Getting a final attendance forecast from registrations that survives show day

The venue wants a number for the peak day. The caterer wants one for the guarantee. The shuttle contractor wants one for the vehicle count, and the security lead wants one for the door plan. All four ask the same question in the same week and all four get given the registration figure, because it is the number that exists.

The registration figure is wrong for every one of those purposes, and it is wrong in the same direction each time. A final attendance forecast from registrations has to multiply that figure down by a turnout rate, because it is too high by an amount that varies by show, by edition and by who is in the file.

Why does turnout differ by registration type?

The reason a single blended turnout rate misleads is that your file contains several populations who made different decisions to be there.

Somebody who paid 1,400 for a conference pass has already spent the money and will make the trip. Somebody who clicked through a free expo registration in ninety seconds on a Tuesday has committed nothing and will come if the week allows. A hosted buyer with a paid flight and a meeting diary is contractually close to certain. Those are different probabilities and averaging them into one rate throws away the only structure the file has.

Maritz's Registration Insights Report 2024, drawn from more than 360,000 attendee registration records across 30 trade shows, found exhibitors slightly more likely to register late than attendees, 48 per cent against 45 per cent in the final four weeks, and traced late registration to first time attendees and to people within driving distance. Registration types behave differently on timing, and the same split that makes them behave differently on timing makes them behave differently on turning up.

Estimate the rate per type from your own last two or three editions: verified attendances of that type divided by registrations of that type. Two or three editions is enough for turnout in a way it is not enough for pacing, because turnout rates are stable to within a few points when the type definitions hold still. A turnout rate is the complement of a no show rate, so the definition and denominator of that figure belong to A21, and predicting which individuals will not turn up belongs to A22.

The worked forecast

Take a registration forecast of 10,000 for the coming edition, produced by the division method that is A9's subject, and split it by type using the mix your pre-show file is currently running.

Free expo visitors, 7,000 registrations at a 68 per cent turnout rate, give 4,760 attendances. Paid delegates, 1,900 registrations at 88 per cent, give 1,672. VIP and hosted buyers, 1,100 registrations at 82 per cent, give 902.

Add them: 4,760 plus 1,672 plus 902 is 7,334 attendances from 10,000 registrations. The implied blended rate is 73.3 per cent.

Now do it the way most shows do it, with last edition's blended rate. That edition took 10,400 registrations, made up of 8,600 free expo visitors, 1,300 paid delegates and 500 VIPs. At the same three turnout rates it produced 5,848 plus 1,144 plus 410, which is 7,402 attendances, and a blended rate of 71.2 per cent.

Apply that 71.2 per cent to this edition's 10,000 registrations and you get 7,120. The type by type answer was 7,334. The gap is 214 people, and no behaviour changed to produce it. The turnout rates are identical in both calculations. The difference is entirely that this edition's file has a higher share of paid delegates and VIPs, who turn up more often.

A blended rate is a weighted average whose weights are last year's mix. Using it next year assumes the mix is frozen, and drift in the registration type mix is the assumption A31 exists to break.

The people who are not in your file yet

The forecast above covers people who have already registered or will register before doors. It misses the ones who register at the desk on the morning.

Maritz found 9 per cent of registrations in its sample arriving on site. On a show with that profile, the pre-show file is only 91 per cent of the eventual registration count, so 10,000 pre-show registrations imply 10,000 divided by 0.91, which is 10,989 registrations in total, of which 989 arrive on the day.

On-site registrants attend by definition. They are standing at the desk. Their turnout rate is 100 per cent and they add 989 straight onto the attendance figure, taking it from 7,334 to 8,323. The blended turnout across all 10,989 registrations is then 75.7 per cent.

Before you use any of that, settle what your historic final registration numbers included. If your closed editions were counted after the show, they already contain the on-site registrations, so a forecast built on their shares already includes them and grossing up by 0.91 would count them twice. If your finals were frozen the day before doors, they do not, and the gross up is required. Get this wrong and you have a 9 per cent error sitting inside a forecast that looks careful.

The on-site share is also the least forecastable number in the whole exercise, since the people producing it decided after your last measurement. Estimate it from the last three editions, note that it moves with the drive-in share of your audience, and treat it as the widest part of the range.

What counts as an attendance?

None of the arithmetic means anything until somebody defines the numerator, and the definitions in use across one portfolio are rarely the same.

The version I would defend is a person whose badge was scanned at a hall entrance on at least one day of the show. That is a measurement, it is repeatable, and it produces a number that can be audited. It is not the same as unique daily entries, which is a bigger number, and it is not the same as booth scans, which measures something else entirely and belongs with the timestamp mechanics of badge scanning in E2.

The definition has to be identical on both sides of the division. If you count entrance scans as attendance and registrations as everyone in the platform including cancellations, your turnout rate is depressed by however many cancellations you leave in. Strip cancellations from the denominator, or leave them in on both sides of every historic edition, and write down which you did. A definition written down once is worth more to attendee analytics than any refinement of the model sitting on top of it.

UFI's Global Exhibition Industry Statistics, published in May 2025, put 318 million visitors across 32,000 exhibitions worldwide in 2024, which averages to a little under 10,000 visitors per exhibition. Every one of those visitor figures is the output of a definition somebody chose, and the industry has spent decades arguing about which. Yours only has to be consistent with itself across editions to be useful for planning.

Two ranges multiplied

The registration forecast is a range and the turnout rate is a range, and the honest attendance forecast is what happens when you multiply them.

Say the registration forecast runs from 9,600 to 10,400 and the blended turnout rate from your recent editions runs from 71 to 76 per cent. The low corner is 9,600 times 0.71, which is 6,816. The high corner is 10,400 times 0.76, which is 7,904. The point estimate of 7,334 sits between them, and the planning range is about 1,100 people wide on a show of roughly 7,300.

Multiplying the extremes assumes the two errors move in the same direction, which overstates the width if they are independent. They are not fully independent. A year where demand is soft leaves registrations short and tends to leave turnout weak among exactly the free registrants who make up most of the file, so the corners are correlated and the wide reading is the one I would plan against.

Give the two corners to different decisions. The catering guarantee is a payment for a number you name, so it goes against something near the low corner. Badge stock, entrance lanes and shuttle capacity are cheap to over-provide and expensive to run out of, so they go against the high corner. A single number handed to both is a decision to be wrong on one of them.

Where this stops

Every turnout rate you have is capped by your entrance scanning coverage. If 12 per cent of the people who came through your doors last year were never scanned, because of a side entrance, a VIP route or a steward waving a group past a jammed reader, then every historic turnout rate is 12 per cent low and so is this forecast. The rates will still be internally consistent, which is what makes the error survive for years without anyone noticing. Audit the coverage before you audit the model.

The second limit is that turnout rates are only stable while the type definitions are. Renaming a registration type, merging two of them, or moving a category of visitor from free to paid resets the history for that type and there is no clean way to carry the old rate across. When a type definition changes, the honest move is to say that type has one edition of history.

The third is that this forecasts a total and most operational decisions need a daily peak. A show drawing 8,300 attendances across three days does not put 8,300 people in the hall at once, and the ratio between total attendance and peak concurrent occupancy is a separate measurement with its own history.

Start by computing turnout by type for your last two closed editions. Registrations of each type, verified attendances of each type, and the division, on one page with the definition of attendance written above it. If any type's rate has moved more than a few points between the two editions, find out what changed in the definition before you use either number in a forecast.