Five ways to use predictive analytics
The era of graphs and spreadsheets as a way of thinking about analytics is beginning to approach its end. Predictive analytics, along with associated artificial intelligence (AI) and machine learning technologies, are changing the way in which we deal with data. These tools are becoming more accessible, and ‘big data’ thinking is no longer limited to firms with billion dollar budgets.
Predictive analytics provides a glimpse into the future, as well as access to strategic insights that can open up new opportunities. Here are five ways you can put predictive analytics to use, and how you can change the way you think about data.
According to Forrester research, predictive analytics has found three main use cases for dealing with leads. Specifically:
- Predictive scoring: This method analyzes how leads are responding to your marketing attempts and how likely they are to take action based on that information. In this way, you can more quickly identify which leads to focus more resources on and which to divert resources from.
- Identification models: This use case is an approach that focuses on comparing leads to customers who have taken actions in the past. In doing so, you can divert resources to those leads who are most promising based on previous actions they have taken, as well as identify new markets that you weren’t previously aware of.
- Personalization: In concert with predicting which leads are most likely to take which actions, the same data can be used to determine which leads respond best to which types of messaging. This advanced form of segmentation can take things deeper than simply splitting leads into groups – instead sending them much more personalized messages.
One prominent example of this was covered in the Harvard Business Review, detailing how a Harley Davidson dealership increased sales leads by 2930% using an AI named Albert.
The AI crunched CRM data to identify characteristics and behaviors of previous buyers. It then split them into micro-segments based on those characteristics. For each segment, it tested different combinations of headlines, visuals, and other elements to determine which worked best for each segment.
The value of your lead qualification is highly dependent on the value and quantity of your data. No matter how good your statistical models are, their abilities are still very limited without access to the information that they need to learn about your customers.
In the digital space – particularly if you are not using a CRM – the best place to start with predictive analytics will almost certainly be an integration of Google Analytics and Google BigQuery.
Modeling customer behavior
While lead qualification and conversion is the most obvious use-case for predictive analytics, and likely the one worth looking into first, it’s far from the only marketing application of this emerging technology. But virtually any use is going to have customer modeling at its core.
You can divide customer modeling into three basic types: cluster models, propensity models, and collaborative filtering.
Clustering is a way of segmenting customers into groups based on many variables. A cluster model looks for correlations between various attributes and identifies a number of equilibria in which certain types of attributes tend to accumulate. What makes clustering special, compared with traditional segmentation, is the sheer number of variables involved. Clusters often use 30 variables or more, far more than would be possible if you were manually segmenting customers, or even if they were manually segmenting themselves.
Clusters come in three forms:
- Product clusters: These are clusters of customers who tend to only buy specific types of products, ignoring other things in your catalog
- Brand clusters: These customers tend to buy from a specific collection of brands
- Behavioral clusters: These are segments of customers with a specific collection of behaviors, such as frequent buyers who place small orders, or customers who prefer the call center over the checkout cart.
What’s important to recognize about these clusters is that they enable predictions about which clusters people belong to – even with limited information. If they buy one product with a specific brand, your brand cluster can predict what other brands they may be interested in, rather than just the more obvious recommendation of simply offering everything else by the same brand.
A propensity model is one that makes future predictions about customer behavior based on correlations with other behaviors and attributes. This may be accomplished using regression analysis or machine learning. A good propensity model controls for as many variables as possible so that correlations aren’t confused for causes.
Here are a few examples of propensity models:
- Propensity to unsubscribe: A model like this allows you to determine the appropriate email frequency, weighing the possibility that a recipient will unsubscribe against any possible positive outcome
- Propensity to churn: These are customers who are likely to move on if you don’t take action, but who may be high value otherwise
- Lifetime value: Modeling the lifetime value of a customer can help you make strategic marketing decisions if it leads you to customers with more lifetime value, or leads to behaviors that extend lifetime value.
Other propensity models include predicting how far through somebody’s lifetime value you are, and how likely they are to convert or buy.
If you’ve seen Amazon’s “customers who liked this product, also liked…” recommendations, you know what type of model this is. At first glance collaborative filtering might sound similar to product-based cluster models, but collaborative filtering is a bit different. Rather than grouping customers by the types of products they are likely to buy, collaborative filters make recommendations based on aggregate behavior.
In other words, this is less about the user’s product preferences and more about the behaviors that products tend to cause for users.
There are three types of collaborative filters:
- Up-sell recommendations. These are recommendations for a higher tier version of a product before the sale is made
- Cross-sell recommendations. Also offered before the sale is made, this is a recommendation for a product that is often bought at the same time as the initial one
- Follow-up recommendations. These are recommendations for products that people tend to buy a certain time period after buying a prior product, such as replacing a product that runs out, or buying dishes after buying a table.
Connecting the right product to the right market
Working backwards from customer modeling, it’s possible to identify markets for your products that you may not have been aware of. Here are just a few examples of how this use case can play out:
- Incorporate referral sources into your cluster models. This will allow you to identify which traffic sources correlate with which types of products, brands, or behaviors. From this, you can immediately identify a new market for these products or brands
- Incorporate referral sources into your lifetime value propensity models. This will allow you to determine which locations to invest more of your marketing resources into, since you roughly know what the ROI will be
- Look for correlations between traffic sources and success with up-sells, cross-sells, and follow-up recommendations
- Look for correlations between keywords and your customer models
- Analyze the attributes that are strong predictors of buying specific types of products and brainstorm other markets that might share those attributes that you have not yet targeted
- Investigate high performing outliers where limited data is available and investigate whether expanding in those markets is a good option.
Connecting the right users to the right content
There are a number of ways that you can leverage your customer models to connect prospects with content in ways that move you toward your goals, some of them more obvious than others. Here are a few examples:
- Matching content related to products or brands based on the appropriate clusters
- Matching users to conversion copy when propensity models predict they are most likely to buy
- Recommending content to users that improves their propensity scores
- Recommending content to users that enhances their likelihood of responding well to an up-sell or cross-sell
- Matching traffic sources to the content that tends to produce high propensity scores for each particular traffic source.
As you can see, the number of approaches you can take here grows pretty quickly. Think strategically about how best to put your models to use and make the most of your models.
Discovering strategic marketing insights
While some predictive analytics tools can automatically streamline your marketing process and generate results (like Albert did for Harley Davidson), it’s important to remember that human decisions still play a very important part in the process.
Where predictive analytics and related AI tools often fail is in a propensity to ‘over-fit’ the data. They can get stuck at local maximums and minimums, incapable of making the leap to new terrain.
Escaping from traps like these, and making the most of these tools in general, requires you to find strategic insights from within your predictive analytics models.
For example, suppose you discover that a specific piece of content has a tendency to raise your prospects’ propensity scores; any automation you have in place can be applied to customize how your users are marketed to, and push them toward that piece of content. But what predictive analytics can’t tell you is whether there might be other traffic sources you haven’t tried yet that would be a good fit for that content. Using your experience and brainstorming capabilities, you can identify other potential markets for that content, feed them into your model, and see how the exposure changes things.
Your goal in working with these kinds of models must always be to find insights like these and test them to see if the results are as expected. If your model runs on autopilot it will not discover any new opportunities alone.