Full SKILL.md
Value Driver Modeling
Builds a value-driver tree, reconciled baseline, tornado ranking of drivers, scenarios and a value-at-stake number using economic profit, ROIC and the key value driver formula (Koller, Goedhart and Wessels), with DuPont at the ROIC node. Use whenever a recommendation rests on a financial number whose operational drivers are unstated, or the user says "where is the value", "size the prize", "which levers matter", "why did margin fall", "sensitivity on our plan", "build me a model", "value at stake", "is this worth doing". Not for tracking value once a plan is agreed (kpi-architecture), industry profit pools (market-competitive-intelligence), or a product business case (product-strategy, which may call this for the model).
Connects a root financial metric to the levers a manager controls, sizes each lever with sourced ranges, and states value at stake as a range. The method and its sources are in references/frameworks.md; read it only when a step needs defending.
When not to use
- Value question is really about where profit sits in the industry:
market-competitive-intelligencefirst, then return here. - Plan is agreed and needs a scorecard and benefits ledger:
kpi-architecture(this skill supplies its tree). - Product proposition economics inside a product strategy:
product-strategyowns the answer and calls this skill for the model.
Inputs
Ask for, or infer and label "assumed": the decision the number serves and who makes it (sets root metric and required accuracy); three years of P&L and balance sheet or nearest equivalent (ask separately for receivables, inventory, payables, fixed assets if only a P&L exists); volume and price one level below revenue, by segment if segments differ; cost of capital or enough to estimate it (default 8–10% for mature mid-market, flagged); any prior model or plan to reconcile against; horizon and terminal growth.
Ask one round of questions if the decision or the accounts are missing. Otherwise proceed on numbered, labelled assumptions.
Workflow
-
Root metric and horizon. Start from the decision. Owner or sponsor: economic profit, capitalised to value. BU head with capital allocated centrally: EBITA, with working capital as a check branch. Single initiative: change in economic profit or NPV. Turnaround: operating cash flow. Margin diagnosis: EBITA margin bridged between two periods. Done when the root is written with its definition and one sentence says why not EBITDA or revenue. Test: if management maximised only this metric, would the owner be happy?
-
Reconstruct the baseline. Latest full year. Rebuild NOPAT from EBITA and a stated tax rate; invested capital as operating working capital plus net fixed assets plus operating intangibles, excluding excess cash. ROIC on average capital where possible (say which). Normalise: strip one-offs, annualise partial-year items, restate leases/capitalised development/rebates to where the tree needs them, note policies that flatter year-end. Reconcile to reported EBITDA less D&A with every adjustment named (
references/templates.md§2). Done when the CFO could trace every number to their accounts in five minutes. -
Build the tree, identities first. Work down from the root using identities only until each leaf is a metric in an operating report: economic profit → NOPAT and capital charge → EBITA and tax → gross profit, opex, D&A → revenue × margin → volume × price, split how the business is actually run (accounts × orders × order value; units = capacity × utilisation × yield; heads × utilisation × rate; customers × ARPU; sites × transactions × basket). Working capital as days on the right base: receivables on revenue, inventory and payables on COGS. DuPont at ROIC: NOPAT margin × capital turnover. Only then hang causal links below identity leaves ("gross margin driven by discount discipline, mix, rebates"), drawn differently and each recorded as a hypothesis with a test. Four or five levels; a leaf no manager owns is decoration. Done when every node is labelled identity or causal, identities reconcile to the baseline, and changing any leaf by 10% traces to the root.
-
Envelope before model. For the five to eight most plausible levers, a two-line estimate: value of a permanent change = Δ annual economic profit / (WACC − g); one-off capital releases counted once. Rank. If the ranking is clear and the decision does not need precision, stop and write up. Build a model only if levers interact, ranges are wide, or the number will be scrutinised. Done when the envelope table exists and a written decision on whether to model. Test: which single number, if wrong by 2×, changes the answer? That gets the first hour of diligence.
-
Model hygiene (spec in
references/templates.md§5–6; hand toxlsxfor the file). Assumptions sheet (one input per row: value, unit, source, date, owner, low, high, range source); calculation sheet as the tree, formulas only, no typed numbers except labelled constants; outputs sheet; a checks cell reading OK or listing failures. Units in a column, percentages as decimals, one sign convention, scenarios as switches not tabs, inputs blue / formulas black / links green. Done when a random calculation cell has no typed number and the checks cell reads OK. -
Tornado. Candidate inputs: assumption rows a manager or the world could move. Low and high for each, with a source (3–5 year history, peer benchmarks, dispersion across the client's own branches or segments, or a labelled judgement), asymmetric where evidence is asymmetric, oriented so high is favourable. One input at a time, others at base; swing = |root at high − root at low|; sort descending; draw bars around the base line. Bracket check: base lies between low and high for every input, else a sign error. Report swing in the root and in value. One-paragraph reading. Note what one-at-a-time hides: interactions and correlated inputs. Done when the sorted table has a source on every range and the chart reads aloud in one sentence.
-
Scenarios and value at stake. Three to five switch sets: base carried forward, one or two coherent management cases, a combined plan, a sceptical downside. Report revenue, EBITA, NOPAT, invested capital, ROIC, economic profit, value for each. Value = NOPAT × (1 − g/ROIC) / (WACC − g), g and WACC stated; add one-off capital release once. Value at stake = combined plan minus base, with downside-to-plan range beside it, never a point. Done when the table exists, the value-at-stake sentence has a range, and the three assumptions carrying the plan are named. Test: if the downside happens, is the recommendation still right? If not, it needs a hedge or a stage gate.
-
Drivers that matter. Keep drivers that are large (top of tornado), controllable (a named manager can move them in the horizon) and evidenced (range from data). Usually two or three. For each: driver, current level, target, evidence for the target, value, owner. Then the headline: worth today, could be worth, what has to be true. Done when a CEO could act on it tomorrow without asking what it means.
Output
Lead with the headline conclusion in the first 150 words. Then, each once, 1,500 to 3,500 words including tables:
- Root-metric statement (one line) and baseline page with reconciliation table
- Driver tree (one page, nodes labelled identity or causal)
- Envelope table
- Tornado table (rank, input, base, low, high with sources, root at each, swing) and chart, with one-paragraph reading
- Scenario table and the value-at-stake sentence with range and the assumptions that carry it
- Drivers-that-matter note (half page)
- Model specification when a file is required (
references/templates.md§5)
Assumptions and data gaps follow the conclusion they affect; never open with them. No method vocabulary, checkpoint labels, skill names, framework authors or the worked example's client in the deliverable. Full example with all numbers, tornado and five scenarios: references/worked-example.md. Working notes only on request or when they change a decision, under one heading after the deliverable.
Quality bar
Run silently; fix, do not report.
- Recompute the root from the leaves by hand; it reproduces the baseline to the last decimal.
- Three random baseline numbers trace to the accounts or a numbered assumption.
- The tree has an invested-capital branch with a cost attached, or the first line says why capital is out of scope.
- Every tornado range has a source; "assumed" ranges are flagged in the reading.
- Bracket: root at low < base < root at high for every input.
- Model headline agrees with the envelope within 1.5×, or the difference is explained.
- No typed numbers in five random calculation cells; checks cell OK.
- Each driver that matters has an owner and an evidenced target.
- Value at stake is a range.
- Baseline reconciles to reported EBITDA less D&A with every adjustment listed.
- Headline in the first 150 words; no leaked method vocabulary.
Anti-patterns
- EBITDA tree with no capital: add invested capital and a charge, or state why capital is out of scope.
- Symmetric ±10% on every input: source every range; asymmetric where the evidence is.
- Causal leaves as siblings of identity nodes ("customer satisfaction" under revenue): label and hang causal links below identity leaves.
- Growth as pure upside: show the capital it consumes; below WACC, growth destroys value.
- Tornado read as a to-do list: filter by controllability and evidence; the top bar may be the market.
- Working-capital days on the wrong base.
- One-off cash release counted as annual, or not at all.
- Baseline from the budget rather than actuals.
- Model built before the envelope.