A club general manager looking at retention numbers at the end of a financial year is looking at the past. The members who left, left months ago. Most of them stopped using the courts twelve to sixteen weeks before they cancelled their direct debit. The membership team had no idea. The general manager, when she sees the churn number in the May board pack, also has no idea — until she goes back through the booking records and notices the pattern.
The pattern is unmistakable once you know where to look. Court usage cadence drops first. Frequency of partner-pairing changes second. Cancellation follows. The window between the first signal and the cancellation is wide enough that an intervention has time to work. The signal is sitting inside data the club already collects.
This article describes the signal, the model that surfaces it, and the operational changes that make the model worth building. It connects to the broader theme covered in federation talent pipelines — that the data layer for sport operations is often already collected but not legible. Clubs face the same problem federations face, on a smaller scale.
The two ratios that predict churn
The model output is, in operational terms, simple. Two ratios per member, computed weekly, plotted against the member’s own baseline.
Court usage rolling-30 / rolling-365. A member who booked twelve hours a month on average over the past year and is currently booking three hours a month is on a trajectory. The rolling-30 over rolling-365 ratio is the cleanest signal of an active member moving toward inactive. Below 0.4, the member’s behaviour has changed materially. Below 0.2, the cancellation is usually six to ten weeks away.
Partner-pairing diversity rolling-90. Members who churn often do so after their regular doubles partner stops playing. The partner-pairing diversity ratio — distinct partners played with in the past ninety days divided by distinct partners played with in the prior ninety days — drops sharply before the member’s own usage drops. The signal leads the usage signal by two to four weeks. This is the part most clubs miss completely.
The two ratios in combination, scored against the member’s own baseline rather than against a club average, surface the at-risk member earlier and with fewer false positives than any single-variable model. The maths is not difficult. The data infrastructure to compute it weekly is the work.
Why most clubs do not build this
Three reasons. The first is data integration. Court booking, billing, CRM, and the social audience system are typically four separate platforms with four separate exports and four separate ID schemes for the same member. The first job of any practical model is to reconcile these IDs. The reconciliation is grunt work, not modelling work, but it is the work that determines whether the model ever ships.
The second is operational. The model output is a weekly list of at-risk members. The list is useless if the membership team is not staffed to act on it. Most clubs are not. The membership manager has front-desk responsibilities, billing query responsibilities, marketing responsibilities, and the prospect of additionally working a weekly outreach list lands badly. Without an operational owner, the list goes unread.
The third is cultural. Some clubs, particularly traditional ones, prefer not to be seen to “track” members. The reservation is real and worth respecting. The framing that works: the model surfaces members for whom the club’s offer no longer fits. The outreach is a service conversation, not a sales conversation. The wording of the outreach script matters more than the technology underneath it.
What the outreach actually looks like
The most effective interventions, from operations that have run this work, are small. A note from the head coach saying “I noticed you haven’t booked the women’s morning group in a while — is there a time that suits better now?” The member who feels seen continues. The member who has genuinely moved on says so, the club books the cancellation in an orderly way, and the member often returns six to eighteen months later. The membership team does not pretend the data did not exist; the conversation is direct.
What the outreach is not: a marketing email. Marketing emails to at-risk members do worse than no contact. The signal is interpersonal. The intervention has to match.
SportBusiness has covered this kind of operational analytics in adjacent sports — golf clubs, fitness operators — and the pattern holds. Personal contact beats programmatic contact. The model surfaces who should be contacted; the conversation has to be human.
Vertical AI consulting and the build path
Building this model is not a SaaS purchase. It is a small bespoke project. The reason is that the model needs to be calibrated against your specific club’s booking patterns and your specific member base. A model trained on a country club in the Cotswolds will not generalise to a 600-court municipal operation in Madrid. The data shape is different and the baseline behaviour is different.
This is the kind of work that sits neatly inside vertical AI consulting: defined-scope, defined-data, defined-decision. The model is built in two weeks once the data integration is done. The operationalisation — building the membership team’s weekly workflow, drafting the outreach scripts, training the staff on what to do with the list — takes the remaining time. Most of the project is the operationalisation.
Sportico reporting on club retention work in the US suggests similar findings; the modelling is the smaller part of the work, and the operational integration is where projects either land or do not.
The signal you do not need
Some vendors will try to sell you sentiment analysis on member social media, demographic clustering, propensity-to-pay scores. None of these add meaningful accuracy over the two ratios above for member-churn prediction in a tennis club setting. The reason is that the two ratios above are themselves a behavioural signal — they are what the member is doing with her time, which is the leading indicator. Adding inputs that proxy intention is correlated with the behavioural signal but does not improve the model meaningfully. Save the budget for the operationalisation.
The MIT Sloan papers on consumer behaviour modelling point in the same direction: behavioural signals dominate stated-intention signals, in the sports context as elsewhere.
Sponsorship implications
A club that has built this model has, as a side effect, built an audience layer. The same data infrastructure that powers churn prediction can power audience segmentation for sponsorship and partnership conversations. A sponsor presented with “we have 2000 members, average age 47” has been given marketing language. A sponsor presented with “we have 320 active women members aged 25-40 who book on average 14 hours per month and who attended at least three on-site events last year” has been given a defensible audience that can be priced. The infrastructure is the same.
This dimension is covered in more detail in the broader operations and audience intelligence cluster, and the federation-side analogue appears in the data layer gap article.
What this means for your operation
If your club has a retention problem you want to address, start with the question that surfaces budget: what did churn cost you in the last financial year? Sum the lifetime value of the members who left in the prior twelve months. The number is usually large enough to justify the project several times over.
To scope what this would look like with your specific data, book a 60-minute call. Come with two artefacts: a sample export from your court booking system covering at least six months, and your current churn rate. We will tell you whether the data is computable, where the integration friction sits, and what the realistic timeline looks like.