In previous blog posts, we’ve highlighted Velocidi’s ability to run machine learning predictive models on first-party data. This allows you to increase conversion rates, optimize ad spend and personalize your customer’s journey. This time, we will dive into two ways predictive modeling helps accelerate your growth, using specific models as examples.
- Deeper Insight
- Real-time Engagement and Revenue Optimization
But first, some ML fundamentals.
ML vs. AI
“AI” is a term used broadly to apply to any and all cases where a computer performs a function in a way that mimics human reasoning. But it isn’t just one thing applied universally. AI is really an umbrella term for numerous individual algorithms that each have their own purposes and use cases. Each algorithm is trained to solve for a specific problem.
“Machine Learning” is yet another umbrella term, but nested within the larger umbrella of “AI.” Algorithms that fall under machine learning are designed so that you only need to provide them access to data and some foundational logic, rather than try to anticipate every bit of knowledge it will need at the outset. Machine learning algorithms learn from the data available and self-improve over time based on their own output.
“Predictive Modeling” is type of statistical analysis. It can be done without AI or ML, but applying machine learning to it allows us to solve more complex problems, and make our models much more accurate. Different algorithms are applied to predictive modeling depending on the use case. Most of the algorithms used for predictive modeling fit into the machine learning sub-category.
Therefore, when we use the term “machine learning predictive models,” we mean that our predictive models are designed using machine learning. This makes them more powerful, more accurate, and allows them to constantly self-improve.
ML begins and ends with high-quality data.
In order to function at its intended capacity, AI needs a continuous high volume stream of data. This is the first and most fundamental thing Velocidi delivers. Our Private CDP captures real-time data from your online domains and stitches it together with data ingested from external systems. By bringing all of your data together, Velocidi gives you a unified view of your customers’ journeys. And, of course, the more data you collect, the more accurate your predictive models become.
Machine Learning is Math – Not Magic.
If there’s one thing we want you to take away from this article it’s that despite the hype and confusion going around, ML is as practical as any other data analytics tool. It’s a sensible and methodical process involving people and experiments and ongoing improvement. Even as the algorithm improves itself, our programmers also make improvements based on their expert observations of the output.
Two different businesses trying to solve the same problem might start with the same base algorithm. But as each algorithm is trained and begins to self-improve, they will soon diverge enough that they would only product accurate results for the business in which it was trained. Using ML means you’re getting a bespoke solution that continually works better and better for you as it keeps learning.
And now, here are two practical examples of how ML predictive models can help your business grow.
1. Deeper Insight
One great example of predictive modeling uncovering elusive insights is the Predictive Discounting model we use in our platform. This model was designed to find out, on an individual or situational basis, when offering discounts is profitable for you, and when it’s not.
There are many complex variables when it comes to discounts that are a massive headache to process without the help of machine learning. A customer’s discount threshold may be different for different products, or it might depend on their total order value.
It’s especially tricky if you are dealing with non-recurring purchases, like consumer electronics or furniture. In this case, you don’t have comparable data to easily determine whether the customer who bought with a 10% discount is likely to have purchased without the discount.
Luckily, that’s where your robust foundation of behavioral data comes in handy. Even with non-recurring purchases, we can experiment with different levels of discounts within the same product, add an additional dimension to experiment with customer segments, and train the algorithm towards reaching the best possible results.
Over time, the model continuously becomes more and more accurate as it learns from its results. Not only do you discover nuanced ways to adjust your discounting strategy, but you also gain deeper insight into your customers’ behavior and purchasing patterns.
2. Real-time Engagement and Revenue Optimization
Likelihood to Purchase
Velocidi’s Likelihood to Purchase model is a customer analytics staple with versatile applications. One of its uses is to determine which of your online visitors are likely to become customers, and which are not — by taking into account all of your customers’ omnichannel interactions with your store. This is valuable information you can use to time your messaging, optimize advertising campaigns, and predict revenue flow.
Based on behavioral indicators in each customers’ website activity, this predictive model gives a score from 1-100 for how likely it is that the customer who just left your store will come back and make a purchase when prompted by a retargeting ad.
Common indicators the model might use include product views, frequency of website visits, cart abandonment, or ad click-throughs. Your store may have other specific indicators that matter more for you. The predictive model’s machine learning capability learns what those are by processing both historical and real-time data.
When it comes to optimizing campaigns, the output of this model can be used as a real-time segmentation attribute. This model is directly integrated with the segmentation and activation tools in the Velocidi platform, so the output is immediately actionable for real-time engagement.
Segments can be created to target customers who are likely to purchase within one day, two weeks, a month, or whatever time frame is appropriate for your sales cycle. You can use these attributes to time the delivery of ads, as you nudge your customers toward purchasing.
This is also helpful in reducing ad waste. It allows you to separate out the customers who are less likely to convert, and either target them with different messaging or dial down your targeting for that segment. This predictive model also helps make sure that customers stop seeing your retargeting ads after making the intended purchase. As soon as a customer makes a purchase, their likelihood to purchase score for the given time frame immediately goes down, which then excludes their unique identifier from your ‘high likelihood to purchase’ targeting segment.
Thanks for reading! Check out the other articles in this series and subscribe to our newsletter below to get future posts delivered to your inbox.