Building sustainable relationships with donors makes fundraising more effective. That’s why it’s important to examine your donor journey and identify opportunities for improvement.
Improve the donor journey? What do I mean by that? I mean two things:
- Improve your donors’ experience with your organization as they move (or jump) up your giving ladder.
- Manage your donors’ journeys so well that your fundraising brings in more in gift revenue in a shorter time and on the same budget.
Machine learning can take care of both needs since its job is to bring out patterns of behavior, like:
- Behavior chains that point out constituents whose engagement is increasing and those who are getting lost in the shuffle;
- Unnoticed segments of your organization’s fan base so that you can follow up with tailored cultivation.
Machine Learning vs Traditional Donor Modeling
Machine learning is different from donor modeling, which scores everyone with the same equation leaving you with a huge, sorted list. Instead, machine learning identifies groups of data points that describe your actual audience segments. Marketers call these groups, “personas”. Since we are used to naming our constituents’ various personas through our culture’s paradigms (as in, “Our alumni don’t give. They are mostly teachers.”), finding new ways of looking at our audience affords fresh engagement, cultivation, and solicitation opportunities.
How Machine Learning Enhances the Journey
A common machine learning technique is random forest: Your audience segments are described as belonging in different trees with different branches of behaviors/indicators, all the way out to specific leaves with very specific audience characteristics. The leaves may have only 3 records, which brings up the worry around overfitting, a of random forest models. However, the sometimes hundreds of trees do split your audience into cohesive groups of prospects with very similar characteristics.
Machine Learning in Action
As an example, consider the phrase, “people of color.” Though it’s used as a polite term, it also lumps together people whose ancestry comes from Africa, Asia, and Latin America. In other words, it’s an exclusionary term rather than a segmenting term (their ancestors are not from Europe).
Project that concept back into your own database. Women in your database are NOT men, the class of 1995 will NOT feel the same way about its 30th class reunion that the class of 1975 did. Or, from a recent client example, people who buy balcony tickets do NOT have the same motivations toward going to the theater as people who buy tickets in the first few rows.
Here is an example of a decision tree from a recent project (random forest is actually a number of decision trees but there are algorithms that will create just one). The client is a high-school level service program providing participants with skills that are much needed in the marketplace. Here is the cluster of characteristics of one of the most interesting segments as derived from over 20 years worth of surveys:
- They indicated that they did not get the mentoring that they needed from the industry volunteers,
- They also indicated that they did not get offered internships very often,
- And they also indicated more often than not that they would be willing to mentor current students.
As you read this, are you guessing that this cohort also selected African American as their ethnicity? I was not surprised. But this cohort – should the organization take their offer to mentor students — would be key to both bringing in more students into the program (the “that guy looks like me so I’ll join” moment) and to provide mentorships and internships that they did not receive.
New insights like this can help any program, as shown by other client results, sampled below:
- The project to find which events to stop holding resulted in showing that any event type that followed reunion was worthwhile.
- For the entire prospect pool, one visit by the college dean was enough to solicit a $1 million gift, except for one specific group.
- A segment from the annual giving pool – renewing donors – needed to be split between those who went to the annual gala vs. those who did not to boost direct mail results. Our work with this client showed the ability to raise an additional $300,000 per year more (15% of their annual yield) using our proprietary optimization tool.
- Identification of three specific characteristics allowed the client to bring in new, multi-million dollar planned gifts.
- The client was sending 25 or more emails each year while only 5 were really needed to solicit the audience well.
How does machine learning work?
The algorithm sifts all the data back and forth, like sand sifting from bucket to bucket, until the data sets that fit best together land in their own buckets, and then the algorithm finds the next best characteristic to sift and sifts that – and then on and on until no more value can be squeezed from the data. The specific math for different machine learning methods can be found online, but that is not the most interesting part. The best part is finding an unknown segment and then building a plan.
Ready to Learn More?
If your analytics team is curious about getting started in machine learning, take a look at our brand new course: https://staupell-analytics-group-online-workshops.teachable.com/p/machine-learning-basics
Or write to me if you have thoughts or questions.