Part 2 | Predictive Modeling | Using Predictive Techniques to Decrease Consumer Returns
This series of posts talks about the problem of customer returns in the retail industry, using return rates as a proxy for customer satisfaction. We explore solving this problem by using a data driven marketing automation approach to increase customer satisfaction, via reducing return rates.
In 2016, US online retailers reported return rates of between 20% and 40%. Research by Accenture estimates that 68% of consumer electronics returns are labeled as NTF (no trouble found), and another 27% are due to buyer’s remorse which can occur for a variety of reasons similar to NTF, including subpar consumer experience and a lack of consumer education. Taken together these figure suggest that 95% of consumer electronics returns are for reasons other than product defects.
According a study by the National Retail Federation, 49% of retailers now offer free return shipping, and high rates of returns are putting an increased financial burden on retailers that are already operating on razor thin margins. Decreasing costs by increasing customer satisfaction and reduce rates of return, has become an imperative for retailers hoping to compete in an increasingly competitive online marketplace.
To create a data driven, automated method to doing this, there are two big ideas — the first is to ensure that the customer purchases the right product configuration upfront. The second is proactive outreach when you have reason to believe that the customer is likely to return. In either case, predictive data & messaging technology that understands customer behavior holds the key.
There are three aspects to this method:
- Data Collection
- Predictive Modeling
- Consumer Messaging
In Part 1 we talked about Data Collection in Part 2, we’ll tackle Predictive Modeling.
PREDICTIVE MODELING
Our returns analysis focused on evaluating customer behavior at the beginning of the user’s time with the product — how the device is used over the first three days, initial week and month to determine when to engage with the customer to shift the retention curve away from a return and into continued usage. To do this, we developed a model to predict likelihood of returns. This predictive model allows us to figure out who to reach out to and when, as well as to measure the impact of that outreach to close the loop.
Before we began, we knew (via customer RMA reasons) that the main reasons for customer return fall into one of three categories.
- Unwanted Gift (1/2 of all returns)
- Hardware issue (1/3rd of all returns). The ideal outcome here is an exchange, not a return. Usually faulty units have battery or screen issues that get resolved in subsequent device builds.
- Software issue (1/5th of all returns). These are rarely caused by an actual defect, but by inadequate onboarding and training — this became a significant focus of optimization.
After observing customer behavior in the first seven days (see Part 1), we built a classification model that could predict whether or not an individual is likely to replace or request a refund for their device, with 80–90% accuracy.
We did this by performing a simple cohort analysis to get a feel for what the retained vs returned user was doing on the device. We stack ranked their events by highest to lowest number of occurrences. Please recall from Part 1 that our database has events captured by email, and we’re able to look across devices, apps and support tickets.
Cohort Analysis
Returned vs Retained Cohort Event Comparison; data illustrative, not actual
Random Forest Model: We then took the variables above and ran them through a random forest model to predict the dependent variable, which predicted whether the user would return the device or not. 70% of sample to train model; 30% of sample to test model. Here is a confusion matrix to see how to model performed on unseen data:
Confusion Matrix for RF Classification Model; data illustrative, not actual
Now we’re able to predict with a high degree of accuracy within n days since device activation (n=3days,7days,30days) to understand if a user has a high likelihood of return. So now that we can predict whether or not a consumer will make a return, how do we change that behavior? Discover messaging strategy and execution in Part 3.
WANT MORE?
Get a Free Copy of our e-book, “Predictive Marketing: All you need to know to increase engagement and revenue” by clicking here.
ABOUT SID CHAUDHARY
Sid Chaudhary is the founder & CEO of Intempt (https://www.intempt.com). Intempt employs machine learning to chomp through large swaths of data, analyze your customer’s unique fingerprint, and build a profile that not only targets your consumer in real time, but predicts their future behavior. Intempt employs AI capabilities to immediately analyze the success of those offers on a consumer-by- consumer basis, and incorporates that data into your database of consumer profiles, empowering marketers to implement sophisticated, data-driven campaigns, that fully optimizes on a level that only multi-billion dollar technology companies have been able to achieve. All on an easy to use platform. Without writing a single line of code.
Sid lives and works in San Francisco.
