Skip to content

A registration desk staffing model built on quarter hour arrival bins

Attendee analyticsUpdated 2026-08-188 min read

In short

A registration desk staffing model divides the opening day into quarter hour bins, converts each bin's expected arrivals into an offered load, and sets agents per bin against a service target. Green, Kolesar and Soares showed in 2001 that this period-by-period approach still understaffs critical periods unless it is corrected for the lag between arrival and completion.

The roster for day one says eleven agents from 08:00 to 17:00. It says eleven because it said eleven last year, and last year it more or less worked, apart from the queue photograph somebody posted at 09:30 that the marketing director still mentions.

A registration desk staffing model that fixes this is neither difficult nor expensive. It needs the arrival profile you already have from last edition, a service time you can measure in an afternoon, and the willingness to roster different numbers of people at different times of day.

What does a flat roster do to a show morning?

Take a desk open from 08:00 to 17:00, which is 36 quarter hour bins, handling 6,480 check-ins on day one. Mean service time is 17.6 seconds, which is 0.2933 minutes.

At 09:15 the profile says 849 arrivals in that quarter hour, which is 56.6 a minute. Eleven agents each complete one check-in every 17.6 seconds, so together they clear 11 divided by 0.2933, which is 37.5 a minute. Arrivals exceed completions by 19.1 a minute, and over the fifteen minutes of that bin the queue grows by 287 people. Nothing about that is recoverable inside the bin. Those 287 people are simply now in the queue, and the next bin starts from there.

At 15:00 the profile says 120 arrivals in the quarter hour, 8 a minute. Eleven agents can clear 37.5 a minute. Utilisation is 8 over 37.5, about 21 per cent, so on average nine of the eleven are doing nothing.

One roster, wrong in both directions, on the same day. The morning failure is expensive in reputation and the afternoon failure is expensive in money, and the flat number produced both.

Bins, and why fifteen minutes

The unit of the model is the bin, and the bin size is a real decision.

Hourly bins are useless for registration, because the entire event you are trying to staff for happens inside one hour. An 09:00 to 10:00 bin containing a violent surge at 09:10 and a lull at 09:50 averages out to something that looks manageable and will not be.

Five minute bins go too far the other way. At 120 arrivals in a quarter hour you have 40 in a five minute bin, and ordinary randomness moves that between roughly 28 and 52 with nothing meaningful behind the movement. You would end up smoothing it, which puts you back at a quarter hour with extra steps.

Fifteen minutes holds the shape of a show morning and carries enough arrivals per bin to be stable. It also matches how people actually work, since you can move two staff between desks at a quarter past and nobody has to renegotiate a break.

Build the bins from last edition's check-in timestamps, scaled to this edition's expected volume. If the show date moved, or the opening time changed, the profile shape moves with it and last year's bins mislead, which is a separate problem worth knowing about before you rely on them.

Turning arrivals into an offered load

Each bin needs converting from people into work.

Offered load is the arrival rate multiplied by the mean service time, both in the same units. At the 09:15 bin, 56.6 arrivals a minute times 0.2933 minutes gives 16.6. That number is in erlangs, and it says the desk needs 16.6 agents working continuously with no idle time at all just to keep level.

That is the floor, and staffing at the floor is a mistake, because at exactly 100 per cent utilisation any randomness at all builds a queue that never clears. You need agents above the offered load, and the margin is what buys you the service target.

A workable approximation for the margin is the square root of the offered load, scaled by a factor that depends on how strict your target is. For a five minute target on a registration desk, a factor of around 0.6 gets you close to what an Erlang C calculation returns, and it has the advantage that anybody can do it on a phone. I use it for planning and check the critical bins properly afterwards.

How many agents does each bin actually need?

Run the arithmetic across the profile.

The 09:15 bin: offered load 16.6, square root 4.07, margin 0.6 times 4.07 is 2.44, so 19.04 agents. Round up to 19.

The 10:00 bin: 560 arrivals, 37.3 a minute, offered load 10.94. Square root 3.31, margin 1.98, so 12.92. Round up to 13.

The 13:00 bin: 180 arrivals, 12 a minute, offered load 3.52. Square root 1.88, margin 1.13, so 4.65. Round up to 5.

The 15:00 bin: 120 arrivals, 8 a minute, offered load 2.35. Square root 1.53, margin 0.92, so 3.27. Round up to 4.

So the requirement across the day runs from 4 to 19, and the flat roster of 11 sits 8 short at the peak and 7 over at the quiet point. Written out that way it stops being a scheduling preference and becomes an arithmetic error.

The interesting thing about the peak bin is how little the margin costs. Going from the bare offered load of 16.6 to 19 is under two and a half extra people, and those two and a half are the difference between a desk that holds level and one that produces the photograph. Going from 11 to 17 fixes almost nothing on its own, because 17 still leaves you working at 98 per cent utilisation, and a desk at 98 per cent utilisation behaves nothing like a desk at 87 per cent even though the numbers look adjacent.

The correction that period-by-period models need

Green, Kolesar and Soares wrote about this in Operations Research in 2001, and they named it the stationary independent period-by-period approach. You divide the workday into planning periods and apply a stationary Erlang C model independently to each one, which is exactly what the bins above are doing.

Their finding is the part worth carrying into your own model. The approach often suggests staffing that is substantially too low to hit the targeted service level during exactly the critical periods you built it for, and they proposed two modifications that reach the target with only modest additional staffing.

The reason is a lag, and it is easy to see on a registration desk. Somebody who arrives at 09:14 is still being served at 09:16, so the work generated by one bin lands partly in the next one. Treating each bin as an independent stationary system ignores that spillover, and the error is largest where the arrival rate is rising fastest, which on a show morning is the steepest part of the ramp.

Green, Kolesar and Whitt reviewed the family of methods in 2007, in Production and Operations Management, and concluded that time-lag refinements improve on both the period-by-period approach and the older peak-hour rule when service times are moderate and the service target is high. A registration desk fits that description.

The practical version needs no additional mathematics. Shift the staffing profile earlier than the arrival profile by roughly one mean service time, then hold the peak staffing level for one bin longer than the arrivals justify. On a 17.6 second service time the shift is small, and the extra bin at the back of the peak is where the real gain sits, because that bin is absorbing the queue the surge left behind.

Rostering against a requirement that changes every fifteen minutes

The requirement curve is the model's output. The roster is a separate problem, because people arrive for shifts and take breaks and cannot be summoned for a single quarter hour.

Three things make the gap manageable. Stagger the shift starts, so that a block of staff comes on at 08:30 and another at 09:00 rather than everyone at 08:00. Schedule every break outside the peak bins, which sounds obvious and is routinely violated because breaks get set by the shift pattern rather than by the profile. And keep a float group of three or four people who are doing something useful nearby and can be on a desk within two minutes.

Then watch the live gap against the plan. Comparing planned agents against required agents for the current bin, alongside the real completion rate, is the entire show office view for registration, and the completion side of it comes from the same registration desk throughput measurement. When the queue does build anyway, converting its length into a published figure is a different calculation, and registration queue wait time needs the arrival rate rather than the roster.

Where this stops

The model assumes the arrival profile is knowable in advance, and the profile is the weakest input in the whole exercise. It comes from last edition, and last edition had different show dates, a different keynote start time, a different transport situation and possibly different weather. A profile that is 20 per cent wrong in the peak bin makes the staffing requirement wrong by roughly 20 per cent in the place where being wrong hurts most.

The honest response is to build the model on last edition's shape and then re-forecast the level once pre-registration pacing tells you the size of the audience, and to size the float group against the uncertainty rather than against a fixed rule. Where the surge itself is the thing you are managing, rather than the desk behind it, show opening surge management treats the doors as the constraint and comes at the problem from the other side.

The second limit is that agents are not interchangeable. The model counts servers, and a server who cannot handle an on-site registration, or who does not speak the language of the delegation that just arrived, is a server for some of the queue and not for the rest. Every specialised sub-queue is its own system with its own requirement, and averaging them together produces a roster that is right on paper and short at the desk that matters.

The step for this week is one export and one pivot. Take last edition's check-in timestamps for day one, bin them into quarter hours, and multiply each bin's rate by your measured service time to get the offered load. Put your actual roster for that day beside it. The bins where the roster sat below the offered load are the ones that produced every complaint you received, and they will be a smaller set than you expect. Everything downstream in your attendee analytics is easier once that comparison exists as a chart rather than as an argument.

Questions people ask about registration desk staffing model

Why is a flat registration roster wrong?
Because arrivals at a show are not flat. A roster of 11 agents can be 8 short at 09:15 and 7 too many at 15:00 on the same day. The morning shortfall builds a queue that never fully clears, and the afternoon surplus pays people to stand still. One number cannot serve both bins.
What size should the bins be in a staffing model?
Fifteen minutes works for registration. Hourly bins hide the shape of the opening surge entirely, since a single hour can contain both the peak and the recovery. Five minute bins produce so much noise from ordinary randomness that you end up smoothing them back to something like a quarter hour anyway.
How do you convert expected arrivals into a number of agents?
Multiply the arrival rate per minute by the mean service time in minutes to get the offered load in erlangs, then staff above that load with a safety margin. At 56.6 arrivals a minute and a 17.6 second service time the offered load is 16.6, and a five minute service target needs about 19 agents rather than 17.

Related reading

All on-site analytics articles