Valuation Consistency Checks

Valuation consistency checks

Most bad valuations contain no wrong number. They are internally inconsistent. Cash flows sit in one currency and the discount rate was built in another. Growth appears that nobody paid for. Terminal growth runs above the economy's. A method the company's own type rules out was used anyway.

Each of those looks defensible when you stare at it alone. Only a check that reads several artifacts at once catches them. That is what this script does.

Treat it as a loop, not a report:

run it  ->  fix what it flags  ->  run it again  ->  no errors left

You are done when the run comes back with zero errors and every warning has a written defence. A warning you cannot defend is an error you have not admitted yet.

When to Use

• Before accepting a DCF, as the mechanical gate between the model and the write-up.
• When reviewing someone else's model, or when a value looks too high.
• When checking terminal value assumptions: growth against the riskfree ceiling, reinvestment against g / ROC, perpetual excess returns.
• When checking growth-versus-reinvestment consistency, currency consistency between cash flows and discount rate, or capital weights and beta range.
• When an equity bridge does not tie, or a company's type imposes hard constraints the model must honour.

How to Run

scripts/validate.py — pure standard library, no installation needed. It reads JSON files by path and prints findings.

python3 ${HERMES_SKILL_DIR}/scripts/validate.py --capital capital.json --dcf dcf.json --classification cls.json
python3 ${HERMES_SKILL_DIR}/scripts/validate.py --dcf dcf.json --capital capital.json --quiet
python3 ${HERMES_SKILL_DIR}/scripts/validate.py --dcf dcf.json --capital capital.json --json

Exit code is 0 when nothing reached ERROR, and 1 otherwise. Nothing else changes the exit code — warnings never fail the run.

The artifacts it reads

Pass any subset. Every check whose inputs are missing is skipped and printed as a SKIP line with the reason, so the gate is useful part-way through an analysis. Running with only a cost of capital artifact is a legitimate thing to do.

Building the capital artifact

The wacc subcommand of cost-of-capital-toolkit returns weights and currency but not the riskfree rate or the beta, because those were its inputs. Merge the input payload and the output into one file:

{
  "currency": "USD",
  "riskfree_rate": 0.0275,
  "levered_beta": 1.0013,
  "wacc": 0.0781,
  "weights": {"equity": 0.8988, "debt": 0.1012, "preferred": 0.0}
}

riskfree_rate may also sit at cost_of_equity.riskfree_rate, and the beta may sit at beta.levered. Either nesting is read.

Building the DCF artifact

The output of dcf.py value is already the right shape, with one gap: it carries no method field. Add one so the constraint check can see which model was used.

python3 ${HERMES_SKILL_DIR}/../dcf-valuation-engine/scripts/dcf.py value --in drivers.json > dcf.json
# then add "method": "fcff" (or "fcfe", "ddm", "excess_return") to dcf.json

Set currency in the DCF drivers file. It is copied through to the result, and without it the currency check has nothing to compare.

Writing the classification artifact

Nothing produces this. You write it after deciding what kind of company this is. It records the hard stops that the company's type imposes, so that a later run cannot quietly ignore them.

{
  "company": "Deutsche Bank",
  "buckets": ["financial-service"],
  "constraints": [
    {"rule": "no-fcff-valuation", "because": "debt is raw material for a bank"},
    {"rule": "require-failure-probability", "because": "crisis bank, equity can be wiped"}
  ]
}

Constraints may be plain strings or objects with a rule key. Three rule names are enforced:

Any other rule name is recorded and counted but not enforced. That is deliberate: writing down a constraint the script cannot test is still better than not writing it down.

The severity model

An ERROR means the model contradicts itself. There is no assumption set under which both halves are true at once. Terminal growth above the riskfree rate is not aggressive, it is inconsistent with the rate you already chose.

A WARN means the model is possible but is making a claim you have not stated out loud. The right response is usually a sentence in the write-up, not a change to the numbers.

What each check does, and what to do about it

currency — ERROR

Compares the currency field across the mandate, capital, forecast and DCF artifacts. Skipped when fewer than two of them declare one.

Currency is a measurement unit, not a value driver. A company valued in reais, in dollars and in francs must come out to the same number once converted. That holds only when cash flows and the rate that discounts them live in the same currency, because a currency's riskfree rate carries its own expected inflation.

Fix: pick one currency and rebuild the other side. To move the rate rather than the cash flows, use convert-rate in cost-of-capital-toolkit, which applies the inflation differential. Do not convert the finished value at the spot rate and call it done.

terminal_growth — ERROR

Compares terminal.growth_rate against riskfree_rate from the capital artifact.

The riskfree rate is expected inflation plus the expected real rate. Nominal economy growth is expected inflation plus expected real growth. The two share the inflation term and their real terms converge in a mature economy, so the riskfree rate is the working ceiling on perpetual nominal growth. Above it, the company eventually becomes the economy.

Fix: lower terminal growth to the riskfree rate or below. Setting it lower is often right — a mature firm in an economy that still contains young firms probably grows slower than the aggregate. Disney's November 2013 valuation set 2.5% against a 2.75% riskfree rate. A negative rate is fine too: Heineken in euros used −0.5% terminal growth against a −0.5% riskfree rate.

If terminal growth genuinely has to be higher, the riskfree rate is the thing that is wrong, and probably in the wrong currency.

terminal_discount_rate — ERROR

Fires when terminal.cost_of_capital is at or below terminal.growth_rate.

The perpetuity has a zero or negative denominator, so it returns no finite number.

Fix: this is arithmetic, not judgment. Either growth is too high or the terminal rate is too low. A terminal cost of capital below the terminal growth rate usually means the fade schedule pushed the rate down too far.

terminal_reinvestment — ERROR

Checks that terminal.reinvestment_rate equals terminal.growth_rate / terminal.return_on_capital.

Growth has to be bought. In the stable phase there is no efficiency growth left to harvest, so reinvestment is the only remaining source of growth, and the identity binds exactly. The classic sleight of hand is to assume capital spending merely offsets depreciation, assume no working capital need, and then apply positive real growth anyway. Zero net reinvestment is consistent only with roughly zero real growth.

An ERROR also fires when the required rate exceeds 100%, which says the firm must raise capital forever just to stand still.

Fix: do not set growth and reinvestment independently. Choose terminal growth and a perpetual return on capital, then let the reinvestment rate follow. Disney: 2.5% / 10% = 25%. Baidu: 3.5% / 15% = 23.33%. Heineken with negative growth: −0.5% / 5% = −10%, meaning the firm releases capital as it shrinks. The value subcommand of dcf-valuation-engine computes this rather than accepting it, so an error here usually means the DCF artifact was edited by hand after the run.

terminal_excess_return — WARN

Compares terminal.return_on_capital with terminal.cost_of_capital. Warns when the spread is wider than 2 points in either direction.

A return above the cost of capital in perpetuity is a claim that competitors never arrive. That claim can be true, and Disney's valuation made it — terminal return 10% against a 7.29% terminal cost of capital, on the argument that brand advantages would not have fully dissipated. But it needs a named barrier to entry, not a spreadsheet default.

A return more than 2 points below the cost of capital is the mirror image. Growth then destroys value, and the faster the firm grows the more it destroys.

Fix: either name the moat in the write-up, or set the terminal return equal to the terminal cost of capital. Setting them equal makes growth exactly value-neutral, which is the honest default. When the spread is negative and intended, terminal growth should be zero or negative to match.

terminal_value_share — WARN

Warns when terminal_value_share_of_total exceeds 90%.

Above that, the answer rests almost entirely on assumptions beyond the forecast horizon, and the explicit forecast is decoration.

Fix: lengthen the explicit forecast until the company has actually reached maturity inside it. Tie the length to how long the competitive advantage lasts, not to the number of columns in a template. If the share stays high with a long horizon, the company may be an option rather than a going concern, and a DCF is the wrong tool.

capital_weights — ERROR

Checks that the values in capital.weights sum to 1, within 1e-4.

Fix: rebuild the weights at market values. Weights that miss usually mean a preferred stock component was dropped, or that book equity crept in where market capitalization belongs.

beta_range — ERROR and WARN

A levered beta at or below zero is an ERROR: equity risk cannot be zero for an operating company. A beta outside 0.3 to 3.0 is a WARN.

Fix: for the WARN, check the comparables and the debt-to-equity ratio used to relever. A very safe or very leveraged firm can sit outside the band legitimately. A regression beta usually cannot — use a bottom-up beta from cost-of-capital-toolkit.

growth_reconciliation — WARN

For each forecast row it computes reinvestment / invested_capital and compares it against that year's revenue_growth. It reports the widest gap, and warns above 2 percentage points.

This is the fundamental growth identity, g = reinvestment rate × return on capital, rearranged. When revenue grows faster than the capital base, the model is assuming efficiency growth — the same assets earning a better return. That is real and it is the cheapest growth there is, but it is a one-off improvement spread over a transition, not a permanent source.

Fix: decide which of the two you meant. If efficiency growth is intended, say so and say when it stops. If not, lower the sales-to-capital ratio so growth costs what it should. A gap running the other way means capital is being spent with no growth to show for it.

Two notes on reading this one. The check compares against revenue growth, so a deliberately expanding margin will open a gap that is not an error. And invested_capital in the DCF rows is end-of-year, while the identity wants start-of-year capital, which understates implied growth by roughly 20 basis points. Neither effect is large enough to explain a wide gap.

equity_bridge — ERROR

Recomputes the walk and compares it with the reported equity_value:

equity value = operating assets − debt − minority interests + cash + non-operating assets

The less_debt and less_minority_interests fields hold positive magnitudes and are subtracted.

Fix: the arithmetic is deterministic, so a mismatch means a line item was edited without the total following. This is where most disputes between competent analysts actually happen, so it is worth getting exactly right. Beyond arithmetic, watch for the double counts the check cannot see: brand value added on top of cash flows that already reflect brand pricing power, or goodwill added as an asset when it is an accounting residual.

value_per_share — ERROR

Checks that bridge.equity_in_common_stock / bridge.shares_outstanding equals value_per_share.

Fix: if you subtracted employee option value in the bridge, divide by actual shares outstanding. Dividing by diluted shares as well counts the options twice. Value the options properly with option-valuation-toolkit and pass the result in.

tax_rate — ERROR

Flags any forecast year whose tax_rate falls outside 0 to 1.

Fix: usually a percentage entered where a decimal belongs. A genuinely negative effective rate belongs in the loss carryforward, which dcf-valuation-engine handles through net_operating_loss_carryforward.

constraints — ERROR

Enforces the three rules listed above and reports how many constraints were recorded.

Fix, by rule:

• no-fcff-valuation fired. For a bank, debt is raw material rather than financing, so operating and financing decisions cannot be separated and there is no meaningful cost of capital. Value equity directly: a dividend discount model by default, or an FCFE model against regulatory capital when payout no longer reflects capacity.
• require-failure-probability fired. A going-concern DCF prices only the branch where the company survives. Add an explicit probability and a distress value rather than raising the discount rate, which double counts. JC Penney: a 20% failure probability with proceeds at 50% of book took operating assets from $4,841m to $4,357m.
• no-earnings-multiple fired. A multiple of negative or trough earnings is meaningless. Use revenue or book value multiples, or normalize the earnings first and say so.

Thresholds

All of them sit as named constants at the top of scripts/validate.py. None is settable from the command line, on purpose — a gate you can loosen per run is not a gate.

A run, and the loop it starts

A Disney-shaped valuation, run against its own cost of capital artifact:

[WARN ] terminal_excess_return   The terminal period assumes a return on capital 2.71
                                 percentage points above the cost of capital, forever...
[WARN ] growth_reconciliation    In year 1 the forecast grows revenue 6.4% while the
                                 reinvestment and return assumed there support about 4.2%...
[INFO ] currency                 All artifacts agree on USD.
[INFO ] terminal_growth          Terminal growth 2.50% is within the riskfree ceiling 2.75%.
[INFO ] terminal_reinvestment    Terminal reinvestment of 25.0% of income is consistent
                                 with 2.50% growth at a 10.00% return.
[INFO ] terminal_value_share     Terminal value is 63% of total value.
[INFO ] equity_bridge            The equity bridge adds up.

0 error(s), 2 warning(s).

Exit 0, so the gate passes. The two warnings are the whole analytical argument. The first says the model claims Disney out-earns its cost of capital forever, which the write-up has to justify by naming the brand. The second says revenue grows faster than the capital base, so returns drift up over the forecast, which has to be either intended or removed.

Break the same set deliberately — shift the DCF currency to EUR, lift terminal growth to 3.5% without touching reinvestment, add 5,000 to the equity value, and record a require-failure-probability constraint — and the run returns five errors and exit 1.

That round trip is worth doing once on your own artifacts. It confirms the gate is reading the files you think it is reading, which matters because of the trap below.

Wiring it into a pipeline

python3 ${HERMES_SKILL_DIR}/scripts/validate.py \
  --mandate mandate.json --classification classification.json \
  --capital capital.json --forecast forecast.json --dcf dcf.json \
  --quiet || { echo "valuation gate failed"; exit 1; }

For a machine consumer, --json returns findings, skipped, error_count and a boolean passed. Store the JSON alongside the valuation. A finding list with zero errors and two defended warnings is a record that the checks ran, which is worth more later than the memory that they did.

One trap. A path that does not exist is treated as an artifact that was not supplied. The run then skips every check that needed it and exits 0. A typo in a filename therefore looks exactly like a clean pass. In a pipeline, test that each file exists before calling the validator, and read the SKIP lines rather than only the exit code.

What it does not check

The gate tests internal consistency. It has no view on whether your assumptions are any good. It cannot tell you that a 22% target margin has never been sustained in the industry, that the sales-to-capital ratio came from a different business, or that the comparables are the wrong ones. It will not notice a return on capital inflated by expensed research spending, because it never sees the accounts.

For those, use financial-statement-normalization before the DCF, and the implied subcommand of dcf-valuation-engine afterwards. Reverse-engineering the market price into the growth or margin it already assumes is the check that tests your assumptions against the world rather than against each other.

Common failures

Verification

validate.py has no selftest; the gate proves itself on your own artifacts. Run it on a clean set and confirm exit 0, then break one field deliberately — shift the DCF currency, or lift terminal growth above the riskfree rate — and confirm the same run exits 1 with the finding named.

python3 ${HERMES_SKILL_DIR}/scripts/validate.py --dcf dcf.json --capital capital.json --quiet; echo "exit $?"