Looking at the Numbers: How to Calculate and Leverage LTV


Importance of Understanding Lifetime Value

Many Ecommerce challenges come down to two things: knowing who your customers truly are and how to efficiently acquire new ones. Most companies have a mix of customers from the extremely loyal to one-time buyers. Yet they struggle to understand their customers on an individual level, and are unable to use more targeted advertising.

One model that has proven its efficiency with some of our Ecommerce clients is lifetime value (LTV). To illustrate this model, we’ll take a look at one of our clients, an Ecommerce retailer for planners, stationery, and other accessories. Their varied mix of customers, from loyal devotees to one-time buyers, requires advanced targeting techniques, but for this reason they are also an ideal candidate for LTV modeling.

Shifting from Total Spend to LTV Strategy

Many companies group their customers by total spend. This can be a good first approach but it has many flaws. For example, some of your historical top spenders may not have purchased anything in several years. What are the chances that these customers are still engaged with your brand?

What if your top customers haven’t bought in several years?

Another approach is to try to add rules of thumb like, “if a customer hasn’t purchased in the last 12 months, that person is considered churned.” While these types of rules are an improvement, they are not ideal since they don’t consider the statistical properties of customer transaction patterns. To address this issue, researchers develop statistical models to describe the transaction patterns of customers.

The model that has proven to be most efficient in our work is the customer lifetime value (LTV) model.

Ready to become an industry innovator
and take an LTV approach to marketing?

Overview of LTV

One of the leading researchers on LTV is Peter Fader, a Professor of Marketing at Wharton Business School. Peter Fader not only has several research papers on this topic, but also is the co-founder of Zodiac, a predictive analytics firm for customer valuation models and insights. But what is LTV? Well, let’s dig deeper.

One of the attractive features of LTV is that it requires a minimal amount of data. In fact, it only requires transaction log data, which is simply a dataset with the following columns:

  • Unique customer identifier
  • Transaction date
  • Value of each transaction

The beauty is that most companies already collect this information in one form or another. Here is an example:

RFM Analysis for Customer Segmentation

Using this data we can construct the “RFM object,” which summarizes each customer using 4 metrics from the transaction log. Here is a description of each:

  • Frequency: the number of repeated transactions, this is basically the total number of transactions the customer made minus 1. It measures how often a customer buys.
  • Recency: the number of months between the first and last transaction. It measures the time span a customer has been actively making transactions. A higher value of this metric is better than a lower one.
  • Time as Customer: the number of months between the first transaction and today. This measures how long ago a customer made the first transaction and scales the other variables. For example, a customer that made the first transaction a few months ago will necessarily have a low recency value.
  • Monetary Value: the average spend per transaction, also known as Average Order Value (AOV). This measures how much a customer spends when they make a transaction.

The great thing about summarizing everything in an RFM object is that we don’t need to keep track of specific dates and transactions. Here is what a dataset looks like in an RFM format:

Now let’s go over how the statistical models work. The main idea is that while customers make transactions they stay active and will follows rules that will predict their future spending. We're taking a high-level look, but if you're curious about the distributions and academic research, find check out the resources at the bottom of the page.

We fit these distributions with the RFM object and then use the estimated parameters to calculate the future spend of each customer. For each customer, we can calculate the likely number of future transactions, the probability the customer will still be active, and the average spend per transaction. Then the predicted Lifetime Value of a customer is equal to:

Expected # of Transactions
x
Probability of Being Active
x
AOV
=
LTV

Implementation of LTV Research

Many of our clients come to us with a goal of better understanding their customer base to make smarter marketing decisions. Through deep analysis by our Digital Intelligence team, Wpromote is able to deliver this insight and provide strategies on how to leverage across channels.
  • Removing entries incorrectly formatted (for example characters where we should have numbers),
  • Removing all the entries with zero or negative spend (those are most likely refunds and the model can’t incorporate that),
  • Removing outliers (customers with a significant amount of transactions and/or spending a lot per transaction, as these are most likely test accounts or retailers), and
  • Making sure all the remaining entries are in the correct format (for example dates are all in the same format, spend doesn’t have $ signs that can cause it to be read as a string, as so on)
Next, we compute their RFM object. One important variable you need to decide is the time interval. In this example, we decide to calculate everything in terms of months since people buy planners once or twice a year. The goal is to have a limited number of events in a single time period.

If people were buying 10 planners per month on average, the monthly time frame would not work, and we would look at weekly or daily time periods. On the other hand, if people only bought once every 3 or 4 months, you wouldn’t want to use a daily time frame because the time between transactions would be too large and this makes the estimations harder.

Key idea: High-value customers are more than 10x as valuable as Less Valuable customers.

After establishing the time interval, we need to make sure the model can predict the transactions and spend accurately. For this we divide the dataset into two: a training dataset and a test dataset. The training dataset is the data from 2015 and 2016, while the test dataset is the data from 2017. We use the training dataset to create the model and then to predict the transactions and spend for 2017. We then compare the predicted spend for each customer with the actual spend in 2017 to make sure the predictions were accurate.

Once we’ve passed the accuracy test, we adjust the model to fit the entire dataset and make predictions for the next 12 months. The final output for each customer contains:

  • Predicted number of transactions for the next 12 months,
  • Predicted average order value per transaction,
  • Total spend for the next 12 months (which is just the multiplication of the two previous metrics), and
  • Cohort group, which we’ll explain in the next section.

Ready to start using LTV to your advantage?

Cohort Strategy for Paid Media Teams

These groups were built based on the predicted spend for the next 12 months. Customers were placed into 3 different groups: high value, good value, and less valuable.

As the names suggest, the high value group contains the customers with the highest predicted spend, which was also the smallest group in number.

The less valuable group contains the customers with the lowest predicted spend. Customers in this group are predicted to have churned already, and this was the largest group in terms of number of customers.

Finally, the good value group contains the customer that have predicted spend lower than the top performers, but unlike the less valuable group they are predicted to spend money in the next 12 months.

At this point, we have 3 distinct client lists to share with the paid media teams. They can match the customer IDs with customer emails and created audiences in AdWords and Facebook. Here are just a few of the strategies the media teams can implement using our outcome from the LTV model:

High Value Group: create lookalike audiences based on this list, and use these lookalike audiences for prospecting campaigns. Our main goal is to find new customers that look similar to the current best customers.

Good Value Group: use this list for remarketing campaigns. We can use tailored campaigns to influence these customers to spend more. We want to move them to the top tier and avoid them moving to the bottom tier.

Less Valuable Group: remove them from campaigns’ targeting or reduce the bids. We want to save marketing dollars on customers that are predicted not to spend in the future.

Conclusion

By developing the LTV model, we are able to divide our client’s customers into three distinct groups and predict future spend. While all channels benefit, paid media in particular has a number of strategies based on this information. Having this invaluable insight informs all of our client’s marketing decisions and puts them ahead of the competition.

More Resources:

 
Become An Insider! Never Miss Our Industry-Leading Content

Thanks for signing up to be a Wpromote Insider.
You’ll be the first to get the scoop on our latest services, promotions and industry news.

CONNECT