Ask Amount recommendations are based on a large-scale analysis of ask ladders and outcomes. Channel Recommendations predict which channel a contact will contribute the most value to.
Ask Amounts
What is an Ask Ladder?
Typically, a solicitation for one-off donation will prompt the recipient with multiple options for a donation amount. These options are known as an 'Ask Ladder'. A three-step ladder will consist of an 'anchor' amount and two higher amounts. A typical approach to the ladder is: Anchor, Anchor * 1.5, Anchor * 2.0.
For example:
- $20 (anchor)
- $30 (anchor * 1.5)
- $40 (anchor * 2.0)
As you can see, the anchor is the critical value in setting up the Ask Ladder. It makes sense to set a higher anchor for certain donors and a lower value for others. Dataro's Ask Amount Recommendation provides a rigorous way to set a unique anchor value for every contact in your CRM.
How does the Ask Amount recommendation model work?
In an ideal world, we could pose an ML problem like so:
Based on a donor's history, which anchor value will result in the highest possible donation value?
However, in our analysis we have determined that historical Ask ladder information has not been recorded with sufficient fidelity and in adequate volume to train a model to solve this specific problem.
Instead, we have analysed several dozen different ask strategies and the giving behaviour of millions of donors to develop what we believe is a highly effective method to estimate the best the Ask Amount.
The model takes into account a donor's recent single giving behaviour and finds a good balance between the average and the maximum donation. The model will generally ask for a reasonable increase in donation value, but is more conservative in cases where the donor has given less, given less frequently or if they have not donated a single gift recently.
If a donor has only made recurring donations in the past, then the value is based on their average RG gift.
If the donor has never donated before, then the Ask Amount is based on the organisations median value of recent, successful, first-time single gifts.
How does Dataro recommend storing Ask ladder data?
Ask ladder data should be recorded against each constituents solicitation record for each campaign. This will facilitate more advanced analysis of ask recommendations in future.
This will look different depending on your CRM.
In Salesforce, you should create custom fields on the CampaignMember object and store the values there when a Contact is assigned to a campaign. Please the below API names and types:
Ask_Amount_1__c Currency
Ask_Amount_2__c Currency
Ask_Amount_3__c Currency
In Blackbaud, you can use the comments field on the Assigned Appeal record with the following structure:
Ask1:50.00,Ask2:75.00,Ask3:100.00
Channel Recommendations
Fundraisers know that some donors have a preference for a particular channel (eg Mail, Email, SMS, or Phone) and that when asked on that preferred channel they are more likely to receive a response or a gift. Some channel preferences are obvious (eg if you do not store a phone number, you can’t reach out via phone/sms), but some are more difficult (a donor who has given to an email and a mail outreach, which is best). Fundraisers lack the ability to analyse all this data and decide which channel is best for each and every donor.
The goal of this new propensity is to provide fundraisers with an easy way to see which channel a donor may give to so they can tailor their communications. When combined with Dataro’s other propensity models, fundraisers know who to reach out to and which channel to use.
Channel Recommendation can be structured as a more typical ML problem. However, there are two novel aspects here compared to Dataro's other propensity models.
Multiclass Classification
The first novel aspect of this prediction is that it is a "categorical" prediction instead of a binary prediction (i.e. we are predicting Mail / Email / Phone / SMS vs. True/False). In the ML world this is called a ‘multiclass’ problem. The output of the model is a probability distribution across the four potential classes. For each contact we are publishing the class with the maximum probability to the CRM.
Defining the target
The second tricky part is how to define the 'target' i.e. what we are training the model to predict. We have structured it as follows:
- Which channel has a donor donated the most amount of money to in the next twelve months? Perhaps the donor has donated $10 6 times via SMS, but given a $75 mail gift… the model is going to learn from the training that this donor is a Mail donor.
- There is an argument to be made that the number of gifts is more important than the total amount of gifts, however, our analysis and customer research indicated that using the total amount as the target results in a model which will be most helpful for fundraisers.
- If a contact in the training set doesn't make any donations in the next 12 months from the reference date, then we will look at which communication channel they have the highest number of responses to.
Model Factors
The factors for the model are relatively similar to the other propensity models, however, we are also including:
- Binary factors about contactability: has_address, has_mail, has_phone. How do we get these when we don’t capture PII? For NXT / SF we are hashing the data before it is saved to disk, so we never save (and it's impossible to un-hash to recover) the raw data.