Building an at risk exhibitor call list reps will actually work through

Two weeks after the show closes, somebody exports the risk scores into a spreadsheet, sorts descending, colours the top rows red and sends it to the sales team with the subject line "renewal priorities". By the following Friday three reps are working their own accounts in their own order and the spreadsheet has been opened twice.

This happens every cycle and it is not a discipline problem. The list was sorted on the wrong thing, it was too long to finish, and it did not tell anybody what to say. Any one of those is enough to kill it.

An at risk exhibitor call list is a queue with a rule for where it ends. Get the ordering and the ending right and reps work it, because working it is easier than deciding for themselves.

Risk is the wrong sort key

Sorting on churn probability puts your most certain losses at the top. That feels correct and it is a good way to spend a fortnight talking to accounts that were never coming back.

Take four accounts from a real-shaped book.

Account A has a churn probability of 0.62 and spent 38,000 last edition. Expected revenue at risk is 0.62 times 38,000, which is 23,560.

Account B has a churn probability of 0.88 and spent 9,500. That is 0.88 times 9,500, or 8,360.

Account C sits at 0.41 with 71,000 of space and sponsorship. 0.41 times 71,000 is 29,110.

Account D sits at 0.79 with 12,400. That comes to 9,796.

Sorted on risk the order is B, D, A, C. Sorted on expected revenue at risk it is C, A, D, B, which is close to the reverse. The account carrying 29,110 of exposure sits fourth on the risk list, below two accounts that between them carry 18,156, and in a fortnight where a rep gets through sixty conversations the difference is most of the year's renewal revenue.

There is a second reason risk is the wrong key, and it is the one that took the field a while to accept. Eva Ascarza, in the Journal of Marketing Research in 2018, combined two field experiments with machine learning and found that the customers a model identifies as highest risk are not the best targets for a proactive retention programme. The paper's framing is that retention campaigns sometimes fail because the targeting rule is wrong and not because the incentive is wrong. If the account at 0.88 is going to leave whatever you do, the call is spent. The account at 0.41 might be genuinely undecided, and undecided is where a conversation has somewhere to go.

Testing that properly needs a holdout and an uplift model, which is its own piece of work and belongs with G24. Expected revenue at risk gets you most of the way there without an experiment, because value and risk are only weakly related in an exhibitor book and multiplying them moves the mid-risk, high-value accounts up where they belong.

What has to be on every row?

A ranked list of account names is a to-do list, and reps already have one of those. The row has to answer three questions before anyone picks up the phone: why is this account here, who am I calling, and what am I opening with.

Which means the row needs the account, the parent company if the exhibitor sits under one, the last edition's total value split between space and everything else, the score with its two or three largest negative contributions, the named contact and whether that contact has changed since the last edition, the date and outcome of the last logged conversation, and a single next action.

That is seven fields and it fits on a screen. The temptation is to add the other fifteen columns because they exist. Do not. Every column you add is a column a rep has to decide to ignore, and the decision costs more than the column is worth.

The reason codes matter more than the score. A rep who can see that the account paid 41 days late twice and cut its stand by a third has an opening. A rep who can see 38 has a number, and how that number was built and where its amber line falls are G18's and G19's problems rather than the rep's.

The stopping rule, and why the list has to be short

A list of 2,400 accounts is not a list. It is the book.

Whatever your capacity works out to, and there is a proper way to compute it that belongs with G23, the list has to end somewhere visible. I would put a hard line in the interface with the count above it, so a rep opening the queue sees "142 accounts, 96 remaining" and can tell whether they are winning.

Below the line, accounts do not disappear. They go into the automated side of the renewal programme: the renewal email sequence, the early-bird floor plan release, the portal reminder. What they do not get is a rep's fortnight. Pretending otherwise produces a list nobody finishes and a set of accounts that got neither the call nor the sequence because everyone assumed somebody was handling it.

Re-ranking mid-cycle is the other thing that kills a list. Scores update, the order changes, a rep comes back on Wednesday to an account that has moved forty places and cannot find where they were. Freeze the ranking when the cycle opens. Add new accounts at the bottom as a separate tranche if something material happens.

Does the churn model need to be better before you ship?

There is a temptation to keep tuning the churn model before shipping the list. Some of that is worth doing and the returns fall away quickly.

Neslin, Gupta, Kamakura, Lu and Mason ran a churn modelling tournament reported in the Journal of Marketing Research in 2006, where academics and practitioners downloaded the same data, built models and were scored on two validation sets. Two findings from that paper are useful here. Method choice genuinely mattered, with the spread across submissions large enough to change the profit of a retention campaign substantially. And the models held up well over time, losing very little accuracy when used to predict on a database compiled three months after the calibration data.

The second finding is the practical one for a renewal cycle. You do not need to rebuild the model between the show closing and the calls starting. Last cycle's model, applied to this cycle's data, is close enough to work the list with, and the time you would have spent retraining is better spent on the value field, which is usually wrong in ways that matter more than a two-point difference in area under the curve.

The value field is where these lists actually break

Expected revenue at risk is a product of two numbers and everyone concentrates on the probability. The value is the half that goes wrong.

Last edition value is not last edition invoiced revenue. An exhibitor on a three-year deal has a contract value that has nothing to do with this year's renewal decision, and multi-year agreements distort every retention figure you report, which is G38's subject. An exhibitor who took a 40 per cent partner discount is worth less than the list price on the contract. An exhibitor whose parent company also holds three stands on your other shows is worth more to lose than their line on this show suggests.

The fix is unglamorous. Define the value field once, write it down, and make it net revenue attributable to the renewal decision in front of you. For a single-show renewal that is space plus services plus sponsorship for that edition, net of discount, excluding anything contracted beyond the decision point. For a parent-level conversation it is the sum across shows, which needs the account view that G25 covers.

Then check it. Pull the twenty highest-value rows on your list and have someone who knows the accounts read them. In every book I have looked at, two or three of the top twenty are wrong by more than half, usually because of a duplicate account or a sponsorship booked against the wrong entity, and those errors are sitting at the top of the list where they do the most damage.

Where this approach stops

Expected revenue at risk assumes the call is worth the same to every account, and it is not. A conversation with an exhibitor who has already signed with a competing show is worth nothing. A conversation with an exhibitor who is annoyed about something you can fix in a week is worth a great deal.

Nothing in the ranking knows the difference, because the data that would tell you sits in an email nobody logged. Which is why the list should have a manual override that a sales director can use to promote fifteen accounts, and why those overrides should be recorded as overrides. If the promoted accounts renew at a better rate than the ranked ones, your director knows something the model does not and you should go and find out what field it is.

The second limit is that a good list makes the model look better than it is. Reps call the top of the list, those accounts renew, and the renewal rate among high-exposure accounts improves. You cannot tell from that whether the calls worked, and neither can anyone else, until you hold some of them back.

Take last cycle's call log, join it to the renewal outcomes, and count how many calls landed on accounts that renewed anyway and how many high-value accounts you lost without a single logged conversation. That second number is usually the one that changes how the next list gets built.