Project Investment Analysis

Project and acquisition investment analysis

A project decision is one arithmetic question wrapped in two judgments. The judgments are what belongs in the cash flow stream and what rate discounts it. The arithmetic is here.

The engine will not decide what is incremental for you, and it will not hide an awkward answer. A stream with two internal rates of return comes back with both roots and a note saying the IRR rule does not apply, rather than one root dressed up as the return.

When to Use

• When deciding whether to take a project, ranking competing projects, or comparing projects with different lives.
• When sizing a capital budget under rationing, or testing whether a cash flow stream is truly incremental.
• When computing NPV, IRR with every root reported, MIRR, payback, project return on capital or EVA.
• When valuing synergy or control, or setting the maximum price for an acquisition with the four-number acid test.
• When pricing a majority or minority stake: control premiums and minority discounts.

How to Run

scripts/project.py — pure standard library, no installation needed. JSON on stdin (or --in FILE), JSON out.

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

Cash flow convention throughout: cash_flows[0] is the year-0 outlay and is not discounted; cash_flows[t] lands at the end of year t.

Build the stream before you run anything

Every number below is only as good as the stream feeding it. Work through this first.

Incremental cash flow checklist:
- [ ] Write the counterfactual: what the firm's cash flows look like without the project
- [ ] Exclude sunk costs, and the depreciation tax shield on any capitalized sunk asset
- [ ] Exclude allocated fixed overhead; include genuinely new overhead the project causes
- [ ] Charge the opportunity cost of every resource the firm already owns
- [ ] Charge cannibalization of existing products, at the share that would have been kept
- [ ] Credit side benefits, valued at the receiving business's rate
- [ ] Close the stream with salvage, working-capital recovery, or a terminal value

Sunk costs are out. Money already spent and unrecoverable does not change with the decision. So does its depreciation tax shield: that shield exists whether or not the project goes ahead, so leaving it in quietly keeps part of the sunk cost in the analysis. The behavioural research says managers find this nearly impossible to do. The rule is easy; the discipline is not.

Allocated fixed overhead is out; incremental overhead is in. A share of a central G&A pool charged to the project on a sales basis is not caused by the project. New headcount and new systems are. Split the pool empirically by regressing company G&A on company revenues: the slope is the variable rate you charge, the intercept is the fixed pool you do not. Treating the whole allocation as non-incremental is the opposite error and just as wrong.

Opportunity cost is in, priced at the best alternative use. Never at zero because the firm already owns the asset, and never at book value.

Excess capacity is the case that looks free and is not. "Already paid for and nobody else wants it" is a sunk-cost argument in a different hat. Ask when capacity runs out without the project, when it runs out with it, and what the firm does then. If the answer is "build earlier", charge the present value of building earlier less the present value of building later. If the answer is "cut production", charge the after-tax cash flow on the lost sales.

Cannibalization is in, at the share that would have been kept. For an exclusive differentiated product, diverted sales were not going anywhere else, so count all of them. In a fiercely competitive market those customers might have left for a rival regardless, so count only the share you would truly have retained.

Side benefits are in, at the receiving business's rate. See "Synergy" below.

The incremental subcommand runs the adjustment route: start from cash flows as the accounting system reports them and back out what the firm would have spent anyway.

python3 ${HERMES_SKILL_DIR}/scripts/project.py incremental --example | python3 ${HERMES_SKILL_DIR}/scripts/project.py incremental

Building the stream directly from incremental revenue and cost lines is the other route. The two must reconcile up to rounding. Do not mix them inside one model, or the adjustments get counted twice.

Choosing the rate

Match the rate to the cash flows on four dimensions: claimholder, business, geography and currency. Cash flows to the firm discount at the cost of capital; cash flows to equity discount at the cost of equity. A project in a different business from the parent takes that business's risk, not the parent's. Build the rate with cost-of-capital-toolkit (wacc subcommand, read cost_of_capital) and pass the number in here.

Never blend rates across streams that carry different risk. The npv subcommand takes a streams list precisely so a project and its synergy can each carry their own rate and still produce one total.

NPV and IRR

echo '{"cash_flows": [-2000, -1000, -859, -267, 340, 466, 516, 555, 615, 681, 11990],
       "discount_rate": 0.0846, "profile_rates": [0.08, 0.12, 0.16, 0.20]}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/project.py npv

NPV is the expected increase in firm value from taking the project. Report it as a dollar statement, not a sign. A positive NPV that is thin relative to the investment is a thin recommendation: the Netflix Fit case clears its hurdle by $106 million on a $2.4 billion outlay, and small assumption changes flip it.

irr solves for the rate that drives NPV to zero, by scanning the whole plausible rate domain for sign changes in NPV and bisecting each bracket it finds. Three outcomes:

• One root, one sign change. irr is the answer and reliable is true.
• More than one root. irr is null, roots lists them all, and the note tells you to decide on NPV at the actual cost of capital. This happens when the stream changes sign more than once — decommissioning, cleanup, restoration, or a mid-life capacity investment.
• No root. The stream never crosses zero. Usually the year-0 outlay is missing or has the wrong sign.

Read sign_changes even when a single root came back. A stream can change sign three times and still have one root in the searched range; the rate is then descriptive rather than a decision rule, and the engine marks it reliable: false.

Set hurdle_rate and the output carries a decision, falling back to NPV automatically when the IRR rule does not apply.

Ranking projects against each other

Independent projects are the exception. When candidates compete, NPV and IRR can disagree for three reasons, and each has its own remedy.

MIRR replaces the IRR's assumption that intermediate cash flows are reinvested at the IRR itself. Outflows come back to today at the financing rate, inflows go forward to the horizon at the reinvestment rate, and the gap between IRR and MIRR is the size of the illusion. Both rates default to the hurdle rate, which is the standard treatment.

Profitability index is NPV / initial investment — value created per dollar of scarce capital. The output also carries pv_index_including_investment, which is the same number plus one, because both conventions are in circulation.

python3 ${HERMES_SKILL_DIR}/scripts/project.py rationing --in projects.json

Use the index only when capital is genuinely rationed. Applying it otherwise biases the firm toward small projects and leaves value on the table. And check whether the constraint is real: surveys put roughly 70% of capital rationing down to borrowing limits set by the firm's own management, which is a policy, not a law of nature.

Give a budget and the engine also reports best_selection, the highest-NPV affordable set found by exhaustive search when there are 16 projects or fewer. Projects are indivisible, so taking them in index order can leave money idle; when it does, the output sets greedy_is_suboptimal and quantifies the gap.

Projects with different lives

Raw NPVs are not comparable across different lives. The longer project accumulates value over more years and ties up capital for longer, and the comparison rewards it for both.

python3 ${HERMES_SKILL_DIR}/scripts/project.py different-lives --example \
  | python3 ${HERMES_SKILL_DIR}/scripts/project.py different-lives

The engine reports both repairs. Equivalent annuity divides each NPV by the annuity factor over that project's own life, giving a level annual value. Replication repeats each stream to the common multiple of the lives and discounts the whole thing; the reinvestment at the start of a new cycle nets against the last receipt of the previous one. The two should rank the projects identically, and routes_agree says whether they did. They can part company when annual cash flows are irregular, in which case trust replication.

Before using either, ask whether repetition is realistic. A one-off licence, a unique site or a first-mover position cannot be repeated on the same terms, and then neither repair applies. IRR needs no life adjustment at all, since it is a rate rather than a total.

Payback

payback reports both the simple and the discounted period, interpolated within the crossing year, and the gap between them as cost_of_time_years. That gap is often large: Rio Disney pays back in 10.3 years undiscounted and 16.8 years discounted.

Payback describes how long capital is at risk. It ignores every cash flow after the payback date, and in its simple form ignores the time value of money entirely. Use it as a supplementary read, never as the decision rule.

Accounting return and EVA

python3 ${HERMES_SKILL_DIR}/scripts/project.py accounting-return --in project-books.json

Feed after-tax operating income by year and beginning-of-year book capital by year, plus the cost of capital. Book capital for a project is undepreciated fixed assets plus the book value of non-cash working capital; for a firm it is book debt plus book equity minus cash, taken from the previous year's balance sheet.

The output carries three averages, because the corpus itself uses two of them and they can differ by several points:

• average_roc_start_of_year_basis — mean of the annual ratios, income over opening capital
• average_roc_average_capital_basis — mean of the annual ratios, income over average capital
• roc_on_average_income_and_capital — average income over average capital

On Netflix Fit those give 14.25% and 10.76% on the same numbers. Name the convention when you quote the figure.

EVA converts the spread into dollars: EVA = (ROC − cost of capital) × capital, which is identically the income left after every dollar of capital is charged its cost. The output also discounts the EVA stream, which should track NPV when the book capital roll-forward is consistent with the cash flows.

Read the level, not the trend. Book returns climb mechanically as the asset base depreciates even when nothing improves — Netflix Fit's ROIC runs from −2.91% in year 1 to 44.06% in year 10 on an unchanging asset. And averaging over a truncated window on a long-lived project loads the early losses in and leaves the mature years out. Rio Disney's 4.18% average ROC over ten years says reject; the perpetual-life DCF says +$3.3 billion.

Synergy

A project can make the firm's other businesses more valuable. That benefit is real and it is routinely abused, invoked qualitatively late in the process to override a negative NPV. The discipline is to value it explicitly, in the original analysis, at the cost of capital of the business that receives it.

Two routes, both in the synergy subcommand.

Cash flow route — for a project side benefit or a synergy you can schedule:

echo '{"mode": "cash_flows", "receiving_business": "Netflix Entertainment",
       "cash_flows": [56.25, 56.81, 57.38, 57.95, 58.53, 59.12, 59.71, 60.31, 60.91, 61.52],
       "discount_rate": 0.0893}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/project.py synergy

cash_flows starts at year 1. Put zeros in front for the lag before the benefit arrives — synergies that "start immediately" are almost always wrong. Add a terminal block with cash_flow_next_year and growth_rate for a synergy that continues in perpetuity, and an exchange_rate when the synergy is earned in another currency.

Combined-firm route — the definition, for a merger:

synergy = value of the combined firm with synergy − (acquirer alone + target alone)

Pass combined_value_without_synergy as well and the engine runs the sum-of-parts check. That value must equal the sum of the two stand-alone values exactly, because merely adding two firms together creates nothing. A difference means an inconsistent assumption reached the combined-firm model. Use the target's restructured value in the baseline when the deal also claims control value, or the restructuring gains get counted twice.

Force every claimed synergy onto a valuation input before you value it. Higher returns on new investment or more new investment raise growth. Cost savings raise the operating margin. A more durable advantage lengthens the growth period. Tax benefits lower the effective rate. Added debt capacity raises the debt ratio. If a claim moves none of these, it is a buzz word. Diversification is not a synergy for a public firm, whose investors diversify more cheaply on their own account.

Report the stand-alone number next to the combined one, always. Netflix Fit is $106 million alone and $483 million with synergy, and the whole recommendation turns on that gap.

The value of control

Control is worth the gap between the firm as it is run and the firm as it could be run. Nothing else. Value the target twice, once under existing management and once under the policies you would set, and subtract. Both valuations are full DCFs and belong in dcf-valuation-engine. control-value does the arithmetic that follows.

echo '{"status_quo_value": 955, "restructured_value": 2323, "shares_outstanding": 186.3,
       "probability_of_change": 0.595, "market_price_per_share": 9.50}' \
  | python3 ${HERMES_SKILL_DIR}/scripts/project.py control-value

A gap is only worth having if someone can close it, so the second input is the odds that management actually changes.

expected value of control = probability of change × (restructured − status quo)

That one product explains four things at once. It sets what a hostile bidder can pay. It puts a market price between the two values. It makes a voting share worth more than a non-voting one. And it is the part a minority stake does not get.

A buyer who takes control can force the change, so its probability is 1 and its ceiling is the whole restructured value. Blockbuster in 2005 was worth $5.13 a share as run and $12.47 a share run well. At the $9.50 market price the most a bidder could justify was a $2.97 premium, and paying all of it would have handed the entire improvement plan to the seller.

Pass market_price_per_share and the engine inverts the identity to show the odds the market is already paying for. Blockbuster traded at $8.20 before Carl Icahn's challenge and $9.50 after, which reads as 41.8% and 59.5%. That 18-point move is the value of activism, measured directly. Read a figure above 100% as a broken input rather than as near certainty; the restructured value is usually the input at fault.

Two adjustments matter. Changes that take years arrive late, so pass implementation_delay_years with a discount_rate and quote adjusted_value_of_control. And where a firm has voting and non-voting shares, pass a share_classes block. Every share owns the same cash flows, so the status quo spreads across all of them and only the expected control value attaches to the voting class. Embraer's voting shares carry a 10.4% premium at 20% odds and a 26% premium at 50%.

Control premiums and minority discounts

control-premium turns the same gap into the two ratios deal practice quotes, and converts between them.

control premium   = (optimal − status quo) / status quo
minority discount = (optimal − status quo) / optimal
minority discount = control premium / (1 + control premium)

One gap, two denominators. A controlling stake is priced off the optimal value because it can make the changes. A minority stake is priced off the status quo because it cannot. At Kristin Kandy a 51% stake is worth $1.02 million and a 49% stake $784,000. Two points of ownership are worth $236,000, and all of it is control.

The flat 20% control premium in circulation is not analysis. It is a survey average of other people's deals, and averages of overpayments are still overpayments. A well-run target carries a premium of zero however highly it is regarded. If you cannot name the change, price it and defend it, there is nothing to pay for.

Acquisitions are projects

Every rule above applies to a deal. The two mechanical ways to overpay are discounting the target's cash flows at the acquirer's rate, and building the acquirer's cheap debt into the target's cost of capital. A risky business does not become safe because a safe buyer owns it. If lowering the discount rate is what makes the deal work, the deal does not work.

maximum price = target's stand-alone value + value of synergy

Pass an acquisition block to synergy and the engine computes that ceiling, the premium over stand-alone value, and how much synergy is left for the acquirer at a given price.

python3 ${HERMES_SKILL_DIR}/scripts/project.py synergy --in deal.json

The ceiling is a walk-away point, not a target. Paying the full synergy value as a premium hands the entire gain to the seller and leaves the acquirer with exactly nothing, which the output shows as synergy_retained_by_acquirer: 0. Read synergy_required_to_justify_price before bidding: on Tata Motors and Harman the market price demanded about $2.75 billion of synergy and the credible figure was $1.18 billion.

The acid test

There are three value-based reasons to buy a company, and each has its own benchmark. deal takes the four numbers and runs all three at once.

python3 ${HERMES_SKILL_DIR}/scripts/project.py deal --example | python3 ${HERMES_SKILL_DIR}/scripts/project.py deal

The output splits the gains. The seller takes the premium over stand-alone value, the acquirer keeps whatever is left under the ceiling, and premium_share_of_combined_gains says which way the split went. AB InBev paid $104 billion for SABMiller against a ceiling of $70.8 billion, so the premium came to 272% of every gain the deal could create.

deal refuses to run in two cases, both of them ways of reaching a price rather than a value. It needs target_discount_rate_used: true to confirm the target was discounted at its own cost of capital. And when synergy is claimed it needs synergy_baseline: "restructured_target", because baselining synergy on the status quo puts the control gains inside the synergy figure and lets the deal claim them twice.

When all three tests fail, exactly two explanations remain: the synergy is underestimated, or the buyer is overpaying. Pick one and say which.

synergy-haircut applies the post-merger evidence to a synergy schedule. Split the schedule into cost and revenue components first — a single blended number hides that the fragile half is the revenue half. Cost synergies land most of the time; roughly 70% of mergers miss their expected revenue synergies. Revenue components also take a customer attrition haircut for losses during integration.

Before trusting any synergy number, test it for concreteness. Can you name the plant, the contract, the headcount and the month? Is a named person accountable for delivering it? Are you one of several bidders, in which case competition bids the synergy into the price?

Two side notes on deal arithmetic. Goodwill is the acquisition price less the target's adjusted book equity, which makes it the acquirer's own public estimate of the value it must now create rather than an asset. And the base year has to be normalized before anything is projected: on Harman, fixing a one-off working-capital spike alone moves free cash flow from −$95 million to +$167 million.

Working with the other engines

A negative NPV is not always a rejection. A project that buys the right to delay, to expand later, or to abandon carries option value that a static DCF does not price. Value it with option-valuation-toolkit and add it, rather than arguing the cash flows upward.

Reference material

• references/reference.md — the corpus worked examples used as test vectors, the decision-rule usage and capital-rationing surveys, side-cost pricing rules, and the full input reference for each subcommand.
• scripts/data/synergy_realization.json — the McKinsey post-merger realization bands that drive synergy-haircut, tagged with an as_of date and a refresh note. Pass a replacement with table_path rather than editing the bundled copy.

Common failures

Verification

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