Lifetime Value Part 3: Tarot cards, crystal balls and fortune tellers? Lifetime value is all you need when predicting success
Summary
After using RFM scoring to segment our customers and getting great results, we might want to build a more sophisticated model to move beyond a ‘score’ and get more granular data points to use when segmenting our customers.
In order to do so, we can build a predictive model to provide a more complete view of customer lifetime value, using historical customer behaviour to predict expected purchases, expected order value, overall spend and likelihood of churn.
This data can be used to segment our customers, and we can optimise our strategy for each segment to drive long term revenue growth
We can also build models to understand how certain customer attributes (such as age, gender, product purchased) and site behaviour (source, time on site, pages visited) influence customer value to allow us to adjust our strategy to focus more on customers likely to have a higher lifetime value.
This data can also be used to predict the LTV of new customers.
Context
If you have been following the series to date (part 1 and part 2) you should have an understanding of the importance of measuring lifetime value, and you may even have tried to calculate RFM scores for your own business.You are now ready for the next stage in the LTV journey: building a predictive model. However, before you get into the nitty-gritty, remember the problem with 'big shiny things'. When thinking about building a more complex model, it is important to identify and get agreement on key actions that you can take using the model, each of which should have an impact on core business KPIs. The model/report is not the output of the project, but simply a means to an end.
Building a predictive lifetime value model
Let’s assume you’ve done the basic RFM segmentation, and found that the model was a bit restrictive. As an example, you notice that customers that buy more than five times are a lot more likely to buy again, but customers that buy less than five items will never buy again, and the model does not take such nuances into account. This is where you may need to seek out your friendly neighbourhood data partner so that you can find a way get more accurate predictions for future spend.
The RFM model that we discussed last week is an example of a historical model. It assumes that if someone had high recency, frequency and monetary value in the past, they will be high value in the future, which is a good proxy in a lot of cases. However, why don't wt model specifically how R, F and M influence customer value? This is where a predictive model comes into play. To develop these models, we look at historical data on customer behaviour and model how recency, frequency and monetary values influenced overall customer spend over that period, and use this data to provide a prediction of future lifetime value. This provides a more accurate understanding of future customer value than building broad buckets of R, F and M, as it takes into account specific business nuances.
To build this model, we start with a survival curve. This type of curve was historically used in biology, to predict the likelihood of a virus surviving over time, but we are going to use it to understand the likelihood of a customer staying with your business over a period time. The curve below is an example of a survival curve for an individual customer, with the y axis representing the likelihood of a customer being ‘alive’ and the vertical line representing a customer purchase. When a customer makes a purchase, their likelihood of being alive increases to 100%, and then decreases over time until they make another purchase; note that with each additional purchase this likelihood decreases at a slower rate (the gradient is less steep) as their loyalty is assuming to be increasing.
Figure 1: Survival Curve
To build this model we also need to model a ‘negative binomial distribution’ curve (NBD) to predict the number of purchases a customer is likely to make. When we multiply expected purchases by an estimate of average order value, we get the LTV: the amount we predict that the customer will spend in a given time period. Once you have the LTV you can paint this beautiful curve, which is sure to impress your team. It tells us that 20% of our customers drive 70% of our future revenue. Again, we must remember to ask: what is the action we can take from this insight?
Figure 2: LTV curve showing the distribution of customers and LTV
Segmenting your customers with predictive models
Using your fancy model, you now have a set of new data points: expected purchases, expected order value and most importantly, predicted LTV. The model will also output another metric called P(Alive) which shows the probability of a customer churning. The actions you can take is the same as that mentioned in the previous post: segment the data! However, now you can have a much more accurate data set to use for your segmentation: as an example, you can create a high LTV customer group of customers who will spend more than $200, for example. Or your can use the P(Alive) metric to create a segment of ‘At-Risk customers, customers with mid-high LTV and high P(Alive). Keep these actions front and centre throughout this process as models don’t drive results, but data-driven actions do!
Understanding which features drive high lifetime value
The rubber really hits the road with this last idea...
Let’s say you know the LTV of your customers but now you want to understand more about your different customers. Are they older or younger? Male or female? Did they buy electronics or makeup on their first purchase? What possible actions might you take with these insights? Well imagine you knew that customers who bought electronics were likely to be high value customers. You might target electronics websites and spend more money advertising the electronics products., knowing that these customers are likely to spend more in the long run.
This is where your data partner comes in again. They will aggregate all the data in a privacy compliant manner, with features on one side, and the lifetime value on the other. So we will know, for example, that a 25 year old male who first bought a beard trimmer has a high LTV. But a 55 year old female buying makeup has a low LTV. Then they will build a model to find out the relationship between each feature and the LTV to answer the all-important question: What type of customer is likely to have a higher LTV? We can then personalise and target new customers that are likely to have a higher value to your business in the long term.
You can even use this data to predict the lifetime value of a new customer. If a 25 year old male visits your site and buys a beard trimmer, you should have an idea of how much they will spend over a defined time period based on the behaviour of current customers with similar behaviour. Thus, in this final stage of the journey, we can predict LTV of a customer upon acquisition. Often however, we are limited by the data we collect at point of sale, so such a model may not have a high enough accuracy to be useful. We have used this type of model successfully in industries that collect more data, such as banking. In this specific use case, we could predict how much a user would invest over their lifetime based on their first investment (What type of account did they set up? What type of investment was it? What products did they invest in? What other investment options did they use?)
Moving from data to decision with LTV
Hopefully these examples show you that LTV isn’t a buzzword, but a new way of thinking. Through this series we have shown some easy ways to use the LTV model based on segmentation: using detailed understanding of historical customer behaviour to predict future behaviour and segment your customers. You can then take appropriate action based on the customer segment to drive long term results for your company.
How have you used predictive LTV modes? What have been your learnings?
