Big data meetup budapest adding data schemas to snowplow
Customer lifetime value
1. Customer lifetime value
What it is
Why it matters
Using it in practice
SnowPlow Analytics Ltd
2. What is customer lifetime value?
• Prediction of the net profit attributed to the entire future relationship with a
customer (wikipedia)
£50 £10 £1000 £100
• The most important metric in business analytics (incl. digital)?
• Not widely used… (Because it is hard to calculate, esp. in digital)
• Example: using CLV to acquire customers for a mobile game
SnowPlow Analytics Ltd
3. Why is customer lifetime value important?
20% of our customers Customer acquisition
account for 80% of costs keep rising
our sales
The best customers might be It is often more cost effective to
– Brand loyal spend money retaining existing
customers than acquiring new
– Don’t “shop around”
customers
– Rich
– Different from the average
SnowPlow Analytics Ltd
4. Where is customer lifetime value used?
Customer acquisition Customer relationship management
1. Use average CLV to inform • Maximize customer lifetime value
acquisition cost – Instead of maximizing other metrics
– E.g. pay more for a customer than e.g. utilisation
recoup on first purchase, based on – E.g. email marketing to encourage
Increasing sophistication
likelihood that he / she will make a repurchase
second / third / forth purchase)
• Differentiated approach for different
2. Calculate CLV per channel customer segments
– pay more more to acquire customers – Spend more cultivating loyalty in the
on channels with higher CLV most valuable customers
– E.g. search engine marketing vs price (personalisation) e.g. loyalty
comparison sites schemes
Acquire valuable customers Retain valuable customers
SnowPlow Analytics Ltd
5. Calculating customer lifetime value: 2 challenges
• We need to be able to attribute profit to a customer over his / her entire lifetime
– Profit across sales channels (on and offline)
– Single customer view?
– Web analytics packages visit rather than customer-centric
• We need to be able to forecast lifetime value based on past behaviour to date
– Need a model that matches the data (reasonably well)
– Needs to be done fast if used to acquire customers
– Limited data set
– Prediction is an art, not a science
SnowPlow Analytics Ltd
6. Meeting those challenges:
1. Measuring actual customer lifetime value
1. Identify the moments in a customer journey where value is generated
2. Tie records for a specific customer together into a complete journey
– E.g. using sales records, loyalty programmes, cookie IDs
– If it is not possible to do at a customer level, then do at a segment level (and infer
average CLV from segment lifetime value / number of customers)
3. Measure the profit made at each point
– Normally use gross profit for simplicity Doing this is getting easier all
the time:
1. Improvements in
4. Sum them over the customer’s “lifetime” analytics solutions e.g.
Universal Analytics
2. Companies are getting
better at getting user’s to
identify themselves e.g.
via logins
SnowPlow Analytics Ltd
7. Meeting those challenges:
2. Forecasting value based on past behaviour to date
1. Identify the moments in a customer journey where value is generated
2. Examine the value created at each moment: what is it a function of?
– Does it vary much by customer / segment/ time / anything else? (I.e. wide variance in
values)
– If that variation is significant, what is it a function of?
3. Examine the likelihood of moving from one moment to-the-next: what is it a
function of?
– Does it vary much by customer / segment / time / anything else?
– If that variation is significant, what is it a function of?
Developing a model is likely
much easier for a telecoms
operator (reliable subscription
revenue) rather than an online
clothing retailer
SnowPlow Analytics Ltd
8. An example: using CLV to drive customer acquisition
• Mobile game
• Free to download, monetise by in-app purchases or virtual goods
• Virtual goods can be bought at any stage of playing the game (i.e. very frequently or
never at all)
• Wide variety across customer base in terms of customer lifetime value
– Zero value from majority of users. (Who play without ever buying an item.)
– Small fraction account for disproportionate amount of value
• Crucial to acquire users from channels where a high proportion of acquisitions
have high CLV
SnowPlow Analytics Ltd
9. Calculating CLV: the steps
• Measuring the lifetime value of existing customers was easy:
– All the data in a single system
– Easy to track customer consistently (through single account)
• Forecasting value based on behaviour to date was hard:
– Massive variation number of purchases by customer (from 0 to a very high number)
– Massive variations in the length of time consumers play game (download and never play
vs download and play for months / years)
– However, limited variation in each purchase value (all virtual goods cost roughly the
same)
SnowPlow Analytics Ltd
10. One key insight led to a simple model for CLV
• Customer lifetime value varied widely between channels
• The best predictor of whether a customer would purchase a virtual good in future was
whether they had purchased a virtual good in the past
• Within each channel, the likelihood that a customer would make another purchase was
constant (i.e. independent of the number of purchases they had made to date)
– This means lifetime value can be modelled as a geometric series where each term in the series
represents a purchase event
– The ratio between terms represents the probability that a user makes an nth purchase having made
an (n-1)th purchase. That ratio, r, is what needs to be measured for each different channel
– Once you have r for a channel, then the lifetime value of the customers acquired can be estimated:
(p = average price of virtual good)
Value of 1st purchases Value of nth purchases
SnowPlow Analytics Ltd
11. So what?
• Easily prediction lifetime value by channel:
– Measuring r is easy: it is calculated simply from the ratio of 1st purchases, 2nd purchases etc.
Keep the model as simple as possible. Use intuition about
customer behaviour to derive key modelling insights
• Fast results:
– Purchase events were, as a whole, frequent enough that a value could be calculated for r based on
only a few days worth of data
• Accurate results:
– Estimations of lifetime value were found to be accurate to 12%
• Powerful results:
– Marketing budget was optimized to those channels driving the most valuable users
If you have large variation in customer lifetime value
between segments, your CLV prediction might not be very
precise but canAnalytics incredibly useful
SnowPlow still be Ltd
12. Questions
• Where do you use CLV? Where do you want to be using it?
• What type of models have you built?
– What worked?
– What didn’t?
– Why?
• Any other questions or insights?
SnowPlow Analytics Ltd