In today’s session of Analytics in Action, I want to show how to use predictive analytics to create digital marketing success.
We’ve talked a lot about predictive analytics in prior posts, such as:
- What are predictive analytics?
- Improving ROI with predictive analytics
- and, steps in developing predictive analytics
- Also, check out what our friends at Marketo have to say about using predictive analytics here
Take at look at these prior posts if you want to understand what we mean by predictive analytics or how they contribute to ROI. I’m gonna focus on creating an example of using the 7 steps in developing predictive analytics from this last post.
Step-by-step predictive analytics
Step 1: The problem – develop a lead scoring algorithm
Here, I’m using an example from a client, we’ll call them The Widget Company. Widget offers a broad line of B2B software solutions sold through a dedicated sales force. Every 2 weeks they publish an e-newsletter with about 630,000 subscribers. Greg, the Marketing Director at Widget, has limited information on their subscribers — this is often a tradeoff between building a subscriber list and getting information at sign-up. Greg knows many of the subscribers are competitors, students, or other tire kickers. He knows if he asks the sales force to follow-up with subscribers, he’s going to hear a lot of grumbling because many are not prospects for any Widget product.
Enter Hausman & Associates. Greg hired us to develop a scoring mechanism to not only identify which subscribers were prospects, but what product(s) they might find valuable. Subscribers scoring above a predetermined threshold were assigned to a salesperson specializing in the particular product of interest.
Greg was also interested in knowing which articles were favored by subscribers representing a good prospect for 1 of Widget’s products. This required modifying the data readily available from the email provider, which only indicated opens and clicks for the entire list.
Step 2: Identify metrics
The first step was to identify which metrics we had available and which might be meaningful in developing a lead generation algorithm.
Widget’s e-newsletter data from their email client existed in 4 relational databases.
- Subscriber information
- Newsletter data — opens, opt ins, unsubscribes, email clicks, bounces
- Article data — title of each article in the newsletter
- Opens by users – the email address of everyone who opened a particular article
Data from these databases wasn’t clean. The most significant data problems were there was missing data and some data translation information was missing. Often a significant problem in predictive analytics (or any analysis technique), data collected without a clear notion of how it’s going to be used means a lot of data cleaning.
In this case, we first had to create a dictionary to correlate information between databases. For instance, database 3 contained articles by name, while database 4 contained coded data for the articles containing information about which issue contained the article and it’s position within the newsletter.
If Widget hired Hausman & Associates before starting the project, we’d have made sure the relational databases contained the key necessary to relate them. If you’re using social media as the data for predictive analytics, developing a key for relational databases is critical for success.
Step 3: Determine analysis technique
We determined regression analysis was the best option for generating a lead score for each subscriber. Commonly, predictive analytics uses regression analysis. We first needed to transform the raw data into something meaningful.
Working with Greg, the Hausman & Associates team identified metrics that likely impact lead scoring — opens and clicks. We weren’t able to use subscriber demographics, such as position and company, because much of this data was missing..
Opens and clicks resulted in points, with actions more highly correlated with buying signals getting more points and those showing little or no buying signal resulting in fewer points. For instance, if a prospect opened an article about 1 of Widget’s products, they received a certain number of points. If the subscriber opened the contact information from the newsletter, they got more points, opening a general interest article resulted in fewer points.
Step 4: Collect historical data
Hausman & Associates analysts went through the entire article database, scoring each article with points from 1 – 10 — we called this the impact factor.
The impact factor was used to score each subscriber based on the impact factor of each article they opened or clicked. Because some subscribers joined the list years ago, while others joined more recently, we used a weighting factor to ensure older subscribers didn’t get an artificially high score simply because they’d been on the list longer and new subscribers weren’t overlooked as prospects because they’d just joined the list.
Step 5: Analyze data
The actual algorithm we developed is proprietary, using several econometric models in construction. But, it was easy to create a score for each subscriber using it.
Step 6: Communicate findings
Hausman & Associates prepared a report for Greg and Widget. The report contained our methodology and the scoring mechanism used to calculate the impact factor.
Step 7: Implement changes
We also gave them the subscriber list — with each subscriber assigned a score. Greg determined a cutoff value he felt would effectively distinguish between viable prospects and tire kickers. The automated system we generated a lead form for subscribers who scored above this threshold, which was sent to the appropriate sales person based on geographic territory and product specialization. This solved the first problem Greg had with the existing system of handling subscribers.
The impact factor for each article, when multiplied by the number of subscribers opening a particular article, gave Greg the information he needed to improve newsletter performance. He set his team to work crafting (or curating) more content similar to articles scoring a high impact factor.
Step 8: Iterate
There’s actually an 8th step in using predictive analytics — feedback. Your goals should always include improving your predictive analytics.
We worked with Greg after he and his team implemented our findings. We monitored the conversion rates of leads generated to the sales force, their average order size (AOV), and the attitude of the sales force to the leads we generated. When we found a problem, we revisited the scoring mechanism used to create the impact factor for each article and tweaked the algorithm used to score subscribers.
The results speak for themselves. Widget increased sales and closed many of the leads generated by the predictive analytics algorithm. Sales people were happy because we generated good leads that closed without sending them a bunch of useless leads that weren’t prospects.
Final thoughts on predictive analytics
As you can see, using predictive analytics isn’t something you put together in an afternoon. Predictive analytics takes some skill, training, and the right software. Using predictive analytics for digital marketing is compelling because website visitors, social media connections and mentions, and email campaigns generate a wealth of valuable data that businesses never had from traditional media. And, the payoff from effectively implementing predictive analytics far outweighs the benefits of using descriptive analytics alone.
Whether you need a complete content marketing strategy, some help with predictive analytics, or some consulting to optimize your existing social media marketing, we can fill your digital marketing funnel. We can help you do your own social media marketing better or do it for you with our community managers, strategists, and account executives. You can request a FREE introductory meeting or sign up for my email newsletter to learn more about social media marketing.