Here’s How Little Data You Need to Calculate Customer Lifetime Value
There is no doubt that predictive analytics are the way of the future for marketers, particularly LTV or CLV calculations (Lifetime Value or Customer Lifetime Value). CLV is the measure of a customer’s net worth to your business during your entire relationship with them: past, present, and future.
Lifetime value measurement helps marketers make a wealth of informed decisions, from the amount that should be spent on acquisition and retention to cross-channel allocations to sales forecasting to overall corporate valuation. Many businesses fail to make acquisition and retention decisions based on CLV and end up overspending to attract or keep customers that will end up being a net loss.
Lifetime Value Calculations Don’t Require a Lifetime of Data
Lifetime value does not end at today. But because we cannot see into the future, we need to make some predictions based on the data we have from the past leading up until the present.
Below are a few options you can use to measure lifetime value based on the type of data to which you have access.
If You Have Transaction Logs
In an ideal world, you have the following type of visibility into your customers’ activities, with the horizontal lines representing each person’s lifespan with your brand until today:
How do you take this historical transaction data and use it to project what will happen in the future?
If you look at a cohort (a group of customers acquired during the same time period) and plot what they do over time, they will generally purchase less as time goes on. In the customer analytics universe, this truth is about as certain as the law of gravity to a physicist. No matter how you market to them, engage them, etc., the collective purchases of any given cohort are going to slow down. Although seasonality or promotional activity might introduce temporary variations, the purchase pattern will return to an ever lower baseline when things get back to normal.
Buy Till You Die (BTYD)
Although it sounds like another way to say “Shop till you drop,” Buy Till You Die is actually a statistical methodology that describes a customer’s pattern of purchases until they stop buying from a particular brand. It assumes that each person has an average time between purchases, which isn’t always consistent. A customer may average one transaction every three months, but suddenly purchase multiple times in a row. At some point the purchases stop and the customer “dies.” From a predictive standpoint it’s better to use these customer stories in the aggregate.
It turns out that this simple but “harsh” model not only does a good job of capturing the steady drop-off process, but it does a better job than anything else of showing how that process will continue to evolve in the future. The fact that it’s so simple almost seems like a problem. Humans are complex! It’s hard to understand our behavior. However, the best model to capture and forecast it is surprisingly simple.
If we look at how many purchases a cohort makes over time after the initial purchase (with the first purchase made in Q1 in our example), we will notice that the bars on the right side of the histogram representing heavy purchasers tend to be a lot shorter. Meanwhile, the zero bar will grow from one quarter to another.
If we look at an aggregate tracking plot for a cohort, we can look to see how rapid the drop-off is and compare those characteristics from one cohort to another. The way those downward curves vary between cohorts tell us about the BTYD characteristics of that group. You might even layer in demographic or machine learning variables here. Then you can start to look at the cohorts that die more slowly and figure out what defines them. And here are your actionable insights as a marketer: you can see things like —
- Which channels this cohort came from- social media, referrals, mobile ads, etc.
- How much the customers in this cohort cost to acquire
- How much this cohort is worth to your business over the next several quarters or years (sales forecasting)
- How much you should spend to acquire and retain these customers
If You Don’t Have Transaction Logs
If you don’t have a transaction log, or don’t have the means to analyze so many data points, there are a few simpler options that will still do the job quite nicely:
- Recency, Frequency, and Monetary Value (RFM): Most retail marketers are familiar with the concept of RFM. This is an individualized measure of how recently a customer purchased from your brand, how frequently they did so, and how much they spent. It’s an oldie but a goodie for a simple reason: it’s consistently great for diagnostic and predictive value. The most important part about it is that every customer is being looked at in the same way, with three values attached to them. You don’t really need to know who bought what and when down to the day or minute, even if your POS system can tell you that. It’s a great way of distilling what could be thousands of data points down to three standard numbers per customer.
- Recency and Frequency (Just RF; drop the M): If you ask an economist, he or she will say that people have a finite amount to spend, so you might think purchase size would lower for customers who buy often. But in the real world, this is rarely true: for the most part the “flow” of transactions over time (captured by R and F) and the size of those transactions (M) are largely independent of each other. Thus, we recommend looking at the two separately – building separate models for transaction flow (which is more important) and transaction size. You can then join these two forecasts together on the back end to get a meaningful picture of customer behavior
- Recency Alone (just R): It turns out that if you have R but not F, you can still model and forecast customer behavior reasonably well. You might ask why any business would need to do this? Let’s take a mom and pop coffee shop that is still writing out receipts and does not have a POS system. Instead, they have data from comment cards or maybe those hole-punch loyalty cards that show the date of last purchase. They can use this simple information to forecast how many lattes they might sell to each customer next quarter.
I have news for you—if you’re doing any of the above, you’re practicing REAL data science. You’re not racking up data points just for the sake of it and making pretty models or complex databases to query. Data science is about finding what exists below the surface of the data and using that information to explain and predict human behavior. The question should always be: What are the most efficient ways to gather and analyze the data that gets to the truly important insights underneath? Too often, the question today is, “How can we make use of all the data points we’re gathering?”
We’re not pretending it’s easy to perform the type of analysis above, unless you’re a data scientist or statistician. The good news for marketers is- if you have simple transaction data, or RFM, you don’t need “big data” to get big insights from your data science team or vendor. In fact, we often find that layering in too many variables such as demographics or behaviors like website visits and social activity merely clouds the forecast.
Of course, a tool like Zodiac can go deeper and get more exact than what we’ve described above. But our point is that instead of feeling like you have to make your data fit into a rigid CLV model, it’s quite the opposite: you can change the models to accommodate the data you already have. It’s important to keep in mind that the purpose of manipulating all this data is to achieve your business goals. And if it’s not working toward that end, it’s just noise.
Special thanks to Peter Fader for his help with this post. You can find another version of it on the Zodiac Blog.
