How the additive pickup model forecasts your final registration number

It is Tuesday, sixty days before doors, and the show director wants a final registration number for the operations meeting on Thursday. The registration file says 2,010. Last year's final was 4,585. Somebody in the room will say "we are tracking about the same as last year" and somebody else will say "we were later last year", and neither of them can settle it.

The additive pickup model settles it in one line of arithmetic. Take the number of registrations each past edition gained between the same point in its campaign and its close, average those gains, and add the average to what you are holding today.

The whole method is one subtraction and one addition

Four prior editions of a machine tools show, each measured at 60 days before doors and again at close.

The 2022 edition held 1,712 registrations at day 60 and finished at 4,015. The 2023 edition held 1,905 and finished at 4,410. The 2024 edition held 1,760 and finished at 4,190. The 2025 edition held 1,983 and finished at 4,585.

Subtract each pair. The gains are 2,303, 2,505, 2,430 and 2,602. Those four numbers are the pickup, which is the quantity the model estimates. Their average is 2,460.

Add it. Today's 2,010 plus 2,460 gives 4,470 registrations at show open.

Two things about that number are worth saying out loud in the meeting. It is arithmetic anyone can redo on the back of an agenda, which is most of why the method survives. And it is built on four observations, which is a fact the forecast should carry with it rather than shed on the way to the slide.

Where the method comes from, and why it is called pickup

The vocabulary is borrowed. Airlines and hotels have been forecasting a final number from a partial booking file for decades, and they call the bookings still to arrive the pickup, because that is what gets picked up between now and departure.

Lee set out the statistical treatment in 1990 in an MIT Flight Transportation Laboratory report on airline reservations forecasting, modelling the booking process itself and separating methods that use the shape of advance bookings from methods that use only the history of final totals. Weatherford and Kimes (2003), comparing forecasting methods on hotel data in the International Journal of Forecasting, sort the field into three families on exactly that distinction: historical booking models, which look only at the final numbers; booking curve or pickup models, which use the build-up pattern during the lead time; and combined models, which mix the two.

A registration campaign is the same object. You have a build-up pattern per edition and a final total per edition, and the pickup family is the one that uses both.

The one difference worth noticing is the number of past instances. A hotel forecasting next Tuesday has hundreds of previous Tuesdays. An annual trade show has four or five previous editions. Everything awkward about applying this method to a show follows from that.

What does the spread of the four pickups tell you?

More than the average does, and it is the part people skip.

The four gains were 2,303, 2,505, 2,430 and 2,602. Their deviations from the mean of 2,460 are minus 157, plus 45, minus 30 and plus 142. The standard deviation of those four numbers is about 126 registrations.

Take two of those as a rough width and the forecast becomes 4,470 with a band of roughly 4,218 to 4,722. That band is 504 registrations wide on a forecast of 4,470, which is about 11 per cent, and it is the honest answer to "how sure are you". It is also, and this matters, a band estimated from four numbers, so it is itself a shaky estimate. Three residual degrees of freedom is not much to build a variance on, and what five editions can honestly pay for is O11's argument.

Publish the band. A point estimate presented alone invites the operations lead to order badges for exactly 4,470 people, and the range is the thing that tells them whether to hold the second badge run.

Which day out should you measure the pickup at?

The choice of measurement point is a real decision and most teams make it by accident, using whichever snapshot happens to exist.

Two forces pull against each other. Measure late and the pickup is small relative to the count in hand, so an error in the pickup estimate moves the forecast very little. At 14 days out the machine tools show would be holding most of its file and the remaining gain might be 700 registrations, so being 20 per cent wrong about the pickup costs 140 registrations on a forecast near 4,500. Measure early and the reverse happens. At 150 days out the count might be 400 and the pickup 3,900, and being 20 per cent wrong about the pickup costs 780 registrations, which is 17 per cent of the answer.

Late measurement is more accurate and less useful, because by 14 days out the badge order has gone and the hall plan is fixed. The measurement point that earns its keep is the last one at which somebody can still act on the answer, which in most operations calendars is somewhere between 60 and 90 days.

You do not have to pick one. Compute the pickup at 120, 90, 60 and 30 days out from the same four editions and watch the four forecasts converge as the show approaches. A forecast that jumps around between those points is telling you the campaign shape moved this year, which is information the single number would have hidden.

What does the additive form assume about growth?

The additive model contains an assumption that is easy to miss because it never appears as a line in the calculation.

Adding a fixed increment says the registrations still to come do not depend on how many are already in. Whether you are holding 1,712 or 1,983 at day 60, the model expects about 2,460 more. That is a real claim about how the campaign works, and on a stable show it is close enough to true to be useful.

On a show that is growing hard it is wrong in a direction that costs money. Suppose the show adds 15 per cent a year. The 2022 edition's gain of 2,303 came from a campaign that finished at 4,015, and reusing it as a prediction for a campaign heading for 5,300 imports the smaller show's absolute volume into the larger show's forecast. Averaging four such gains drags the forecast toward the middle of the historical range at precisely the moment the show has left that range.

The fix is to model the ratio instead of the difference, which is the multiplicative form of the same idea and belongs to O2. Deciding which of the two your own file supports is a measurable question with a residual comparison behind it, and that is O3's diagnostic. This post stops at the additive version and its own failure modes.

The two things that will actually break your pickup number

Neither is exotic and both have happened to every show team I have worked with.

The first is a bulk load at an unusual point in the campaign. An association hands over a member list, or an exhibitor uploads 400 guest passes in one afternoon. If that landed at day 40 in 2024 and lands at day 75 this year, the 60 day snapshot is not measuring the same thing in the two editions, and the pickup you computed from 2024 includes an event that has already happened this time. The forecast then double counts, or misses, several hundred registrations. Keep a note of bulk loads with dates against every edition and inspect it before you trust a snapshot.

The second is a change in the registration window. If registration opened 150 days before doors in 2024 and 210 days before doors in 2026, the day 60 counts are not comparable, because one campaign had 90 days of accumulation behind it at that point and the other had 150. This is a bigger problem than it looks and it has its own treatment in the cluster.

There is a third, quieter one. Deduplication runs and cancelled block bookings can make the cumulative count go down between two snapshots, which produces a negative pickup for that interval. That is a data artefact rather than demand, and the smoothing question it raises is a separate subject.

Where this stops

The additive pickup model has no opinion about anything except volume, and volume is not the only thing an operations meeting needs.

It cannot tell you the mix. A forecast of 4,470 says nothing about how many of those are buyers rather than exhibitor staff, or how many are international, and the registration type mix moves the badge order, the catering order and the hall plan more than the headline does. If the type mix is drifting, the aggregate pickup is a weighted average of type-level pickups whose weights are changing under you, and the aggregate can be right while every component is wrong.

It also has nothing to say about turnout. Registrations are not attendance, and multiplying a registration forecast by a show-up rate introduces a second error that is usually larger than the first.

The deeper limit is the one Weatherford and Kimes were working with the opposite of. Their hotel comparisons had years of daily arrivals behind them. You have four increments. Four numbers can tell you roughly where the centre is and can barely tell you how wide the spread is, and no amount of method sophistication creates a fifth observation. That is an argument for keeping the model this simple until the data justifies more, and for spending the effort instead on making the snapshots comparable, which is where most of the recoverable error lives.

Pull your registration file's snapshot history for the last four closed editions, find the count at the same days-to-open value in each, and subtract it from each edition's final. Write the four gains and their average on one line. That line is your pickup, it takes twenty minutes, and it gives Thursday's meeting a number with its own history attached. Where a system holding those snapshots has to be careful is covered in the rest of the forecasting methods work.