What is Predictive Segmentation And Why Does It Matter
By Danny Shayman
Summary: Predictive segmentation is a modern approach to grouping customers by their propensity to take action, such as the probability that they will visit your store in the next 30 days, churn from your subscription service, or be a better fit to be upsold product A versus product B. This is much more valuable than previous segmentation approaches because it enables marketers to better understand each customer and how to best message and market to them.
Descriptive segmentation is the clustering approach that has been used by marketers for decades in which customers are divided manually into groups by surface-level differences between them. Examples of these would be Moms Between 30-40, Loyalty Card Holders, or simply Sports Enthusiasts. They are called descriptive segments because customers are assigned their grouping based on somewhat surface-level criteria decided upon by marketers.
The more modern method is called predictive segmentation.
What is it?
Leveraging machine learning, customer data is analyzed by self-optimizing algorithms to understand how dataset attributes (and combinations of attributes) correlate to specific business objectives. Customers are then automatically placed into segments based on their likelihood or propensity to exhibit a future behavior. Examples of these would be Likely to be a Repeat Visitor, Unlikely to Churn, or Predicted to Double their Annual Spend.
As you can see by the examples, the main difference between descriptive segmentation and predictive segmentation is whether or not the clustering is focused on predicting a future action or not. If it is, it’s predictive segmentation. If it isn’t…well, why are you still using descriptive segmentation?
A more useful way to think about predictive segmentation is that it answers questions that marketers have about their customers which can help them better market to them.
Predictive segmentation: a marketer’s crystal ball
What if instead of having descriptive segments by age/gender (e.g. Men 18-34, Women 50-60), you had a segmentation approach that could predict…
- …who would or wouldn’t come to your restaurant more often if they knew about your happy hour specials
- …who might or might not click on your latest social advertising campaign
- …which existing customers would or wouldn’t be more apt to buy your latest software upgrade
- …who is or who isn’t more likely to buy a certain product of yours on Amazon
Wouldn’t that be more powerful and useful to you than simply Men 18-34?
After all, men between 18-34 vary from man to man. Some might be your best customers. Some may have never bought from you at all. Some may have heard of you but not purchased. Some may have heard of you but won’t ever purchase. Why segment your customers by insignificant [descriptive] variables when you could segment them by significant [predictive] ones?
When armed with this information, you can avoid wasting your marketing budget on Men 18-34 that the data is clearly showing aren’t interested in your product/service instead of spending it on the Men 18-34 that the data shows they are most likely to convert into a customer.
With predictive segmentation, knowing, for example, that someone might cancel your monthly subscription service would make them better suited for an email campaign offering a new, lower price versus a loyalist who is exhibiting signals that they will never churn. Don’t offer a discount to a loyal fan who is fine paying your current price. The discount should go to the customer who is predicted to most likely churn. You might also be willing to pay a higher bid price for them across your biddable, digital advertising. You certainly would use a different messaging approach with someone who is predicted to churn versus a customer is likely won’t.
What other things can you predict with predictive segmentation? Over the years, we’ve predicted for our customers things like:
- Customer Lifetime Value
- Being a customer / visitor / guest
- Purchasing a specific product
- SKU level sales volume
- New location sales volume
- Product mix
- Upgrading / downgrading service
- Detecting fraudulent transactions
- International shipping tariffs
As long as you have the customer data available, simMachines can predict each of your customers’ likelihood to take (or not take) that action.
How does it work? A predictive segmentation walk-through for “likelihood to churn” in 4 steps
I’ve explained predictive vs descriptive segmentation to many marketers over the last five years at simMachines. Most of the time, the value is understood immediately. Of course it makes sense to understand each customer’s likelihood to do something in the future.
So, the next question is always “how does it work?”
Let me show you. As you will see, our self-service platform makes this process easy for any marketer with a little training. With simMachines, you don’t need to be a data scientist to run sophisticated segmentation models.
Step 1: Analyze the data
We pull historical customer records with known outcomes (churn / no churn) into our similarity engine, which learns which data interactions are important for predicting the behavior.
In this example, the telecom marketer first links their 1st party data (customer plan info, usage, service quality, customer service calls, etc.,) with commercial 3rd party commercial demographics data which results in a massive and more robust dataset for simMachines to analyze
Step 2: Each customer is scored
Every record/customer is scored and the driving predictive factors learned by the engine are displayed.
Each customer’s unique data receives a unique prediction. In some cases dozens of data points are used, like spotting heavy users who are growing frustrated with their bandwidth limitations and inconsistent service quality. For other predictions a small selection of variables may be all that’s required, like demographics data providing a recent move flag and an updated billing address to find customers who just moved outside your service area.
Step 3: Intelligent machines cluster customers by prediction
We then use machine learning clustering techniques to group together historical customers who exhibited similar behaviors based on the attributes which are predictive of that behavior.
An analyst reviews the clusters and determines how Customer Retention should react to different predicted churn scenarios. In this example, customers that have a “month-to-month contract” versus an annual one is one of a handful of signals that has a high likelihood to churn.
Step 4: Refresh often to update each customer’s prediction
New records can now be scored on demand to determine likelihood of churning and segment assignment, with automated actions being triggered in response to different segments.
While a customer today may not exhibit signals that they will churn, next month—based on fresh data— their likelihood to churn may go up. Every time the data is refreshed, the predictive segmentation model starts over so that Marketers can react quickly before the customer actually cancels their service.
Actionability: What can a marketer do with predictive segmentation?
Having predictive segmentation makes your marketing smarter, but how does a practitioner act on this information? simMachines has been built from the ground up to be able to send your predictive segmentation to anywhere you need it to go to get the maximum benefit from your segmentation approach. We have customers that connect their SimMachine segments to multiple platforms across their organization and partners.
For marketers, there are so many ways to use this data!
simMachines can data onboard your predictive segments to media companies such as LiveRamp for media activation. From there, your segments can be used for audience targeting, setting bid prices, guiding messaging, and other uses for media buying.
Some common marketing use cases for predictive segmentation include:
Upselling/cross-selling. Have a variety of products or services that you want to sell to your existing customer base? Predictive modeling can help you understand what you should recommend to each customer based on their likelihood to convert. Deliver the best next product, content, or offer every time.
Increase customer engagement. Get a better understanding of when and how to reach out to your customers based on their likelihood to engage with that email, text, catalog, or message. Send notifications at the right time with the best chance to be read.
Personalization. Automatically adapt the journey for each individual consumer, along a predefined funnel based on their propensity to take the action you desire.
Lookalike modeling. Once you have identified your most valuable audiences, marketers can build lookalike audiences that aren’t yet their customers.
The marketing applications for predictive segmentation are limitless!
Interested in telling the future?
The level of computational power needed to run advanced predictive segmentation such as the use cases I’ve described here isn’t small. Even a decade ago, to run predictive modeling on a customer dataset of a million records and a thousand columns of data could take time—marketing performance is directly correlated to how fast that marketers can react to incoming signals.
It was also very expensive and only for the biggest companies in the world who could afford the data scientists and tools needed.
But, because of recent innovation in machine learning, partners like SimMachines have democratized this powerful approach, and now companies like yours can mine predictive segments within their own data set using a self-service, web-based platform fast enough for your marketing team to react.
What are you waiting for?
Reach out now via our online form, email us directly at firstname.lastname@example.org, or call us now at 260-632-7378. Let’s chat so that we can answer any questions you have to see if predictive segmentation is a good fit for your business.