Machine Learning for Charities and Fundraising

Retention is where the money is

A national charity with 60,000 active supporters spends a great deal to recruit a regular donor, whether through street fundraising, television or paid social. Many of those donors cancel within the first year or two. Every early cancellation means the recruitment cost was never recovered.

That is why donor retention is usually the best first use of machine learning for charities. Predicting which regular givers are likely to lapse, while there is still time to thank them properly, show them their impact or offer a lower amount rather than a cancellation, protects income the charity has already paid to acquire.

Where machine learning helps fundraising

• Lapse prediction. Scoring regular donors by the likelihood of cancelling in the next few months, using tenure, recruitment channel, payment failures, contact history and engagement.
• Conversion to regular giving. Identifying one-off and emergency appeal donors most likely to respond to a regular giving ask.
• Legacy propensity. Finding long-standing supporters who may be open to a conversation about gifts in wills, a slow and very valuable income stream.
• Appeal response and ask amounts. Predicting who will respond to a mailing and suggesting a sensible ask level based on past giving.
• Gift Aid and data quality. Flagging likely duplicate records and missing declarations, which costs charities real money.
• Income forecasting. Projecting regular giving income by month, allowing for attrition, which finance teams need for planning.

A worked example on lapse

Take a charity with 20,000 regular donors giving an average of ten pounds a month. Suppose, purely as an illustration, that around a fifth lapse each year. That is 4,000 donors and roughly 480,000 pounds of annual income walking out.

A lapse model does not stop that. What it does is rank donors so the supporter care team can call the highest-risk few hundred each month, send a well-timed impact update to the next tier and leave everyone else alone. If that saves even a modest share of the at-risk donors, the income retained is measurable within a year, as long as the charity keeps a control group that receives normal communications. Without that comparison, nobody will know whether the calls made the difference.

The kindest retention tactic is also usually the most effective one: thank people properly before they have decided to leave.

Profiling donors and the law

Charities have particular reason to be careful here. In 2016 and 2017 the Information Commissioner's Office fined a number of well-known UK charities, partly for screening donors' wealth and profiling them without telling them. The Fundraising Regulator's Code of Fundraising Practice also sets expectations on treating donors fairly.

None of that makes modelling off limits. It does mean the privacy notice must explain clearly that the charity analyses supporter data to decide how to communicate, the lawful basis must be sound, and people must be able to object.

1. Update the privacy notice before building the model, in plain language
2. Complete a legitimate interests assessment or data protection impact assessment
3. Avoid buying in wealth data unless supporters have been properly told
4. Honour objections to profiling across every system, including email and mailing house lists
5. Take particular care with supporters who may be vulnerable

Vulnerable supporters must be protected, not targeted

Some of the most loyal and generous supporters are elderly, and some are vulnerable. A model optimising response rates will happily find people who say yes to every ask, and those are exactly the people a charity has a duty to protect.

We build in the opposite logic: contact frequency caps, suppression of anyone flagged by supporter care as potentially vulnerable, and a rule that legacy propensity scores only lead to a gentle, informative message, never pressure. A charity's reputation rests on donors trusting it, and one story about aggressive fundraising undoes years of work.

When a smaller charity should not bother

Recency, frequency and value segmentation is old-fashioned and still very effective. Most CRM platforms can do it, and it covers much of what a first model would find. Our piece on AI integration for charities covers the language model side, such as drafting thank-you letters and summarising supporter notes.

The CRM is usually the first project

Donor databases accumulate duplicates, inconsistent campaign codes and gaps from migrations over the years. A lapse model trained on that will learn the quirks of the data rather than supporter behaviour. The same person recorded three times looks like three low-value donors.

Deduplication, consistent campaign and channel coding and a clean history of payment failures are dull, and they make everything downstream possible. They also usually improve Gift Aid claims and mailing costs straight away, which helps the case for the modelling.

How we would work with a charity

At SpiderHunts we would begin with an export of several years of giving history and a conversation with the supporter care team about what they could realistically do with a weekly list. The model is then built to fit that capacity, not the other way round. We are also happy to tell a charity that segmentation in its existing CRM is enough, which is often the honest answer.

Where modelling is justified, it sits within our data science service, with the scores pushed back into the CRM so fundraisers use them without learning a new tool.

Frequently asked questions