Upside surprises and honest accounting

Every forecast that is wrong overall is wrong by different amounts in different places, and occasionally one part of it is wrong in the business’s favour by a wide margin. That part tends to get a great deal of attention. An upside surprise is good news, it is usually somebody’s achievement, and it is tempting to let it stand for the whole investment.

This chapter follows one such surprise at the canning plant: a functional drinks brand that grew to eight times the volume planned for it. It measures the surprise against the 2026 shortfall in two ways, by segment and by customer, and shows why only the second is honest. It separates three accounts that a single variance line mixes together: the error in the original forecast, luck, and the decisions made after approval with their costs. It then returns to the Hiding Hand, the argument that ingenuity makes up for underestimated difficulties, and tests it on the plant’s own numbers.

The customer nobody planned for

This is an illustration. All money is A$ excluding GST. Among the 20 named prospects in segment C, the non-alcoholic drinks, was a small functional hydration brand, called H here. H had not signed a letter of intent. In the business case it was simply one of the 12 non-alcoholic customers the managing director expected, at the segment’s standard 600,000 cans a year and A$0.14 a can of contribution.

H wanted its drink in 250 mL slim cans, a format the line could not run. In June 2025, month 6, the business spent A$185,000 on 250 mL change parts and a seamer tooling set to take the work, after a series of trials, and agreed a contribution of A$0.12 a can, below the segment’s A$0.14, to win it. H’s first commercial run was in July 2025, month 7, at 100,000 cans a month after removing the season. Its sales then grew quickly: its volume rose by about 25,000 cans a month until month 19, July 2026, when it reached 400,000 cans a month and levelled off.

At month 21, H’s run-rate was 4.8m cans a year: 8.0 times the 0.600m planned for it, 28.3% of the whole plant’s run-rate of 16.950m and 55.6% of segment C. In 2026, including the reforecast for October to December, H filled 4.229m cans, 26.6% of the plant’s 15.886m. One brand that nobody had treated as important had become the plant’s largest customer by a wide margin.

The 2026 gap by segment

The register compares each segment’s 2026 contribution, nine months of actuals and three months reforecast at the month-21 rate, with the business case’s 95% year.

Figure 12026 contribution by segment against the business case

	Plan volume, m cans	Actual and reforecast, m cans	Plan, A$k	Actual and reforecast, A$k	Gap, A$k
A Craft brewers	3.078	1.236	615.6	247.3	−368.3
B RTD and seltzer	11.400	5.621	1,140.0	562.1	−577.9
C Non-alcoholic	6.840	7.829	957.6	1,011.4	+53.8
D Overflow	2.736	1.200	328.3	144.0	−184.3
Total	24.054	15.886	3,041.5	1,964.8	−1,076.7

Read by segment, the picture is of three segments well short and one slightly ahead. The segments below plan lost A$1,130.5k between them; segment C gained A$53.8k, offsetting 4.8% of that. On this view the upside is a rounding error, and segment C did about what the business case said it would.

Both conclusions are wrong, and for the same reason. A segment total nets a very large gain against a large loss inside the segment.

The customer-level view

Splitting segment C into H and everyone else shows what the segment total hides.

• H contributed A$507.4k in 2026. As planned, one of 12 customers at 600,000 cans, 95% of steady state and A$0.14 a can, it would have contributed A$79.8k. H’s excess over its plan was A$427.6k.
• The other non-alcoholic customers, eight regular brands against the 11 the plan still needed after H, contributed A$504.0k against a plan of A$877.8k, a shortfall of A$373.8k.

The segment’s +A$53.8k is A$427.6k − A$373.8k. Put the other non-alcoholic customers’ shortfall back beside the other segments, and the gross shortfall of everything except H’s excess is A$368.3k + A$577.9k + A$184.3k + A$373.8k = A$1,504.3k. That is the amount by which the customers the forecast described fell short in 2026.

Against that gross shortfall, H’s excess of A$427.6k offsets 28.4%. But the excess was not free: it needed the A$185,000 of change parts. Net of that outlay, H’s excess was A$242.6k, which offsets 16.1% of the gross shortfall. These are the honest figures for what the surprise did for 2026: it made up about a sixth of the shortfall after paying for itself, not the 4.8% the segment table suggests, and not the whole of it, which is what a sponsor’s narrative tends to imply.

Figure 2The 2026 gross shortfall against H's excess

Gross shortfall, A$1,504.3k

• Craft brewers, −A$368.3k
• RTD and seltzer brands, −A$577.9k
• Overflow, −A$184.3k
• Other non-alcoholic customers, −A$373.8k

H's excess over its plan

• Contribution A$507.4k against A$79.8k planned
• Excess A$427.6k: 28.4% of the shortfall
• Less change parts A$185.0k
• Net excess A$242.6k: 16.1% of the shortfall

The same arithmetic at the month-21 run-rate gives the steady picture. H contributes A$576.0k a year against A$84.0k planned, an excess of A$492.0k a year, and the change parts paid for themselves out of that excess in about 0.38 of a year. Over the plant’s life the decision to take H’s work was clearly a good one. That is a separate question from whether the forecast was good.

Three accounts, not one

A single variance line, “segment C: +A$53.8k”, mixes three different things. Honest accounting keeps them apart, because each answers a different question and each leads to a different lesson.

The forecast’s error. The business case described customers: how many, of what kind, ordering how much. Measured against what those customers delivered, the forecast was short by A$1,504.3k in 2026. That is the number that belongs in the business’s forecast record, beside the reference-class evidence from Chapters 3 and 6, and it is the number the post-completion review should use when it asks how good the forecast was.

Luck. H’s growth was a market event. Demand for functional hydration drinks rose and H was well placed to catch it. Nobody at the plant forecast it, and nothing in the record could have: the largest realisation among the 17 rig 2 customers was 1.24, and H realised 8.0. Luck of this kind is real value to the business, but it says nothing about forecasting skill, and it may not repeat. It also cuts both ways: the RTD brand that dropped its seltzer range after month 15 was bad luck of the same kind.

Later decisions and their costs. Converting H’s luck into contribution took decisions made after approval: trials of a new can format, A$185,000 of change parts, and a price of A$0.12 a can instead of A$0.14. Each has a cost that should sit beside the gain. The change parts are visible; the price concession is not, and at the run-rate it is worth A$96.0k a year, the difference between A$672.0k at A$0.14 and the A$576.0k actually earned. Whether that concession was needed to win the work is a fair question for the review. Introducing a new format on a running line is a small first-of-a-kind problem in its own right; Buying down technology risk: qualification, guarantees and separable modules in First-of-a-Kind Technology by the Numbers covers how to qualify such a change before committing to it.

The Hiding Hand on the plant’s results

The Hiding Hand and the evidence on optimistic forecasts set out the debate. Hirschman argued that planners who underestimate a project’s difficulties also underestimate their own ability to solve them, so that the two errors tend to offset. Flyvbjerg’s 2016 test, on a large sample of mostly large public projects, rejected the principle at p < 0.0001 and found cost overruns and benefit shortfalls arriving together. Ika (2018) answered that a fair test must count full life-cycle benefits, unintended effects and the problem-solving that projects provoke. Flyvbjerg (2018) accepted that the Hiding Hand is found, in about one fifth of projects in the author’s data, but not as the typical case.

The plant is one project, not a sample, but it can be read the same way. Its capital cost came in at A$6.10m, a ratio of 1.130 to budget, an overrun of 13.0%. Its steady-state contribution at month 21 was A$2,109.0k a year, a ratio of 0.659 to the business case, a shortfall of 34.1%. That is a cost overrun and a benefit shortfall together, the pattern the 2016 study reported for the average project.

Ika’s critique deserves a fair hearing here, because H is exactly the kind of unintended effect and problem-solving that it says a test should count. The business did find a way, as the managing director had promised: it qualified a new can format, priced the work to win it, and turned an unremarkable prospect into its largest customer. That ingenuity is real and should be credited to the plant’s results. Counted fully, it recovered 16.1% of the 2026 gross shortfall net of its cost, and it is the main reason the plant’s month-21 run-rate of 16.950m beat the outside view’s expected 15.367m.

What the ingenuity did not do is make up the shortfall. The plant is a case of a partial Hiding Hand: a real offset, smaller than the difficulty it met. Flyvbjerg’s “about one fifth” leaves room for a business to hope its project will be among the rescued ones; this plant shows that even a rescue of this size, an eightfold customer, need not be enough. Ingenuity is a capability worth having. It is not a forecasting method.

Concentration: the other side of the surprise

An upside of this size changes the business’s risk as well as its income. At the month-21 run-rate, H provides 27.3% of the plant’s contribution and, because fixed costs do not change, 71.2% of its EBITDA: A$576.0k of A$809.0k. If H were lost, run-rate volume would fall to 12.150m cans, contribution to A$1,533.0k and EBITDA to A$233.0k a year. Chapter 10 shows what that does to the plant’s re-appraised NPV.

A business that owes most of a plant’s earnings to one brand’s growth is exposed to that brand’s next decision: to move to a larger co-packer, to build its own line, or to lose its market as quickly as it found it. That exposure is part of the honest account of the surprise.

The error: a segment ahead of plan

This is an illustration. The managing director’s report to the board for the quarter to June 2026 showed segment C “ahead of plan” for the year, on a forecast of +A$53.8k, and presented the plant’s diversification beyond beer, benefit B4 at 85.0% against a target of 75.8%, as the business case’s strategic aim achieved. The report did not mention that the other non-alcoholic customers were A$373.8k below their plan, and it did not include the A$185,000 of change parts, which had been booked to plant maintenance in 2025.

The error was detected through the register described in Tracking benefits against a phased plan. The sales manager tracked active customers and volume by customer as leading indicators, and the customer-level figures did not reconcile with the segment line: eight non-alcoholic customers at about 0.8 of their planned volumes cannot add up to a segment ahead of plan without one customer far above it. The finance manager then traced the change parts out of maintenance.

The correction was to report segment C as three lines: H, the change parts and other costs of H’s work, and the remainder of the segment. B4 was annotated to show that 4.8m of its non-beer volume came from one customer. Neither change made the plant look worse than it was; both made the board’s next decision, about a second line, rest on what had actually happened.

What an upside should change

An honest account of the surprise still leaves the business with something valuable, and it should act on it. Three things should change.

• Capacity planning. H’s volume uses line hours, and the question of whether the line can absorb further growth is a real one. Chapter 10 tests it.
• Pricing. The next contract with H should be priced on what the work is worth now, including the format’s change parts and changeovers, not on what was needed to win a small prospect.
• Contract terms. A customer that is 71.2% of a plant’s EBITDA should be on terms that share risk both ways: minimum volumes or take-or-pay payments that protect the plant if H leaves, and volume tiers that share the benefit if it grows. Target cost with pain and gain share: worked share calculations in Project Teams and Contracts by the Numbers shows how such sharing arrangements are calculated.

One thing should not change: the forecast’s record. The business case was short by A$1,504.3k in 2026 for the customers it described, and that is the figure that should go into the forecast register and inform the next forecast. H’s realisation of 8.0 should also be recorded, flagged as an extreme value with its cause, so that a future simulation can include the possibility of such a customer without treating it as typical.

Figure 3Questions to ask of any upside surprise

• Do: Split the variance by customer as well as by segmenta segment total can hide a large loss behind one large gain
• Do: Put the costs of capturing the upside beside the gainchange parts, trials, price concessions, extra changeovers
• Do: Record the forecast's error without the surprisethe gross shortfall is the forecast's score
• Ask: Was the upside in the forecast, or in the record behind itH realised 8.0; the record's highest was 1.24
• Ask: Is it luck, a later decision, or botheach needs its own line and its own lesson
• Ask: What happens if the customer behind the upside leavesH is 71.2% of run-rate EBITDA
• Ask: Do the contract terms share risk both waysminimum volumes, take-or-pay, volume tiers

Exercise: a better price for H

Suppose H had agreed to the segment’s standard A$0.14 a can instead of A$0.12, with everything else unchanged. Recalculate H’s 2026 contribution, its excess over plan, and the share of the gross shortfall it offsets, gross and net of the change parts.

Worked answer. H’s 2026 volume of 4.229m cans at A$0.14 gives a contribution of A$592.0k. Its plan is unchanged at A$79.8k, so the excess becomes A$592.0k − A$79.8k = A$512.2k. The gross shortfall is also unchanged at A$1,504.3k, because it excludes H’s excess, so H now offsets 512.2 ÷ 1,504.3 = 34.0% of it, and net of the A$185,000 of change parts, (512.2 − 185.0) ÷ 1,504.3 = 21.7%. The two cents a can were worth A$84.6k in 2026 and about five points of offset. Even at the full segment price, H recovers about a fifth of the shortfall after paying for its change parts: a much better outcome than the segment table suggests, and still a long way from the business case.