Building a registration forecast to show open your team will use

The show director wants a number by Friday. Operations needs it for the badge order, finance needs it for the revenue reforecast, and the sales team wants it for a slide they are showing to a prospective exhibitor next week.

The number all four want is a registration forecast to show open. The temptation at this point is to build something impressive. Fit a curve, add a seasonal term, put a confidence interval on it that nobody will interrogate. My experience is that the elaborate forecast gets built, gets shown once, and then somebody in operations quietly keeps using last year's number plus a bit, because they can explain that one.

The forecast a team actually uses is one that fits on a line and can be checked in the meeting.

The division that is the whole method

Divide the current registration count by the share of the final registration number your show has historically held at this point in the campaign.

At day minus 45 the file holds 6,200 registrations. Across the last five editions, the cumulative share of final registrations at day minus 45 was 0.58, 0.60, 0.62, 0.64 and 0.66. The median is 0.62. The forecast is 6,200 divided by 0.62, which is 10,000 registrations at show open.

That is the entire method. Two numbers and a division, both of which anyone in the room can question: the count, which comes off the platform, and the share, which comes off your own closed editions and needs a band built from five of them, which is A7's subject.

The method has a formal name and a literature. Hotels and airlines have run it for decades as a multiplicative pickup model, scaling the current on-books figure by a historical ratio, and the diagnostics that tell you whether the multiplicative form or the additive one suits your data belong to O2 and O3. What follows is the operational version, which is what a show team needs in week 45.

Write the assumption down

One assumption is doing all the work: this edition's registrations are arriving in the same order as previous editions, so the fraction of the file that has arrived by day minus 45 is about what it usually is.

That assumption is checkable and it fails in specific, listable ways.

It fails if registration opened at a different distance from doors, since a campaign that started six weeks later has had less time to accumulate and its day minus 45 share will be lower for reasons that have nothing to do with demand. It fails if the price tier boundaries sit at different days out, because a tier deadline pulls volume forward and moves the share at every point after it. It fails if a bulk upload of exhibitor guest passes or an association list landed at a different point in the campaign. It fails if the registration type mix has moved, since types register on different schedules and the blended share is a weighted average of them, which is A31's territory.

Print those four checks next to the forecast. A number that arrives with the conditions under which it is wrong is a number people will keep using, and one that arrives alone is a number people stop trusting the first time it misses.

Freeman's end of year trends recap, released in January 2026 from its 2025 research, found 50 per cent of the events it tracked running behind their regular registration pace that year. That is the fifth failure mode and the hardest to see, because it hits every edition at once. If the whole market is registering later than it used to, every show's share at day minus 45 is lower than its own history says, and a ratio forecast built on that history will read high across the board. The defence is to check whether your recent editions' shares are drifting down and to weight the recent ones if they are.

Why is a point estimate the wrong deliverable?

The single number is what gets asked for and the band is what should be handed over.

Divide by the extremes of the share band as well as the median. Six thousand two hundred divided by 0.66 is 9,394. Divided by 0.58 it is 10,690. The forecast is 9,400 to 10,700, centred on 10,000.

Notice what happened to the symmetry. The five shares were symmetric around the median, 0.04 below and 0.04 above. The implied finals are not: 606 below the point estimate and 690 above it. Dividing by a share is a reciprocal, so equal uncertainty in the share produces more room on the high side than the low. Anyone who takes a symmetric plus or minus off the point estimate has quietly moved the forecast down.

Which end you plan against depends on the decision, and the costs are not symmetric either. A catering guarantee wants the low end, because you pay for what you guarantee. Badge stock and registration desk staffing want the high end, because running out on the first morning is a visible failure and a box of unused badge stock is not. Handing one number to both decisions makes one of them wrong on purpose. Carrying a range through to each decision is the part of attendee analytics that operations actually uses.

The forecast should get narrower every week

Recompute weekly and the band should tighten, because the share you are dividing by is larger and better determined the closer you get to doors.

Take the same show at day minus 14, holding 8,300 registrations, with historic shares at that point of 0.80, 0.82, 0.83, 0.85 and 0.86. The median gives 8,300 divided by 0.83, which is 10,000 again. The band runs from 8,300 over 0.86, which is 9,651, to 8,300 over 0.80, which is 10,375.

The band was 1,296 registrations wide at day minus 45 and it is 724 wide at day minus 14. Same method, same show, nearly half the uncertainty, and the point estimate has not moved.

Publish that narrowing. A forecast that visibly tightens as the campaign progresses is a forecast people learn to read, and it sets the expectation that the number in week 45 was never meant to be the number in week 14. It also gives you an early warning that is otherwise invisible: if the band stops narrowing or the centre starts sliding week after week, the shape assumption is failing and you should stop dividing and go and find out why.

Does this beat last year's number?

Before this goes into a pack, run it backwards against your own closed editions.

Hyndman and Athanasopoulos, in Forecasting: Principles and Practice, third edition, published by OTexts in 2021, put the discipline plainly: simple benchmark methods exist so that "any forecasting methods we develop will be compared to these simple methods to ensure that the new method is better than these simple alternatives. If not, the new method is not worth considering."

For a once-a-year show the dumbest alternative is last edition's final number. Run both on the last three closed editions.

The edition that closed on 9,180 held 5,690 at day minus 45, and the median share from the editions before it was 0.61, giving a forecast of 9,328. The error is 148. The naive forecast was the prior final of 8,860, an error of 320.

The edition that closed on 9,540 held 5,980, median share 0.62, forecast 9,645, error 105. Naive was 9,180, error 360.

The edition that closed on 10,120 held 6,410, median share 0.63, forecast 10,175, error 55. Naive was 9,540, error 580.

Mean absolute error for the ratio method is 148 plus 105 plus 55, divided by three, which is 103. For the naive method it is 320 plus 360 plus 580, divided by three, which is 420. The ratio method is about four times more accurate on this show.

Three error pairs is a thin test and I would not defend it as evidence of anything except that the method is not obviously worse. Comparing two forecasting methods on four observations is genuinely hard, and treating the comparison as a permanent scorecard rather than a one-off is what O30 calls forecast value added. Run it every year and keep the record.

Where this stops

The method cannot see anything that has not happened before. A competitor launching into your week, a venue change, a strike at the airport, a date move: each one breaks the shape assumption, and the forecast will carry on producing a plausible number throughout, because the arithmetic has no way to know.

It is also useless early. At day minus 120 with a share of 0.30, the divisor is small enough that a tiny error in it swings the answer hard. Three thousand registrations divided by 0.30 gives 10,000, and the same count divided by 0.27 gives 11,111. A three point error in the share moves the forecast by more than a thousand. I would refuse to publish a forecast before the historic share reaches about 0.50, and say why.

The last limit is the one that matters most on show day. This forecasts registrations. The number operations needs is people in the building, which is the registration forecast multiplied by a turnout rate with its own error attached, and that multiplication is A10's subject.

Start with the backtest rather than the forecast. Take your last three closed editions, compute what this method would have predicted at day minus 45 for each of them, and put those three numbers next to the actual finals and next to the prior year's final. If the method does not beat last year's number on your own history, you have learned something more valuable than a forecast.