Increasing charitable donations for a not-for-profit
We worked with an NGO to optimise their email marketing efforts. We identified the probability of donation for each of the members, allowing the NGO to focus their time more efficiently on members with high probabilities.
A not-for-profit political-activism organisation.
The organisation seeks to create a more progressive, fairer, better society on behalf of its members by helping them to engage in political advocacy on the issues they care about. They engage with their members primarily through email campaigns, and have societal impact mainly via the signing of petitions.
Funding is provided solely by small one-off and regular donations from members, therefore the organisation’s ability to be influential in society is strongly dependent on the donations they receive. Regular donations by Direct Debit in particular provide guaranteed income.
Currently, the organisation takes a rules based approach to decide who to contact to set up a regular donation, resulting in large email lists with low conversion rates. Faculty were tasked with testing the hypothesis that data science can be used to increase the number of members who make regular donations by better tailoring of email campaigns.
We first ingested nine years of anonymised historical donation data, including both one-off and regular donations. Statistical analysis indicated people who give both one-off and regular donations throughout their engagement with the organisation have a higher lifetime value than those who give regular donations only. Therefore, there was the most value in identifying potential regular donors amongst members who already have a history of donating to the organisation.
Distinguishing between members who are likely to set up a regular donation and those that are not is a classification problem, suitable for several machine learning algorithms. We used an XGBoost classifier as it is a powerful technique capable of handling mixed feature types and large structured datasets.
The model was fed donation history, along with data on actions taken by members, such as signing petitions or completing polls, and predicted the probability that they will set up a regular donation within a three month time window. The model outputs a probability per member which allows the organisation to use the results in a flexible manner.
The model was integrated into the organisation’s existing infrastructure and application landscape. The model outputs can be used to allow more efficient use of email lists by not sending donation focussed email campaigns to members with low probabilities. In addition, the organisation is able to test more targeted approaches on members with a high probabilities.
When compared to the rules based approach on a given representative day, the model would have allowed the organisation to email 19% less of their members whilst finding 21% more of the members who go on to set up a regular donation.