Subscription Renewal Prediction for Membership Businesses

Annual renewals behave differently from monthly churn

A monthly software subscription can be cancelled any day, so churn is a continuous risk measured week to week. A professional association, a gym on annual contracts, a heritage trust, a golf club or a trade body works differently. For eleven months the member is a member. Then, at one moment, they renew or they do not.

That single decision point changes the modelling and, more importantly, the timing of the response. By the time the renewal invoice lands, most members have already made up their minds. The window to influence them is months earlier, when disengagement first shows up in the data. Our post on churn prediction models covers the continuous version; this one is about renewal-date businesses.

Engagement signals that predict renewal

• Benefit usage. Visits, bookings, course enrolments, magazine opens, member discount use. The single strongest group of signals.
• Trend, not totals. A member who visited weekly and now visits monthly is at more risk than one who always visited monthly.
• Payment method. Direct debit or auto-renew members lapse far less than those paying by invoice.
• Tenure. First-year members are usually the riskiest; long-standing members lapse for life reasons.
• How they joined. Heavily discounted joining offers or gift memberships tend to renew less.
• Service interactions. Complaints, unresolved queries, a cancelled event they had booked.
• Email engagement. Weaker than it used to be because of privacy features, but prolonged silence still counts.

For a gym the dominant signal is visits. For a professional body it might be CPD tracking, event attendance and use of the member resources library. The model framework is the same; the features differ by business.

How far ahead to predict

We usually build models scored at several points, so a member's risk is visible early and updated as renewal approaches. A member whose risk rose between 90 and 45 days is a different case from one who was always high risk.

Acting on renewal risk without discounting everyone

The instinct is to offer at-risk members a renewal discount. Resist it as the first move. Discounts teach members to wait for the offer, and many at-risk members are not price-sensitive at all; they are simply not using what they pay for.

1. Re-engagement first: personal invitations to events or classes matched to what they used to use
2. A benefits reminder showing the value they received this year, and what they have not yet tried
3. A call from a real person for high-value or long-tenure members
4. Moving invoice payers to auto-renew or direct debit where they are happy to
5. Offering a cheaper tier or pause option before they leave entirely
6. Discounts only for members where testing shows it changes the outcome

That last point is uplift thinking: some members will renew anyway and some will not whatever you offer. Our post on uplift modelling explains how to find the ones in between.

An illustrative example

Imagine a regional professional association with 8,000 members, annual renewals spread through the year and first-year members renewing noticeably less often than established ones. The membership team currently sends the same three reminder emails to everyone.

A model scored 90 days before each renewal might show that first-year members who attended no events and did not log their CPD are the least likely to renew, while long-standing members paying by invoice are a second at-risk group for purely administrative reasons. The first group needs an invitation and a welcome call; the second needs a direct debit form. Neither needs a discount. The numbers are illustrative, but distinguishing disengaged members from merely inconvenienced ones is where most of the value comes from.

When renewal prediction is not worth it

Small clubs with a few hundred members already know who is drifting. A committee member's phone call beats a model. If the organisation cannot record engagement at all, because visits, logins or attendance are not tracked, a model will lean on payment method and tenure and add little beyond a simple report.

And if renewal rates are already very high, effort may be better spent on acquisition. It is worth asking which problem is bigger before commissioning anything. Our post on AI for membership organisations covers other places to start.

What building it involves

Most of the work at SpiderHunts goes into joining engagement data from the membership system, booking tools, email platform and door access or login logs, so every member has a consistent history. The model itself is a standard classifier scored on a schedule, with risk and reasons pushed into the CRM and a simple renewal-risk view for the membership team. Our machine learning services usually deliver a first version in five to eight weeks, and the first renewal cycle after launch doubles as the test, with a control group kept on the old process.

Frequently asked questions