Demand and Benefit Forecasts by the Numbers
Chapter 10 of 10 · 11 min read
The post-completion review and the next decision
Re-appraisal at month 21 gives an NPV of −A$2,479.2k, the 16.950m run-rate sits at P70.4 of the forecast range, the records are updated, and a A$3.9m second line is tested.
The board’s second condition was a post-completion review at about 21 months. In October 2026 the independent reviewer who had tested the forecast in February 2024 led it, with the finance manager supplying the figures. By then the review had a second purpose that nobody had foreseen at approval: the managing director had proposed a second canning line, on the strength of the customer whose growth Chapter 9 examined.
This chapter follows the review through its parts: re-appraising the plant on its actual results, scoring the forecast range against the outcome, adding the plant’s experience to the business’s records, and using all of that to test the next decision. It ends with the error in the second-line proposal and the conditions the review put in its place.
What the review is for
A post-completion review asks four questions: what happened, how it compares with the business case, why it differed, and what the business should do differently next time. Infrastructure Australia’s Assessment Framework includes a formal version, Stage 4, for reviewing a completed project against the benefits in its business case. The canning plant’s review is a much smaller private-sector version of the same idea.
The most important distinction in any such review is between the quality of the decision and the quality of the outcome. A sound decision can turn out badly through bad luck, and a poor one can turn out well through good luck. A review that judges only outcomes teaches the business to imitate its lucky decisions. Reviewing the decisions, not the outcomes in Stage Gates and Stopping Rules by the Numbers treats this question in detail. Here it shapes the order of the review: first the outcome, then the forecast, then the records, and only then the next decision.
Re-appraisal at month 21
This is an illustration. All money is A$ excluding GST. The reviewer rebuilt the appraisal with actual figures where they existed and the month-21 run-rate thereafter, keeping the business case’s structure: capital in 2024, ten operating years from 2025 to 2034, discounted at 9%, no residual value, and the A$57,600 a year of mobile contribution lost as a dis-benefit.
- 2024: capital of A$6.10m, against the budget of A$5.4m: −A$6,100.0k.
- 2025: contribution A$882.8k, less fixed costs of A$1,300.0k, the dis-benefit of A$57.6k and the A$185,000 of 250 mL change parts: −A$659.8k.
- 2026: contribution A$1,964.8k, nine months actual and three reforecast, less fixed costs and the dis-benefit, plus the A$255,000 from the sale of rig 2: A$862.2k.
- 2027 to 2034: the month-21 run-rate contribution of A$2,109.0k a year, less fixed costs and the dis-benefit, with no growth: A$751.4k a year.
The NPV on that trajectory is −A$2,479.2k, and the IRR is 0.3%. Against the business case’s A$6,081.5k, the plant has lost about A$8.6m of value on paper. The comparison that matters more is with the evidence available at the decision: at the simulated P50, the outside view gave −A$2,852.5k. The plant has done slightly better than the record predicted, despite a capital overrun of 13.0%, and the reason is one customer.
The last line puts a value on the concentration risk from Chapter 9. H, the functional drinks brand, accounts for A$2,683.3k of the plant’s present value. If the brand leaves at the end of 2026, the plant becomes worse than the outside view’s central case.
The re-appraisal also shows what would make the existing plant pay for itself. Holding everything to 2026 as it is, the plant needs a steady 21.227m cans a year from 2027, at the run-rate’s A$0.124 a can, for its NPV to reach zero: 125.2% of the month-21 run-rate.
Scoring the forecast range
The forecast that went to the board in February 2024 was a range, and a range can be scored. The month-21 run-rate of 16.950m cans sits at the 70.4th percentile of the 20,000 independent simulated runs from Chapter 6, inside the P10 to P90 range of 11.57m to 19.20m. The business case’s 25.32m had a 0.1% chance on the same evidence. On this single outcome, the driver-level forecast did its job and the inside view did not.
The segment detail tempers that verdict.
| Business case, m cans | Outside view, m cans | Run-rate at month 21, m cans | Run-rate ÷ outside view | |
|---|---|---|---|---|
| A Craft brewers | 3.240 | 1.735 | 1.350 | 0.778 |
| B RTD and seltzer | 12.000 | 7.840 | 5.760 | 0.735 |
| C Non-alcoholic | 7.200 | 4.160 | 8.640 | 2.077 |
| D Overflow | 2.880 | 1.632 | 1.200 | 0.735 |
| Total | 25.320 | 15.367 | 16.950 | 1.103 |
The outside view landed close to the total partly through offsetting errors. Three segments ran at 0.735 to 0.778 of their outside-view expectation; segment C, almost entirely because of H, ran at 2.077. Without H the plant would be running at 12.150m cans, 0.791 of the outside view and at about the 14th percentile of the range. Even the corrected forecast was optimistic about the customers it described, and was rescued in total by a customer it could not have described.
One outcome cannot calibrate a forecasting method. An 80% interval should contain the outcome about eight times in ten, and that can be checked only across many forecasts. The review’s job is to add this forecast and its outcome to the business’s forecast register, with the date, the method and the percentile at which the outcome fell, so that in a few years the business can see whether its intervals contain the outcomes as often as they claim. Deciding, tracking and calibrating the forecast in Reference Class Forecasting by the Numbers shows how such a register is kept and back-tested. Tetlock and Gardner’s advice in Superforecasting is the same in spirit: keep score, and update the method in small steps as the scores arrive.
The review scored the forecast against the version the decision used, the February 2024 forecast, following the rule set out in Measuring forecast error and reading base rates. Later reforecasts are useful for running the plant, but scoring against them would measure how well the business tracks its own results, not how well it forecast before committing A$5.4m.
Updating the records
The forecast for the plant was built from three records: how prospects convert, how much converted customers order relative to what they said, and how fast a new operation ramps up. The plant has added to each, and the review entered the new evidence beside the old.
Conversion. Where the class of each prospect is known, segments B and D, the plant converted 4 of 7 letters of intent: three of the four RTD letters and one of the three overflow letters. Of the five RTD prospects without letters, one became a customer. In segment A, 17 of 40 named prospects had become customers by month 21, 42.5%, against the 50.2% the record implied and the 75.0% the business case assumed. In segment C, 9 of 20 had become customers, 45.0%, against 43.3% from the record and 60.0% in the business case.
Realisation. Measured as each customer’s deseasonalised annual volume divided by the volume the plan assumed for it, the plant’s customers realised: craft brewers 0.750 each, regular non-alcoholic brands 0.800 each, the three RTD letters 0.850, 0.750 and 0.600 (the last before it left), the RTD customer without a letter 0.800, the overflow customer 1.250, and H 8.0. Apart from H, every value lies within or just beyond the range of the 17 rig 2 customers, 0.42 to 1.24. H is recorded with a flag and its cause, so that a future simulation can allow for such a customer without treating it as typical.
Ramp. Measured as a share of the month-21 run-rate, the plant first passed 0.50 in month 8 and 0.90 in month 16, with a first-year average of 0.422. Rig 2 had reached the same points in months 9 and 17, with a first-year average of 0.367. The plant ramped a little faster than the rig, but its month-21 rate is not yet a mature level, so these figures will be revised at the next review. Both are far from the business case’s 70% first year.
The second-line proposal
In October 2026 the managing director proposed a second line at a capital cost of A$3.9m, dedicated mainly to slim cans. The proposal assumed that H would grow from 400,000 cans a month to 800,000 by the end of 2027, an increase of about 33,300 cans a month, and argued that the existing line would be full within a year. The review tested it in three ways: capacity, value and concentration.
The capacity test
At the month-21 run-rate the plant runs about 330 jobs a year: 17 craft brewers at about 10 runs each, three RTD brands at 10, eight regular non-alcoholic brands at 12, H at 26 and the overflow customer at 8. Using the conversion from Chapter 2, 16.950m cans at 9,000 cans an hour is 1,883 hours of running, and 330 changeovers at 1.5 hours add 495 hours, a total of 2,378 hours, 59.5% of the 4,000 scheduled. Even in December, the busiest month with a seasonal index of 1.38, the line needs about 274 of its 333.3 scheduled hours, 82.1%. The line has 1,622 spare hours a year.
If H doubled to 800,000 cans a month, the plant would fill 21.750m cans a year with 26 more runs: about 2,951 hours, 73.8% of scheduled time. December would need 339.3 hours against 333.3 scheduled, 101.8%: 6.0 hours short. A single Saturday shift of 8 hours covers that. Worked on 48 Saturdays a year, an 8-hour shift would add 384 hours and room for 3.456m cans.
The value test
The existing plant needs 21.227m cans a year from 2027 to reach an NPV of zero. With H doubled, it would fill 21.750m, just enough: on the review’s figures its NPV would rise from −A$2,479.2k to about A$204.1k. That outcome needs every other customer to hold its volume. If the anchor RTD brand, at 2.04m cans a year, left at the same time, volume would fall to 19.710m, below the level the existing plant needs. A second line adds A$3.9m of capital to a plant that would only just cover its own cost on the most favourable assumption.
The concentration test
With H doubled, H would be 44.1% of the plant’s volume. A A$3.9m line built for one customer’s projected growth, with no contract behind the growth, would put the business’s capital at the mercy of that customer’s next decision.
The error: a trend and a peak
This is an illustration. The proposal made two errors, each of which made the second line look necessary.
First, it extrapolated a trend that had already stopped. H had grown by about 25,000 cans a month from month 7 to month 19 and had then held at 400,000 cans a month for three months. The proposal assumed growth would resume at about 33,300 cans a month, faster than before, without any order, forecast or agreement from H to support it.
Second, it measured capacity at the peak. The proposal’s capacity table took December’s requirement with H doubled, 339.3 hours, and multiplied it by 12, giving about 4,072 hours a year, 101.8% of scheduled time, and the conclusion that the line would be full. The line’s annual requirement in that case is 2,951 hours, 73.8%; only December is short, and only by 6.0 hours.
The reviewer detected both errors by running the hours test month by month with the seasonal indices from Chapter 8, and the concentration test on the same volumes. The correction was to replace the extrapolated trend with a contract-backed forecast: H’s own forecast of its volume, with minimum quantities it would commit to. The proposal returned to the board as a set of triggers rather than a request for capital.
Conditions instead of a second line
The review recommended three conditions under which the second-line case would be reopened. They are illustrative thresholds, set before the evidence, in the manner of the kill criteria discussed in the stage-gate series.
- Contract first. Any capacity built mainly for H requires a supply agreement with minimum volumes or take-or-pay payments covering at least the capital’s recovery period.
- Concentration limit. The business sets a limit on the share of plant volume from one customer, for example one third, above which further capacity for that customer needs board approval with a downside case in which the customer leaves.
- Utilisation trigger. The case reopens when the plant’s deseasonalised utilisation exceeds a stated level, for example 85% of scheduled hours, for two consecutive quarters. Until then, Saturday shifts cover the peaks.
These conditions do what the original approval’s conditions did not quite do: they name the measure, the threshold and the response before the evidence arrives. The forecast range from From base rates to an honest forecast range, the register from Chapter 8 and the honest accounting of Chapter 9 together give the board the means to apply them. The business that began by trusting a founder’s confidence in ingenuity ends with a forecast register, a set of records, and decisions written down in advance. That is the practical answer to the Hiding Hand: whatever ingenuity the business has will still be there, and the next forecast will not need to rely on it.
Exercise: December with H doubled
Using the review’s figures, how many hours does December need if H doubles to 800,000 cans a month, how many Saturday shifts of 8 hours does the shortfall represent, and does the H-doubled volume make the existing plant’s NPV positive?
Worked answer. With H doubled the line needs 2,951 hours a year. December’s seasonal index of 1.38 scales a twelfth of that to 2,951 ÷ 12 × 1.38 ≈ 339.3 hours, against 4,000 ÷ 12 = 333.3 scheduled, a shortfall of 6.0 hours, which is 6.0 ÷ 8 = 0.75 of one Saturday shift. On value, the existing plant needs 21.227m cans a year from 2027 for an NPV of zero, and the H-doubled volume of 21.750m exceeds it by only 0.523m cans. That holds only if every other customer keeps its volume, and it leaves H at 44.1% of the plant. Neither result supports a A$3.9m second line: the peak is covered by one Saturday shift a year, and the volume that would justify more capacity rests on one customer’s unsigned growth.