Special Situation Models

Special situation models

A standard discounted cash flow model assumes a lot. It assumes the firm survives to reach stable growth, that its debt is a financing choice, that its owner is diversified, that this year's earnings say something about a normal year, and that there are earnings at all.

Five kinds of company break one of those assumptions. This skill repairs the specific break. It does not replace the DCF, and for four of the five branches the answer still runs through dcf-valuation-engine.

When to Use

• When valuing a bank, insurer or any firm whose raw material is money: equity excess return or FCFE against regulatory capital.
• When valuing a startup or other young negative-earnings firm, a distressed firm, or a private business.
• When valuing a miner or deep cyclical from mid-cycle normalized earnings or a commodity price link.
• When pricing an IPO or a sale to a public buyer, walking a private owner's value to an offer price line by line.
• When asked about probability of distress, total beta, illiquidity discount, excess return models, normalized earnings or negative-earnings valuation.

Which branch applies

Answer this before running anything. The branch is a judgment about the company, and the script does not guess it for you.

More than one can apply. A young company usually needs branch 5 and branch 1 together. A private cyclical needs 3 and 4. Run them in the order above, because each later branch consumes the output of the earlier one.

One transition sits alongside the five: a private company on its way to a public listing, or to a sale to a listed buyer. That is the ipo subcommand, and it consumes branch 3.

What each branch forbids

This is the part that gets skipped, and it is where the damage happens.

A bank gets no FCFF valuation and no optimal-debt-ratio analysis. For an industrial firm, debt is a source of capital, so you can separate operating from financing decisions. For a bank, debt is raw material. Deposits are the input the business transforms into loans. There is no meaningful firm value and no meaningful cost of capital, so do not run dcf-valuation-engine value on a bank, and do not run cost-of-capital-toolkit debt-schedule on one. Regulatory capital governs the financing mix, not a tax-shield calculation. A bank that breaches its capital ratio can be taken over and closed however good its earnings look.

Distress risk does not belong in the discount rate. Raise the rate and probability-weight the value and you have counted the same risk twice. Pick the probability weight.

A total beta does not belong in a valuation for a diversified buyer. It systematically undervalues the business and hands the surplus to the buyer. The same applies to the illiquidity discount: a listed acquirer's own shareholders can sell, so no discount applies.

A trailing return on equity does not survive re-regulation. If the regulator is about to demand more capital, the return earned on the old base is not the sustainable one.

Normalized earnings and a recovery in growth are the same recovery. Normalize the base year or forecast the recovery, not both.

A young company with no failure branch is not a valuation. Roughly two-thirds of startups are gone within seven years, and the value of the survivors is not the value of the population.

How to Run

scripts/special.py — pure standard library, no installation needed. Every subcommand takes JSON on stdin (or --in FILE) and prints JSON.

python3 ${HERMES_SKILL_DIR}/scripts/special.py <subcommand> --example    # show the input shape
python3 ${HERMES_SKILL_DIR}/scripts/special.py <subcommand> --in payload.json
python3 ${HERMES_SKILL_DIR}/scripts/special.py selftest                  # verify the engine

Run selftest after editing anything. It reproduces the source models to nine decimals and checks identities that a wrong port would break. Every case it runs is written up in references/worked-examples.md, with the payload alongside the answer.

1. Distress adjustment

A going-concern model prices only the branch where the firm lives. The repair is an explicit probability-weighted blend of two outcomes.

value = going-concern value × (1 − p) + distress-sale value × p

Getting the probability

Five sources, in increasing order of information content. Pass whichever you have.

The bond route prices the promised coupons and principal, weights each by the probability the firm survives to pay it, and discounts at the riskfree rate. All the credit risk sits in the survival weights. Discounting at the bond's own yield instead would count it twice.

echo '{"bond": {"coupon_rate": 0.06375, "maturity_years": 7,
                "riskfree_rate": 0.03, "market_price": 529},
       "horizon_years": 10, "going_concern_value": 8.12,
       "distress": {"basis": "explicit", "proceeds": 2769,
                    "debt_face_value": 11000}}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/special.py distress

Two modelling assumptions are baked in: recovery in distress is zero, and discounting is at the riskfree rate. Both are choices. Assuming a positive recovery would lower the implied probability for the same price.

Report the cumulative probability over your forecast horizon, not the annual one. For Las Vegas Sands the difference between 13.54% a year and 76.66% over ten years is the difference between a $7 stock and a $2 stock.

Getting the distress-sale value

Set distress.basis:

• book — proceeds are a percentage of book equity plus book debt. The default recovery is 50%. Cut it when the economy is weak and every peer is selling the same assets at once.
• going_concern — proceeds are a percentage of the going-concern value, for a firm that would be sold intact rather than broken up.
• explicit — you have an estimate of what the assets would fetch.

Add debt_face_value when you are valuing equity. Equity in distress is a residual: if proceeds fall short of what lenders are owed, shareholders get nothing.

The partial wipeout

A bailout can save the firm and destroy the equity. Set equity_loss_fraction instead of a distress branch, and the blend becomes a single haircut:

adjusted value = going-concern value × (1 − probability × loss fraction)

Boeing in March 2020: a 20% failure probability with a 50% loss to equity is a 10% haircut, not a 20% one.

Where the number goes

dcf-valuation-engine already has a failure block that does the simple blend. Feed it the cumulative probability from here. Use this subcommand's own blend when equity is a residual against the face value of debt, or when the case is a partial wipeout — neither of those fits in the DCF engine's block.

2. Financial service firms

Value the equity directly. See references/financial-service-firms.md for the full argument and for the choice among the three equity models.

value of equity = current book equity + PV of (return on equity − cost of equity) × book equity

Book value is close to irrelevant for an industrial firm. For a bank it is the opposite: assets are marked to market, and regulatory ratios are computed on book equity. That gives a bank a hard definition of reinvestment, which industrial firms lack.

Two reinvestment modes:

retention — book equity grows by retained earnings. This is the standard equity excess return model. Give a high-growth return on equity, a retention or payout ratio, and a stable block. Each driver fades to its stable value over the second half of the horizon unless you set fade_second_half: false.

regulatory_capital — book equity is whatever the capital ratio requires. Give the risk-adjusted assets, their growth rate, and the path of the Tier 1 or CET1 ratio. Use this when ratios are moving or the bank is in crisis. Book equity often sits above regulatory capital, so pass required_book_equity as a year-by-year list instead when the bank has disclosed the path. Subtract any one-off hit to capital, such as a fine, with one_off_capital_hit.

echo '{"book_equity": 64609, "forecast_years": 10,
       "reinvestment": "regulatory_capital",
       "regulatory_capital": {"risk_adjusted_assets": 445570, "asset_growth": 0.01,
                              "capital_ratio": {"start": 0.1241, "end": 0.1567,
                                                "converge_by": 10}},
       "return_on_equity": {"start": -0.137, "end": 0.0944, "converge_by": 10},
       "cost_of_equity": 0.102, "shares_outstanding": 1386,
       "probability_of_equity_wipeout": 0.10,
       "stable": {"return_on_equity": 0.0944, "growth_rate": 0.01,
                  "cost_of_equity": 0.0944}}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/special.py excess-return

Reading the output

value_of_equity and value_of_equity_via_fcfe should match, and route_difference should be near zero. Residual income and discounted free cash flow to equity are the same model written two ways. A gap means the book-equity rollforward disagrees with the cash flows, which is the classic bank-model error.

Early fcfe is deeply negative for a bank rebuilding capital. That is correct, not a bug. Every increase in the capital ratio is reinvestment and it comes out of shareholder cash flow.

Three inputs carry the answer. Set the sustainable return on equity, not the trailing one. Anchor the terminal return on equity on the cost of equity unless a franchise justifies more. Anchor the target capital ratio on the peer distribution rather than the regulatory minimum.

3. Private company adjustments

Two separate repairs, and the script runs both. Detail in references/private-company-adjustments.md.

Total beta, for an owner who holds nothing else:

total beta = market beta / correlation with the market
correlation = square root of the average R-squared of the comparables' regressions

The square root is not optional. R-squared is a share of variance and betas are built from standard deviations. Dividing a 1.18 beta by an R-squared of 0.25 gives 4.72; dividing by the correlation of 0.50 gives 2.36.

An illiquidity discount, applied to equity value after the DCF:

echo '{"buyer": "private",
       "cost_of_equity": {"unlevered_market_beta": 1.18, "r_squared": 0.25,
                          "debt_equity_ratio": 0.1433, "tax_rate": 0.40,
                          "riskfree_rate": 0.0425, "equity_risk_premium": 0.04},
       "illiquidity": {"revenues_millions": 1.2, "positive_earnings": true,
                       "cash_to_firm_value": 0.05},
       "equity_value": 520990}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/special.py private

Three routes come back. The flat 25% rule of thumb, the Silber-refined restricted-stock number, and the bid-ask spread regression. They disagree by a lot — on the restaurant above, by $83,000 on a $521,000 business. Prefer the bid-ask route: it is the most firm-specific and it draws on an unbiased sample. Say which one you used.

Set buyer correctly. A private buyer gets the discount. A public or ipo buyer does not, because the exit already exists, and the script zeroes the discount for them.

Assemble the cost of capital in cost-of-capital-toolkit. Its rating subcommand builds a synthetic cost of debt from interest coverage, and wacc does the weighting. Lever the beta and weight the cost of capital at the same debt-to-equity ratio.

4. Cyclical and commodity normalization

Keep macro out of the micro. If you build your own oil forecast into the valuation, the answer is a blend of two opinions and no reader can tell which is doing the work.

When a usable price driver exists, regress revenues on the commodity price and run the valuation at today's market price or the futures strip:

echo '{"commodity": {"history": {"price": [...], "revenue": [...]},
                     "price": 40, "price_ladder": [30, 40, 50, 60, 80]},
       "base_operating_margin": 0.0301, "target_operating_margin": 0.0935}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/special.py cyclical

Pass intercept and slope instead of history if you fitted the regression elsewhere. The output reports the R-squared and warns when the price explains less than half the variation in revenues, which means the link is too weak to use.

Then state the answer as "worth X at today's commodity price". State your macro disagreement separately, and use the price ladder to quantify it.

When no usable price driver exists, normalize earnings instead. Three approaches:

Approach 3 is the default. The aggregate margin is the sum of EBIT over the sum of revenues, not the average of the yearly margins. The two differ and the aggregate is the one to use.

Normalization is only legitimate when the trouble is temporary. Evidence for: the sector is in a known downturn, peers show the same pattern, the firm earned normal margins for years, and the balance sheet can survive until recovery. Evidence against: falling market share, a structural demand shift, leverage that forces asset sales. If the trouble is permanent, go to branch 5 or branch 1 instead.

The normalized EBIT changes everything downstream. Recompute interest coverage, the synthetic rating, the cost of debt and the return on capital with it. That chain lives in cost-of-capital-toolkit rating, and lease capitalization lives in financial-statement-normalization. Those two are circular with each other, so iterate.

Read dcf_drivers from the output for the base revenue and margin glide to paste into the DCF payload. Set the terminal return on capital to the firm's own long-run average, not a trough number.

5. Young companies with negative earnings

Work backwards from a mature end-state. Current earnings are meaningless and there is no history to extrapolate, so specify the revenue scale, margin and return the business will have when it grows up, then let the intermediate years converge to it.

python3 ${HERMES_SKILL_DIR}/scripts/special.py young-company --example > drivers.json
python3 ${HERMES_SKILL_DIR}/scripts/special.py young-company --in drivers.json > out.json

The two shapes

Revenue. Either state the high-growth rates directly (mode: "explicit", high_growth_rates: [1.50, 1.00, 0.75, 0.50, 0.30]), or state the revenue you expect in a given year (mode: "target_revenue") and let the engine solve for the growth rate that reaches it. Either way, growth then fades linearly to the stable rate by the final year.

Cross-check the answer twice. Excess growth over the industry average dies within roughly five years of an IPO. And market_share_check in the output compares year-10 revenue with the total addressable market — if the implied share is implausible, the growth path is wrong.

Margin. style: "halving" closes a fixed share of the gap to the target each year, which is the shape the Amazon January 2000 valuation uses. style: "linear" runs a straight line to a stated convergence year, which is the Ginzu convention. Set the target margin from the mature sector's margin, not from the company's story.

Survival

Required, not optional. Give survival.probability_of_failure directly, or survival.sector to look up long-run survival, or survival.years_since_founding for the startup table. The probability lands in the payload's failure block.

The handoff to dcf-valuation-engine

dcf_payload in the output is a complete, valid input to dcf-valuation-engine, run as python3 ${HERMES_SKILL_DIR}/../dcf-valuation-engine/scripts/dcf.py value. These fields are produced here:

Fill in the bridge yourself: debt, cash, minority interests, non-operating assets, share count, and the value of employee options. missing_from_payload lists what is still outstanding. Value the options in option-valuation-toolkit and pass the result in; inflating the share count instead understates their cost.

Then run the DCF:

python3 ${HERMES_SKILL_DIR}/../dcf-valuation-engine/scripts/dcf.py value --in payload.json

Read forecast[].roic from that output. If the imputed return on capital drifts to an absurd level, the margin, the sales-to-capital ratio and the growth assumption contradict each other.

Two things not to do. Do not use a regression beta for a young stock — Amazon's had an R-squared of 0.17 and a standard error of 0.50. And do not hold the cost of capital constant for ten years when the whole story is that the firm matures.

6. The private-to-public step

An offering changes who owns the business. That moves the discount rate, the illiquidity discount, the cash in the firm and the share count, all at once. ipo runs the changes as a ladder and prices each one separately, so the reader can see where the value went. Every row of bridge carries the equity value and the value per share before and after one adjustment.

python3 ${HERMES_SKILL_DIR}/scripts/special.py ipo --example | python3 ${HERMES_SKILL_DIR}/scripts/special.py ipo

The discount rate. Give a cost_of_equity block — the same one branch 3 takes, plus pre_tax_cost_of_debt — and the script builds both costs of capital from the one bottom-up beta: the private owner's off the total beta, the market's off the market beta. With a revaluation block it then values the business as a growing perpetuity at each rate. For the restaurant that is 13.25% against 8.76%, and equity of 521 against 1,484 on identical cash flows. If your DCF already ran on a market beta, pass value_of_operating_assets instead and the beta step drops out.

Use of proceeds. Read the prospectus, then classify.

The three must sum to gross_proceeds, or use "use": "retained" for a single bucket. The debt-repayment case is only half-done here: the output warns that the cost of capital has to be recomputed at the new debt ratio and the operating assets revalued.

Prior claims and the share count. Options and warrants come out of the numerator as a value and stay out of the denominator. Everything that becomes common goes into the denominator: common_shares, restricted_stock_units, shares_owed_under_acquisitions, new_primary_shares, and the as-converted shares of any preferred that converts.

Each entry in convertible_preferred carries an as_converted_shares count and a liquidation_preference, and takes whichever is worth more. Leave converts at its "auto" default and the engine iterates to the fixed point, because each round's choice moves the per-share value the others are choosing against. Force it with true or false when the terms do. A round that takes its preference in cash keeps its shares out of the count.

The offering discount. offering_discount is a pricing decision, not a change in value: the bank guarantees the offer price and carries the placement risk, so it prices below the number the model produced. Nothing is applied unless you state it. Average first-day returns run 10–15% and are largest for the smallest deals. The separate underpricing block sizes what that costs the owner — and the loss falls only on the stake actually sold, so a 10% float loses a tenth of what a full float loses.

Reference data

scripts/data/distress_reference.json holds the rating-to-default mapping, the startup survival table and long-run sector survival, each with an as_of date. scripts/data/illiquidity_reference.json holds the pre-computed Silber discount table and the restricted-stock and pre-IPO study results.

Default rates and survival statistics are re-estimated annually, and the bid-ask coefficients date from the end of 2000. Check the vintage. Pass a refreshed file with reference_path rather than editing the bundled copy, and record which vintage the valuation used.

Related engines

Call these rather than rebuilding their arithmetic here.

Common failures

Verification

python3 ${HERMES_SKILL_DIR}/scripts/special.py selftest