Financial Statement Normalization

Financial statement normalization

Accounting statements are not built for valuation. They are built to be auditable and comparable, and those goals push them away from economic substance in three specific ways. Research spending is an investment that the accountant expensed. Leases are debt that sat off the balance sheet for decades and still get reported on rules nobody trusts. One-time items sit inside operating income and get relabelled by management every year.

This skill applies the corrections and then derives what follows from them: cash flows, the capital base, and the return earned on that capital. Every number the script produces is a pure transform of numbers you supply. Which items are non-recurring, and how long research takes to pay off in this business, are your calls.

When to Use

• When cleaning up EBIT before a DCF, or building the base-year operating income and capital base a valuation model consumes.
• When capitalizing research or leases, and rebuilding invested capital and ROIC on the corrected numbers.
• When deciding whether a restructuring charge, impairment or gain is really non-recurring.
• When computing free cash flow to the firm or to equity, or measuring return on invested capital and economic value added.
• When normalizing earnings for a cyclical or commodity company sitting at a peak or a trough.

How to Run

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

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

selftest checks 26 cases, including the algebraic invariant that flat R&D forever leaves operating income untouched, and that spending exactly one amortizable life ago contributes nothing to the research asset. A wrong amortization schedule fails both.

The order of operations

The corrections are not independent. Leases change interest coverage, which changes the cost of debt, which changes the discount rate used on the leases. R&D changes both the numerator and the denominator of ROIC. Run them in this order.

Normalization progress:
- [ ] 1. Update the statements to a trailing twelve months
- [ ] 2. Capitalize R&D and any other multi-year spending
- [ ] 3. Capitalize operating leases, iterating on the cost of debt
- [ ] 4. Strip genuinely non-recurring items
- [ ] 5. Normalize the year if it is a peak or a trough
- [ ] 6. Rebuild invested capital and ROIC on the corrected numbers
- [ ] 7. Derive FCFF and FCFE
- [ ] 8. Read the ratio pack and check the story holds together

1. Update first

An annual report goes stale fast. Build trailing-twelve-month figures from the last 10-K plus the latest 10-Q:

TTM item = last 10-K annual figure − prior-year year-to-date + current-year year-to-date

This matters most for small firms, volatile firms, and firms that recently restructured. The script has no subcommand for this because it is subtraction, not modelling. Do it before anything else, so every correction below lands on current numbers.

2. Capitalizing R&D

Research spending buys benefits over several years, which makes it a capital expenditure in economic substance. Accounting requires it to be expensed. The consequence is that an R&D-heavy firm looks like it earns less than it does and reinvests nothing at all, while its return on capital looks spectacular because the capital it built is invisible.

The fix treats each of the last N years of R&D as an investment being written off straight line over N years. What is left unwritten-off is the research asset.

echo '{"current_rd": 100, "past_rd": [90, 80, 70, 60, 50], "amortizable_life": 5}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/normalize.py capitalize-rd

past_rd runs backwards: index 0 is last year, index 1 the year before. Spending from exactly amortizable_life years ago is fully written off, contributes nothing to the asset, and still contributes its share to this year's amortization. Keep that row in.

Read three fields:

• research_asset — add to invested capital and to book equity.
• amortization_this_year — the charge that replaces this year's expensing.
• adjustment_to_operating_income — current R&D minus amortization. Add to reported EBIT.

The adjustment is positive when R&D is growing and negative when it is shrinking. For a firm whose R&D has been flat for years it is zero, which is the sanity check: the correction only matters when research spending is changing.

Choosing the amortizable life

This is the one genuine judgment in the calculation. Ask how long research takes to pay off in this business, then use the industry table as a default rather than a fact.

The two cases worth naming: a software company amortizes over roughly three years — five if it sells enterprise platforms with long product cycles — because a code base ages fast. A pharmaceutical company amortizes over ten, because a compound spends most of a decade in trials before it earns anything. Using a software life on a drug company shrinks the research asset and inflates ROIC.

If you have fewer than N years of history, pad the missing years with zeros and say so. The research asset is understated by exactly the padding.

The same machinery capitalizes any other spending that buys multi-year benefits: brand advertising, employee training. Run it as a second capitalize-rd call with a different life, and add both assets to capital.

3. Capitalizing operating leases

A lease is a fixed, tax-deductible commitment that must be paid whatever operations do, and non-payment costs the firm the asset. That is the definition of debt. Treating the payment as an operating expense understates both debt and operating income, and the effect is enormous in lease-heavy sectors. Lease expense runs about 12.5% of operating income for the whole market, about 27% for restaurants, about 44% for apparel stores, and about 50% for furniture stores.

Discount the disclosed commitments at the pre-tax cost of debt — not the cost of capital, not the riskfree rate.

echo '{"commitments": [100, 100, 100, 100, 100], "lump_sum_beyond": 300,
       "pre_tax_cost_of_debt": 0.05, "current_lease_expense": 110}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/normalize.py capitalize-leases

commitments is the itemized schedule from the footnotes, usually five years. lump_sum_beyond is the single "thereafter" figure. It is not one payment. The script infers how many years it covers by dividing it by the average of the itemized years and rounding, then discounts it as an annuity starting after the itemized years. Here 300 against an average of 100 implies three more years, so the lease life is eight.

Read four fields:

• lease_debt — add to debt, to invested capital, and to debt in the equity bridge.
• depreciation_on_lease_asset — lease debt divided by the full lease life.
• adjustment_to_operating_income — lease expense minus that depreciation. Add to EBIT.
• imputed_lease_interest — lease debt times the pre-tax cost of debt. Add to interest expense wherever interest coverage is computed.

The asset is created at exactly the same value as the debt. Adding the debt and forgetting the asset understates capital and inflates every return you compute afterwards.

Check the result: net income must be unchanged. The higher EBIT is exactly offset by the higher imputed interest. If net income moves, the lease payment has been counted twice.

The circularity, and how to iterate it

Discounting leases needs a pre-tax cost of debt. When the company has no traded bonds and no agency rating, that rate comes from a synthetic rating built on interest coverage. But capitalizing leases changes interest coverage, in both the numerator and the denominator. The rate depends on the answer that depends on the rate.

Iterate to a fixed point. Two to four passes is normal.

# Pass 1 — first-guess rate from reported coverage
echo '{"ebit": 1012, "interest_expense": 120, "riskfree_rate": 0.04,
       "marginal_tax_rate": 0.35, "table": "large_manufacturing"}' \
  | python3 ${HERMES_SKILL_DIR}/../cost-of-capital-toolkit/scripts/costofcapital.py rating
# read pre_tax_cost_of_debt

# Pass 2 — capitalize at that rate
echo '{"commitments": [899, 846, 738, 598, 477], "lump_sum_beyond": 1425,
       "pre_tax_cost_of_debt": 0.055, "current_lease_expense": 978}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/normalize.py capitalize-leases
# read lease_debt, imputed_lease_interest, adjustment_to_operating_income

# Pass 3 — re-rate on lease-adjusted coverage, then go back to pass 2
#   ebit             = reported EBIT + adjustment_to_operating_income
#   interest_expense = reported interest + imputed_lease_interest

Stop when the pre-tax cost of debt moves by less than a basis point between passes. Record the final rate, because the rest of the valuation has to use the same one.

Two notes on the modern statements. Since 2019 both IFRS and US GAAP require operating leases on the balance sheet, which validates this adjustment. The official rules are far more complicated than a present value, because standard-setters preserved continuity with legacy treatment. When the reported lease liability looks inconsistent with the commitment schedule, recompute it here and use your own number.

4. Stripping one-time items

A charge is non-recurring only if it is both infrequent and variable in amount. An item that appears every year is recurring whatever the firm calls it, even when it flips between gains and losses — a sign flip is variability, not infrequency.

Pull five to ten years of statements and count. For each item type, how many of those years does it appear in?

The middle case is the one that matters most. A firm reports a $500m loss containing a $1bn charge it calls one-time. If the charge is genuinely once-ever, value the firm off $500m of profit. If the firm books a charge like this roughly every five years, it is a recurring cost: build in $1,000m / 5 = $200m every year, giving $300m of sustainable operating profit. Neither headline number is right.

Pass accepted adjustments to full as one_time_items. Sign them from the point of view of operating income: a charge you are adding back is positive, a non-recurring gain you are removing is negative.

"one_time_items": [
  {"description": "2023 plant fire, insurance shortfall", "ebit_effect": 140.0},
  {"description": "gain on sale of the Brazilian unit", "ebit_effect": -85.0}
]

These carry no capital effect. Do not accept management's label, and do not accept an "adjusted EBITDA" figure that has not been reconciled line by line to the audited statement.

5. Normalizing a peak or a trough

Diagnose the cause before you normalize anything. There are five, and they split into two treatments.

Normalizing a firm whose problems are structural values a company that no longer exists. For the three forecast cases, use dcf-valuation-engine with a revenue-driven build and a target margin, rather than this subcommand.

echo '{"method": "average_margin", "historical_margins": [0.12, 0.08, -0.02, 0.05, 0.11],
       "current_revenue": 10000, "reported_ebit": -200}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/normalize.py normalize-earnings

Three methods, chosen by what has changed about the firm:

• average_margin — average the operating margin over the cycle and apply it to current revenue. Use when the firm has grown or shrunk, and you want to keep its current scale.
• average_roc — average the return on capital and apply it to current invested capital. Use when you want to keep its current asset base, and when the margin is unstable because the asset mix changed.
• average_earnings — average the dollar earnings. Only defensible when firm size has barely moved across the window. The script returns a caveat saying so.

Choose the window to span a full cycle. Five years is the working default; commodity cycles run longer. Do not choose a window that starts at the trough and ends at the peak. When the firm was hit by a specific shock, use the pre-shock cycle instead of straddling it.

Then carry the fix through. Normalized EBIT changes interest coverage, which changes the synthetic rating and the cost of debt, which changes the discount rate. It also changes the reinvestment rate and the return on capital that drives growth. Be consistent about which numbers are normalized: a normalized EBIT paired with a depressed current coverage ratio is a broken model.

A normalized, mature firm may deserve no high-growth period at all.

6. Invested capital and ROIC

echo '{"book_value_of_equity": 5000, "book_value_of_debt": 3000, "cash": 1000,
       "lease_debt": 500, "research_asset": 250,
       "ebit_after_tax": 750, "cost_of_capital": 0.08}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/normalize.py invested-capital

invested capital = book equity + book debt − cash + lease debt + research asset

Cash comes out because it earns a financial return, not an operating one. The numerator excludes interest income, so the denominator must exclude the cash that produced it. Leaving cash in understates the return on the operating business.

Measure the capital at the start of the period, so that the income in the numerator was actually earned on it. Using end-of-year capital mechanically depresses the ratio for a growing firm.

The output carries roic, return_spread against the cost of capital, and economic_value_added, which is the spread times the capital. A positive spread means growth creates value; a negative one means growth destroys it, and growing faster makes it worse.

Before trusting the number, run the six-distortion check. ROIC is built from two badly measured numbers.

Compare the result to the firm's own history and to the industry. A ROIC far above the industry that you cannot explain by a durable advantage is a signal to fade it, not to extrapolate it. Empirically, excess returns persist longer than growth does — so fade growth faster than you fade the spread.

The rule that breaks most models

The EBIT adjustment and the capital adjustment always move together. Every correction in this skill has two legs:

Applying the earnings leg alone raises the numerator of ROIC and leaves the denominator untouched. The return looks better than it is, and because fundamental growth is reinvestment rate × ROIC, that error propagates straight into an inflated growth rate and an inflated valuation. This is the single most common porting error when these adjustments are done by hand in a spreadsheet.

The correction usually cuts ROIC rather than raising it, because capital rises proportionally more than earnings. Capitalizing R&D at SAP moved return on capital from 29.9% to 19.2%. Capitalizing leases at The Gap moved it from 12.9% to 9.3%. If your corrected ROIC went up, check that both legs landed.

The full subcommand threads both legs automatically, which is the main reason to use it.

7. Cash flows

echo '{"ebit": 1000, "tax_rate": 0.25, "capital_expenditures": 300, "depreciation": 200,
       "change_in_noncash_working_capital": 50, "net_income": 600, "debt_ratio": 0.3}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/normalize.py cashflow

net capital expenditures = capex + acquisitions + capitalized R&D − depreciation
reinvestment             = net capital expenditures + change in non-cash working capital
FCFF                     = EBIT × (1 − tax rate) − reinvestment

ebit here means the corrected EBIT, not the reported one. Losses are not taxed: when EBIT is negative the script passes it through untouched rather than manufacturing a tax credit.

Two things belong in reinvestment that the cash flow statement hides. Acquisitions sit in the investing section, usually under "other investing activities" — pass a multi-year average as acquisitions, because most firms do not acquire every year. Capitalized R&D is not in the statement at all — pass it as rd_capitalized_this_year.

Use the non-cash working capital definition: non-cash current assets minus non-debt current liabilities. Leaving cash and short-term debt in makes working-capital "investment" swing with the cash balance and the debt maturity schedule, neither of which is operating. Forecast it as a percent of revenue rather than extrapolating last year's dollar change, which is noise.

FCFE arrives by whichever route the payload supports. Pass debt_ratio for the stable-leverage shortcut, in which debt funds a constant fraction of every dollar of reinvestment. Pass new_debt_issued and debt_repaid to model the debt flows explicitly. Use the shortcut only when leverage really is stable; when it is changing, value the firm with FCFF and skip FCFE entirely.

Never put interest expense, interest tax savings, or debt flows into FCFF. The tax shield already lives in the (1 − t) on the cost of debt inside the WACC.

Keeping FCFF invariant to the R&D correction

Capitalizing R&D should leave FCFF unchanged. After-tax operating income and net capital expenditure both rise by the same amount, so the cash flow is untouched — only earnings, capital, ROIC and therefore growth change. Two payload conventions make that identity hold, and the script will not apply either for you.

Add the amortization to depreciation. Adjusted depreciation is reported D&A plus amortization_this_year from capitalize-rd. Otherwise net capital expenditure is overstated by the amortization.

Do not tax the add-back. The firm already deducted the full R&D expense, so the cash taxes it paid were computed on reported operating income. Supply the effective rate on the corrected EBIT:

tax rate to pass = marginal rate × reported EBIT / adjusted EBIT

Worked check, with reported EBIT 1,000, current R&D 100, amortization 60, marginal rate 25%, capex 300, reported depreciation 200 and a working capital change of 50:

adjusted EBIT        = 1,040
tax rate to pass     = 0.25 × 1,000/1,040 = 0.240385
after-tax operating income = 1,040 × (1 − 0.240385) = 790   ( = 750 + 40 )
depreciation to pass = 200 + 60 = 260
net capital expenditures   = 300 + 100 − 260 = 140
FCFF = 790 − 140 − 50 = 600                                  ( = unadjusted FCFF )

Both conventions are payload-level, so they work inside full too: set cash_flow.depreciation and cash_flow.tax_rate explicitly, and full will respect them. Run capitalize-rd first to read the amortization, then build the payload. Left at the defaults, full taxes the add-back and omits the amortization, and the bundled example reports FCFF of 530 where the invariant says 600.

The lease correction behaves differently and does need the tax. Its imputed interest deduction moves into the WACC, so the corrected EBIT is taxed in full. Leave net capital expenditure alone for leases — new leases signed each year replace the depreciation — and remember that the higher cash flow is paid for by the lease_debt you add to debt in the equity bridge.

8. The ratio pack

python3 ${HERMES_SKILL_DIR}/scripts/normalize.py ratios --in statements.json

Every field is optional. Anything the payload does not support comes back as null rather than a crash, so this is useful part-way through an analysis. Feed it corrected numbers, not reported ones — an interest coverage ratio that ignores imputed lease interest will buy a synthetic rating the company does not deserve.

The three readings that carry the most information: sales_to_capital, because it is the input a driver-based DCF needs to convert growth into reinvestment; interest_coverage, because it drives the synthetic rating; and return_spread, because it decides whether growth is worth anything.

Running the whole chain

python3 ${HERMES_SKILL_DIR}/scripts/normalize.py full --example
python3 ${HERMES_SKILL_DIR}/scripts/normalize.py full --in company.json

full applies the corrections in dependency order, threads both legs of each into EBIT and into capital, and returns an adjustments array showing what each correction did to each. Read that array first: it is the audit trail, and a reviewer will ask for it.

Handing the results onward

Run valuation-consistency-checks over the finished artifacts before accepting any of it.

Common failures

Verification

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