Building a multi year registration baseline from five editions of history

Every pacing conversation I have sat in eventually reduces to one sentence. We are 8 per cent behind last year. The word doing the damage is last, because it commits the entire judgement to a single prior edition chosen for no reason except that it happened most recently.

Last year is one observation, and a multi year registration baseline is what replaces it. That edition had its own weather, its own competing events, its own week when the registration site fell over for a day. Judging this edition against it is a comparison of two draws from a process nobody has characterised, and the answer will be wrong in a direction you cannot predict.

What is a registration baseline made of?

The object you want is a band, defined at each day out, describing the range of cumulative shares your show has historically held at that point.

The unit has to be share of final registrations rather than a count, because five editions of a growing show produce five curves that never touch. Divide each edition's cumulative count at each days out value by that edition's final registration number, and every curve starts near zero and ends at one. The only thing left to compare is timing.

That normalisation has a cost worth stating at the front. You cannot compute a share of final until the edition has finished, so the baseline is built entirely from closed editions and the current edition contributes nothing to it. History supplies the shape. The live count is what you apply the shape to. What a curve's shape actually tells you, including what makes one edition's shape break, is A2's subject.

Five editions of one show

Take a components and subassemblies show with five closed editions, finishing on 8,200, 8,900, 9,600, 10,300 and 11,000 registrations. The cumulative share of final registrations held at each days out value:

Work day minus 60 by hand. Sorted, the five shares are 0.55, 0.57, 0.59, 0.61 and 0.63. The median is the third value, 0.59. The lowest edition held 0.55 and the highest 0.63, so the observed spread is eight points of share.

Eight points is a lot. On a show closing at 10,000 registrations, eight points of share is 800 registrations of ambiguity about where day minus 60 should sit, before anything has gone wrong.

Before trusting the median, check the shares for drift. This show grew 34 per cent across the five editions, from 8,200 to 11,000, and if that growth had come from an audience that registers later then the day minus 60 share should be sliding downwards edition by edition. Read them in order: 0.61, 0.57, 0.63, 0.55, 0.59. That is noise around a level, with no direction in it. A median is a fair summary of numbers behaving like that. If the same column had read 0.63, 0.61, 0.59, 0.57, 0.55, the median would be a summary of a moving target and the correct answer would be to use the most recent edition or two and accept a wider band.

What does a percentile mean when you have five numbers?

People reach for the tenth and ninetieth percentiles to define the band, and it is worth knowing exactly what your software hands back when you ask for those from five values.

The common linear interpolation rule places the p-th percentile at position 1 plus p times n minus 1 in the sorted list. With five observations and p of 0.10, that is 1 plus 0.4, which is position 1.4: four tenths of the way from the lowest value to the second lowest. Between 0.55 and 0.57 that gives 0.55 plus 0.4 times 0.02, which is 0.558.

The ninetieth percentile sits at position 1 plus 3.6, which is 4.6, so six tenths of the way from 0.61 to 0.63. That is 0.61 plus 0.6 times 0.02, which is 0.622.

The band is 0.558 to 0.622. The observed minimum and maximum were 0.55 and 0.63. Your tenth and ninetieth percentiles are the extremes pulled in by less than a point of share, and no information exists anywhere below 0.55 or above 0.63, because you have never seen an edition there.

That is not an argument against computing them. It is an argument for calling the result what it is: the range this show has been seen to occupy, shrunk slightly. Say band on the chart. Do not say confidence interval, because with five observations the coverage of anything carrying that label is worse than the label implies, which is O20's subject and worth reading before anyone puts a number in front of a board.

Which editions belong in the band

The temptation is to take the last five and stop thinking. Two decisions deserve five minutes each.

The first is exclusion. An edition with a venue change, a date move of more than a few weeks, a strike, or a registration platform migration mid campaign is not a sample from the same process. Leaving it in widens the band with a spread that has nothing to do with normal variation. Taking it out drops you to four observations, which is worse in a different way. My preference is to keep it, mark it on the chart, and check whether the band changes materially with and without it. If it does, report both and let the reader see which editions are doing the work.

The second is weighting. A show that has changed its audience deliberately over five years has an older history that is less relevant, and there is a case for weighting recent editions more heavily. I would resist it for a first baseline. Weighting introduces a parameter you now have to justify, and with five observations the honest position is that you do not have enough data to tell whether the weights help. Build the unweighted band, use it for a cycle, and revisit when you have seven editions.

UFI's Global Exhibition Industry Statistics, published in May 2025, estimated 32,000 exhibitions held worldwide in 2024 across 1,432 venues, drawing 318 million visitors and 4.7 million exhibiting companies, with 138 million square metres of space rented against 143.7 million in 2019, an average change of minus 0.8 per cent a year. Industry aggregates that stable are exactly why they are useless as your baseline. Individual shows move by an order of magnitude more than the market does, and the variation you need to characterise is your show's, on your registration file. That is the whole case for running attendee analytics against your own history instead of a benchmark deck.

The UFI Global Exhibition Barometer, 36th edition, published in January 2026 from 378 companies across 57 countries and regions, found 47 per cent of respondents reporting activity up more than 5 per cent in 2025 and 10 per cent reporting a fall of more than 5 per cent, with 42 per cent inside that range. More than half of the industry moved by more than 5 per cent in a single year. A baseline band only three points of share wide has almost certainly been built on a quiet run of editions and will be breached the first time the show has an ordinary bad year.

Reading the current edition against the band

At day minus 60 the current edition holds 6,400 registrations. Divide by the band rather than by a single number: 6,400 over 0.622 is 10,289 and 6,400 over 0.558 is 11,470, with the median share giving 6,400 over 0.59, which is 10,847.

The reading is that this edition is consistent with a final somewhere between roughly 10,300 and 11,500. That is a wide answer and it is the honest one at day minus 60 with five editions of history. Converting it into the single number a show director will ask for, and deciding which end of the band each operational decision should be made against, belongs to A9.

The band also gives you a cleaner version of the question people actually want answered. Rather than asking whether the show is behind, ask whether the current share sits inside the band. That is a different question from an index against a single prior edition, which is A1's. If the edition were holding 5,300 at day minus 60, its implied final would run from 8,520 to 9,500, below every edition since the first, and that is a finding rather than a reading.

Where this stops

Five observations is a small number and no arithmetic fixes that. The median of five has a wide sampling error, the percentiles are barely inside the extremes, and adding a sixth edition will move the band more than you expect. Anyone promising precision from this has not done the calculation, and how much model you can afford on five points is O11's subject.

Biennial shows have it worse. Five editions is a decade of history, over which the audience, the venue and the sector have all changed enough that the oldest editions are describing a different show. For a biennial I would use three editions, accept the band is nearly useless, and lean harder on composition.

The last limit is the one that catches people out. The band describes shape and says nothing about level. Five editions can produce a beautifully tight band while the show halves in size, because every edition would still have held about 0.59 of its own final at day minus 60. A current edition sitting perfectly inside the band is not a healthy show. It is a show whose registrations are arriving in their usual order.

Start with the column you can build fastest. Pull the final registration count for the last five editions of one show, then compute the cumulative count at day minus 60 for each from the created dates on the registration records, and divide. Five numbers. Sort them, take the middle one, and note the gap between the lowest and the highest, because that gap is the precision your pacing conversation has actually been running at all along.