Ask your donors for the right amount
How non-profits can maximize small-dollar fundraising using machine learning
Picture this: you open your mailbox and see an envelope from the local 501(c)(3) you gave to last week. “$5, $10, $20, any amount helps” it reads. It has boxes to allow you to select your giving level. It probably offers tiers with benefits that accumulate as you increase your donation.
“$2.70 [] $5.00 [] $25 [] $100 []”: an email with checkboxes from the candidate you gave to just this morning. It asks you to make recurring donation until the election.
When a potential donor receives the mailed tiers or the email checkboxes, she is supposed to have the same reaction to it that she has seeing the wine menu at a restaurant: the urge to impress — to give big, but not too big.
However, this approach suffers from two main faults: it is (1) not personalized to the individual donor and (2) only caters to big-dollar donors with its ever escalating tiers.
Any analyst looking at the donations from previous drives can pretty quickly create this menu of options. Many organizations will already have done so and are trying to fix the above deficiencies. To try to overcome the first fault — the lack of personalization — the fundraising department can cultivate relations with donors. People with relationships to the very top tier of donors know their preferences inside and out from years of conversations. Those donors are asked for the right-sized donation for their situation via individualized contact.
Obviously this relationship strategy cannot scale to a list of thousands of donors. It allows us to overcome fault 1 for a few donors, but not for the rest. It also further reinforces the second fault, with small-dollar donors not giving to their full potential. We need to find a strategy that will be scalable over a large list of potential donors.
Enter the machine learning model. Using machine learning we can overcome both the faults of the check box strategy for the vast majority of our donors. In this post we are going to explore how the general approach above leaves money on the table, and, how using machine learning we can create a personalized ask that rivals a peer-to-peer relationship.
The Ask — sorry for the corporate speak
What is an “ask”? The ask is the amount of money requested in the content. An email might have the subject line: “Can you make a contribution of $5 today?” The ask in that subject is $5. A personalized ask is when the ask changes based on who is seeing it.
Before we get to personalization and machine learning, we first have to understand the properties of using the same ask for every piece of content.
Asking everyone for the same amount
The first trade-off to consider as we determine the ask: a higher ask leads to fewer donations. This is analogous to the demand curve from Econ 101. Non-profits have an advantage over for-profits in this domain though: people can give however much they want — regardless if it is much smaller or larger than the ask. Still, my experience working on one of the largest daily email programs ever — the Bernie Sanders campaign was sending out emails to millions of people a day — bears out the truth of the graph below. The higher the ask, the fewer donations we get.
The second effect to consider when setting the ask is that the more we ask for, the higher the average donation will be. This means that by creating a higher average donation we may actually maximize fundraising receiving while fewer donations.
In the example left, fundraising is maximized at near the highest levels of asks shown. Referencing the previous figure, at a $15 ask we have about half the number donations that we would have gotten at an $8 ask. In this example, we can give up half of our donations and still maximize fundraising.
But anyone familiar with non-profit fundraising sees the fault in this plan: people dissuaded from donating this drive are less likely to donate during the next drive. By asking for too much money, we risk eroding our base of donors. By raising the ask, we can sacrifice a large number of donations to maximize our fundraising for this year only. We will come back to solutions to this problem as we delve into machine learning.
One final effect to take into account is the most common donation will be the amount we ask for. Choosing a symbolic or otherwise meaningful number can magnify this effect.
For any non-profit with a data on historical donations, we can likely use the data to recreate this curve and understand the tradeoff. But even if we show we can increase donations significantly in year 1, we risk the future of the organization by blindly asking everyone for more money.
We need to find an approach that balances today’s fundraising with tomorrow’s viability. It is not enough to increase fundraising on this drive, we need to sacrifice as few of today’s donations as possible while achieving a higher average donation. This is where machine learning comes in to shift the curve by introducing personalization.
Personalizing the Ask — The Machine Learning Approach
In my experience as the fundraising data scientist for one of the largest email donation programs ever, we can use machine learning to both increase the number of donations and the average donation. By learning individuals preferences and asking people for what they are the most likely to give we can keep our donors giving and maximize the amount they give.
To understand how personalization works, imagine two hypothetical donors. The first gave $100 last year and the second gave $5. Neither of these donors have given enough for a non-profit to feasibly reach out to them individually.
Let’s run through a few scenarios:
- We ask both donors for $5. In this case the most likely outcome is that both give $5. We end up with $10.
- We ask both donors for $50. In this case the most likely outcome is the small donor does not donate and the large donor donates $50. We end up with $50.
- We ask each donor for what they gave last time. In this case we are likely to get $105, the most out of any scenario.
From this example we can see how the personalized ask could increase total donations. So why don’t we just ask everyone for what they gave previously?
There are two issues: (1) people who have never donated before need an ask, and (2) the previous donation was specific to the context in which it was given. Because we are changing the context of the next donation ask, we can do better than simply repeating what worked in a previous context.
In the next section we will go over how a machine learning model solves all the challenges we have faced so far.
Shifting the curve
We create a machine learning model by feeding demographic and historical giving data to a predictive algorithm. The algorithm uses the demographics and giving behavior to create a model of individual giving preferences. Using this model, we can predict much each person is likely to give for their donation to a specific drive. Even people who are new to the list can be imputed from similar donors.
The figure below shows the results expected from this model, a higher donation rate at all levels of average ask. The entire curve has shifted up: more people donate across the board.
Note: Because we are personalizing the ask to each individual, there is no one level for the ask like we had above. I’ll be using the average ask to indicate level when talking about machine learning.
Now that we have more donations at every level, we can tune our average ask to maximize our fundraising. We are still bound by the same realities as before: the more money we ask for, the fewer donations we will get. But, this time, we will maximize our fundraising at a higher donation rate than the first example while still increasing the average donation.
The figure to the left shows the trade off curve for both a modeled ask and a one-size-fits-all approach. For this example that modeled ask will maximize fundraising at a $13 average ask. That would correspond to a ~32% donation rate. The traditional approach would maximize revenue at a $12 ask and ~23% donation rate. To achieve the 32% donation rate, the traditional approach would have to lower the ask by half down to $6. Machine learning has allowed us to more than double the ask without sacrificing any donations.
One final effect to mention is an increase in lifetime value, especially for new donors. By increasing the average donation at the start of a donor’s giving lifecycle, you can greatly increase their value in the long term. People often continue giving at the level of their first donation — or slightly below that — so getting the optimal first donation is crucial.
Closing Thoughts
Machine learning personalizes the amount you ask of small-dollar donors making them more likely to give at every level of ask. Using machine learning, non-profits can maximize their fundraising while overcoming the problems with scaling traditional approaches. Machine learning allows an organization to:
- Personalize the ask to each donor
- Scale individualized attention to small-dollar donors
- Create an ask specific to a piece of content or drive—every time
- Increase the number and lifetime value of new donors
- Increase or maintain its donation rate so you secure the future
Anyone with data on historical donations from their small-dollar donors has the ability to increase their fundraising using machine learning.
