MODULE 3 LESSON 2 – MODELING
Modeling is where the magic of predictive marketing happens. Using predictive analytics and machine learning, we take the marketing indicators that we developed and determine the best marketing strategies to get the highest ROI.
In this section of Predictive Marketing University, we will focus on the basic principles of modeling. Modeling is a complex subject that involves advanced mathematics, statistics and computer science. Our goal here is to have you become familiar with the concepts and ideas that will help you achieve your marketing goals.
First, lets discuss the difference between business analytics and predictive analytics. Business analytics uses past data in order to report on events that happened in the past. If we look at a business analytics report for 2014, it will typically have data about 2014 that summarizes key business metrics for the same period.
Business analytics does not use statistical modeling, but rather averages, sums and other simple functions. Business analytics is like looking in the rearview mirror in your car: It helps you see where you’ve been, but is hardly helpful in understanding where you are going.
The goal of predictive analytics is to take past data and apply mathematical models in order to predict future outcomes. Therefore, while both business and predictive analytics use the same data, they deliver very different outcomes.
In every modeling project, it is critical to identify the outcome that we are trying to predict. This outcome should be clear, measureable and impactful for the organization.
Here are the most popular goals for predictive marketing:
- Better funnel efficiency: Help reps call the best leads by identifying those most likely to respond.
- Lead Velocity: Increase the speed of leads passing through the funnel by identifying “low-hanging fruit” leads that will close the fastest.
- Data cleansing or filtering: It is estimated that 30% of your database contains 95% of your wins. By eliminating 70% of your data, you will be able to focus on the leads that really matter.
- Data-Driven Segmentation: Segment buyers, matching each lead to a distinct persona and adding the needed information to your data.
- Cross-sell or upsell: Use algorithms to identify cross-sell and upsell opportunities for prospects in your database.
- Green field discovery: Bootstrap a marketing campaign when you don’t have prior information about your prospects.
After we identified our goals, it is time to get to the model itself. Here is a high level illustration of how a predictive marketing model works.
The best way to explain that is with a very simple example: Imagine that we have a farm and we try to predict the length of the fence that will cover the perimeter of the farm. The only thing that we know about the farm is the length and the width of the farm.
Now, this is a very simple example, but it is only for illustration. If we had a computer that would scan the link between length, and perimeter the computer will find that – Two times length plus two times width equals the perimeter. Two is the coefficient both for length and width that our predictive model found.
Now we can do the same in predictive marketing.
If we want to predict whether a company is likely to become a customer, we look at all of the new customers versus all of the unqualified leads. Then we tell our predictive model to find coefficients that will predict who will become a customer.
As in our very simple example of width and length of a farm, our model also delivers a list of coefficients that weights our data to determine the likelihood that a lead will be come a customer.
The model that we explained here is relatively simple. Many models use more complex methods than just multiplying marketing indicators with coefficient. However, the idea of all of these models is the same, to use past data to predict future outcomes.
Here are some of the challenges with building accurate predictive models:
- Bad underlying data: The worst sin may be bad underlying data, or “garbage in, garbage out”. The quality of your prediction is as good as your underlying data.
- Handling missing values: Even with the most robust data mining and matching, some leads will have missing values. For example, we simply do not know which CRM platform the company is using. Experienced data scientists take missing values into consideration in order to minimize loss of predictive power.
- Scalability: It is moving from thousands to millions of leads. Many models work great on a few thousand leads. Again, it takes a very experienced data scientist to be able to take these models and apply them to millions of leads in an effective way.
- Low statistical significance and over-fitting: It is important to build a model that will deliver results that are robust and repeatable. Over-fitting happens when we have too few observations and the model cannot be extended to our full universe of leads. Again, experienced data scientists can detect that and handle that for you.
How can you achieve the best results with predictive modeling? Here are a few tips:
- Model validation: Make sure that your model has been successfully deployed and tested with dozens of clients.
- Optimization across the funnel: Make sure to optimize separately at any point of the funnel–from marketing campaign engagement to closing deals and lifetime value.
- Understand score drivers: Never accept model results as a black box. Make sure that you get reports with full visibility of what’s driving your score so that you can preform “sanity checks” on your model.
- Full control over leads passed to sales: Ensure that you have the ability to determine the quality versus quantity of leads passed to sales to optimize ROI.
If you have any questions, feel free to contact me directly at PMU@mintigo.com.
To learn more about Predictive Marketing & Big Data, watch this webinar replay presented by John Bara (Mintigo), Megan Heuer (SiriusDecisions) and Russ Glass (LinkedIn).
Get The Slides & Watch Now!