What is predictive modelling?
A typical predictive model takes data about known behaviours and characteristics of your customers at a given point in time. It then uses this to predict a probability for each individual customer who exhibits a behaviour you’re interested in at a given point in the future.
This process can be applied in a variety of ways within marketing, including predicting the likelihood of:
- A customer making a purchase or making a donation in the future
- Someone converting from a prospect to a customer
- An individual reacting to a specific offer or marketing approach
- Somebody upgrading or changing product or service
- A customer renewing a policy early
- Someone lapsing and no longer being your customer (allowing for churn prevention action)
- Being able to reactivate a lapsed customer
- Self-serving and/or switching channels
- Models can also be used as a dimension for segmentation alongside other customer differentiators such as RFM (Recency, Frequency, and Monetary value), ARPU (Average Revenue Per Unit/User) and geo-demographic characteristics.
As well as determining when a customer is going to purchase or leave, and enabling better media selection choices and targeting, predictive models can inform variable content such as imagery and copy, depending on whether a customer is more or less likely to exhibit the behaviour that’s of interest to you.
For example, you may wish to isolate those most likely to convert, those who are on the cusp of converting, and those least likely to convert. You could then treat them differently in terms of messaging, style of email or channel of communication (i.e., DM vs EM). You could ignore those customers whose behaviour is furthest away from your desired goals to not waste marketing time and effort.
What do you need to build a predictive model?
Exactly what kind of information is required for a predictive model will depend on the nature of the model and your precise objectives.
Generally, the more data that can be gathered over a long period of time, the more likely it is that a suitably predictive model can be built.
It goes without saying that understanding and interpreting the models will add far more value if your data has been explored in detail beforehand. Any predictive model scoping should therefore include a data discovery phase in the early stages.
First and foremost, you need to identify the outcome you’re seeking to predict. This is known as the target variable. For predictive models, this tends to be a binary field (only takes the values 1 or 0), i.e., something does or doesn’t happen - although it is also possible to build generalised multi-category models where the target variable can take several values.
This field will often need to be defined from your available data, e.g., a customer is a purchaser if their total spend in the period of interest is greater than 0, or if they purchase at least 1 item in that period.
You can then start to look at the data you’ll use to try to predict an outcome – the explanatory variables.
The kind of data that will usually figure prominently in a predictive model includes:
- Customer transactional behaviour - both current and historical (going back a number of years is ideal)
- Marketing communications activity and levels of engagement with marketing communications sent out, as well as website visits, other response channel data, and conversion data (e.g., sales, donations, behaviour change)
- Demographics - age, gender, and geographic location (often overlaying a demographic tool will enhance a model) Subscription history (if applicable)
- The time of year, accounting for seasonal variations in customer behaviour
- Other indirect factors such as the state of the economy and competitor activity are likely to have a place in a model and need consideration.
For instance, DCX recently built a predictive model for a world-famous coffee product company. The model took in a large number of data feeds, but not all of them turned out to be predictive. So, we were able to implement the model using a reduced number of data sources, making it easier to manage in the future.
How do you build a predictive model?
Well, that’s the complicated bit, and it will vary from organisation to organisation, but basically there are four main principles to follow - all things we can help your brand to deliver:
1. Define scope
- Define objectives and scope in line with agreed budget
- Audit available data and data flows, and update frequency
- Explore additional data sources and data enrichment opportunities
- Consider analysis approach / modelling methodologies
2. Build model
- Carry out data preparation and exploratory analysis
- Define the test and validation data sets
- Build an initial model to assess viability
- Iterate, optimise, and validate the model
3. Implement (in agency or client environment)
- Provide model scorecard/code
- Write up and supply any supporting documentation
4. Evaluate and refine (as required)
- Measure model performance at regular intervals
- Review and audit at agreed intervals, e.g., annually, refining the model where necessary as customer behaviour changes over time
- As well as auditing the data, it’s important to understand the business context behind the data to ensure whoever is running the analysis correctly identifies and clearly defines the target variable that will be predicted.
All available predictor variables will need to be explored to understand their high-level impact on predictive power via a characteristic analysis. Proxies may also need to be created where some data isn’t available, and possibly compensate for gaps in the data and exclude and explain outliers. All this needs to be done before the models are built.
Defining samples of the population to build and validate the model on is vital, ensuring that the model can be tested on different data to that upon which it was built. It may be necessary to generate stratified samples if positive outcomes are rare in the population.
Throughout the build it’s crucial to check that the algorithm converges (if not then the sample size may need increasing until it does) and evaluate the model performance – for instance a gains chart is a good way of visualising this.
How DCX can help
Predictive modelling may sound complex, and there’s a lot behind it – but it can help your organisation understand your customers better, keep your customers from leaving, save you money, drive more conversions, and increase revenue.
Imagine knowing when a customer might make a purchase in the future and being able to do something about it! How powerful would it be to know when a customer might leave you?
Predictive modelling can give you all this and more.
We’re helping a world-famous coffee product company do this right now. And we’ve helped numerous organisations all around the world gather data, make sense of it all, deliver insight, and improve their customer experience.
If you think predictive modelling could be something your organisation would benefit from, or if you’d just like to chat to explore what might be possible, just get in touch. We’d love to hear from you.