt cost, and capital weights.
• Prefer explicit capital weights from the input payload when available.
• Apply the marginal tax assumption only to the debt tax shield.
• Attach a warning when key inputs are manual, defaulted, stale, or missing.

Sector Policy

• Financial institutions often produce unreliable FCFF under industrial-company DCF logic.
• When the company is bank-like, insurer-like, broker-like, or otherwise balance-sheet-driven, downgrade or exclude the result unless the workflow explicitly supports that sector.

Output Requirements

Always return:

• valuation model used
• requested_valuation_model
• effective_valuation_model
• degraded
• major assumptions with source labels
• FCFF anchor and anchor method
• working-capital source used
• WACC inputs and capital weights
• enterprise value, equity value, and per-share value when available
• valuation.present_value_stage1
• valuation.present_value_stage2
• valuation.present_value_terminal
• valuation.terminal_value
• valuation.terminal_value_share
• valuation.explicit_forecast_years
• market_implied_stage1_growth when that block is enabled and valid
• provider cache status diagnostics when provider-backed normalization is used
• diagnostics, warnings, and degradation flags
• by default, return a compact sensitivity summary plus both chart file paths

Execution Notes

• Use {baseDir} instead of guessing the install path.
• Prefer writing a JSON file and passing --input over hand-building one-line shell JSON.
• If the payload only contains ticker/market plus light assumptions, rely on provider-backed normalization instead of fabricating fundamentals.
• Use the default provider cache for repeated runs on the same ticker unless the user explicitly asks for fresh data.
• If the user asks for the latest market or statement snapshot, add --refresh-provider or set normalization.refresh=true.
• If the user asks for a valuation sensitivity table or heatmap, prefer the main runner with its default sensitivity output and auto-generated chart path over a separate second command.
• If the caller needs the full numeric heatmap grid in JSON, set sensitivity.detail=true instead of bloating the default output for every run.
• Only use --sensitivity-chart-output or sensitivity.chart_path when the caller explicitly wants to override the default artifact location.
• The default one-click path assumes matplotlib is installed, because PNG/SVG chart artifacts are rendered automatically.
• If the environment intentionally excludes matplotlib, disable sensitivity first with --no-sensitivity or sensitivity.enabled=false before running the main CLI.
• If per_share_value sensitivity is unavailable because shares_out is missing, try --refresh-provider first or switch the sensitivity metric to equity_value.
• If the user only gives high-level valuation preferences, ask for or derive the missing structured inputs before running the script.
• If valuation_model=three_stage is missing stage1_growth_rate, stage1_years, stage2_end_growth_rate, stage2_years, or terminal_growth_rate, fail with a clear error instead of falling back.
• If valuation_model is unknown, fail with an error containing unsupported valuation_model; do not silently remap it to another model.
• Do not silently degrade a requested three_stage valuation into two_stage or steady_state_single_stage.
• Keep implied-growth scope unchanged in v0.4.0: only one_stage and two_stage are valid there.
• Keep market_implied_stage1_growth separate from implied_growth. It is a one-variable stage1_growth_rate backsolve, not a multi-parameter calibration pass.
• Reject market_implied_stage1_growth on steady_state_single_stage with an error containing market_implied_stage1_growth requires valuation_model in {two_stage, three_stage}.
• If the user wants single-stage market-implied growth, route that request through payload.implied_growth.model=one_stage instead.
• Read references/methodology.md only when you need policy detail beyond this file.

Reference Map

Read only what you need:

• references/methodology.md for the valuation methodology
• examples/sample_input.json for the intended input contract
• examples/sample_output.json for the intended output contract
• examples/sample_input_market_implied_stage1_growth_two_stage.json for a minimal two-stage market-implied stage-1 example
• examples/sample_input_market_implied_stage1_growth_three_stage.json for a minimal three-stage market-implied stage-1 example
• examples/cn_moutai_single_stage.json for a CN provider-backed single-stage example

Quality Bar

• Prefer explicit assumptions over hidden heuristics.
• Prefer auditable fallbacks over brittle elegance.
• Label degraded results clearly.
• Be conservative when provider data is incomplete or inconsistent.
• If a result depends heavily on terminal value, include that in the diagnostics.