What is machine learning?
Ask 10 computer scientists and you will get 10 different answers! But here’s our attempt: machine learning is a subfield of computer science that enables computers to act and make data-driven decisions and predictions, rather than being explicitly programmed to carry out a certain task. Machine learning programs are designed to ‘learn’ and improve over time when exposed to new data.
How do you generate the predictions?
Dataro uses machine learning to predict which donors will churn in the next three months.
If you were going to estimate which donors are likely to churn, you may already have some ideas about what is important: “younger people”, “people who have had a credit card declined recently”, or “people who have not been giving for very long”. But it would be quite hard to know exactly how important each of those ‘factors’ are and how to formulate them (i.e. is a ‘young’ person 18 or 25?). Additionally, even if you did have some idea of what factors were important for one charity, there might be entirely different factors at play at another.
Machine learning essentially is a way of taking all these factors (and hundreds more), then looking at the whole history of your fundraising, to discover exactly how important each one is and how they can be combined mathematically to produce a ‘model’ which predicts the likelihood of, for example, churn for every donor. This process can be used to train models for churn/attrition and many other fundraising campaigns: upgrades, cash giving, and reactivation, amongst others.
How does Dataro compare to us creating targeted lists based on our own lead indicators?
If your organisation has a history of running fundraising campaigns, we expect that the rules you have developed to select a cohort for upgrades and cash giving are sound and reasonable. However, there are a number of advantages to using a machine learning based approach:
instead considering a small set of factors (age, gender, last upgrade), machine learning can take into account hundreds of factors, discovering new and novel drivers for cohort selection that you may not have considered or that seem counter-intuitive.
Machine learning will by definition find the most ‘optimal’ solution to a problem given the data available. In some scenarios it is fairly clear what the optimal solution could be (someone is not going to upgrade their gift again if they upgraded yesterday), but in other cases it is hard to know where to begin (what makes one person a better acquisition than another). A machine learning based approach can be applied identically in all cases, opening up the possibility of new types of fundraising campaigns.
What factors do your models take into account?
There are literally hundreds of potentially relevant factors that may be taken into account – a lot more than just transaction history! However, here is a sample that may be relevant: number of transactions; last transaction date; transaction regularity; days since sign-up; Regular Giving status; recent card declines; missed payments; petitions signed; email clicks; phone call history; website logins; onboarding channel; age; gender; postcode; time of year (e.g school holidays or tax time); days until Christmas; consumer confidence; house or apartment; charitable giving amount*; marital status*; renting or Home owner*; Profession*. *Factors may be estimated from 3rd party data enrichments, such as aggregated data from public sources such as ABS.
How do you assess model accuracy?
Accurate forecasts are incredibly important to us – that’s what our business is built on! We assess and ‘sense-check’ our predictions in two ways: 1) We always test our predictions against historical data. This means that every model is tested on data that it has not ever seen before, which allows us to create a reliable estimate of future performance. By evaluating models using a held out data set and cross validation, it is possible to have a very good idea of how good the model is going to be in practice. Some events are inherently random, but machine learning allows you predict and plan with a greater degree of assurance. 2) We frequently run experiments to evaluate model performance. This involves leaving a ‘control set’ that is not exposed to any sort of campaign or intervention, allowing us to see how the predictions perform in a controlled, scientific experiment.
How often should the model be updated?
We recommend updating models at least every month, or just prior to a specific campaign. In general, while a model can stay accurate for some time, the most important thing is fresh predictions as donors and conditions are constantly changing. For the best results, we recommend running fresh data for every campaign.
How do I upload my data?
It’s simple! Create an account and use our drag and drop framework to import your formatted data direct into our system. If you have any trouble with formatting, our technicians are on hand to troubleshoot and get you started on your way to smarter fundraising!