Customer Lifetime Value Prediction for Growing Businesses

The average CLV number is hiding the useful part

Most growing businesses already quote a lifetime value. It is usually total revenue divided by total customers, multiplied by some guess at retention. It goes into the investor deck and then nobody uses it for a decision.

The problem is that an average blends together customers who will buy once and vanish with customers who will order every month for five years. Spending the same amount to acquire both is how marketing budgets leak. The value of predicting customer lifetime value is in the spread, not the average: knowing, within a few weeks of a first order, which bucket a new customer probably belongs to.

Historical CLV, predictive CLV and why the difference matters

Cohort analysis is honest but slow. You need to wait a year to know what a January cohort was worth. Predictive CLV tries to give you a defensible estimate after the first few orders, which is the only timescale on which you can still change the acquisition spend.

Three ways to model it, from simple to heavy

1. Probabilistic purchase models. Classic 'buy till you die' models (the BG/NBD family plus a spend model) use only recency, frequency and monetary value. They are well understood, need little data and work surprisingly well for repeat-purchase retail.
2. Gradient-boosted regression. A machine learning model predicting 12-month spend from early behaviour, acquisition channel, first basket contents, discount used, delivery experience and so on. Better when you have rich signals beyond order history.
3. Two-part models. One model predicts whether the customer returns at all, a second predicts how much they spend if they do. Useful when a large share of customers never come back, which is most of eCommerce.

We almost always start with the first approach as a baseline. If a model with three inputs gets within a few percent of a machine learning model with sixty, the simpler one is easier to explain, cheaper to run and less likely to break.

A worked example

Take an online coffee subscription and gifting business with 40,000 customers. Historical data shows two very different groups: gift buyers who place one large order in December, and home drinkers who start small and reorder every five weeks.

Looked at by revenue in the first 30 days, the gift buyers look like the better customers. A predictive CLV model that also sees basket contents, delivery address matching billing address and whether a grinder setting was chosen will usually flip that ranking within the first order. If the home drinker's predicted 12-month value is three times the gift buyer's, paid search bids and welcome offers should reflect that, even though the first basket was smaller. These figures are illustrative; the shape is very common.

What to do with the prediction

A CLV score that sits in a data warehouse changes nothing. The decisions it should feed are fairly specific:

• Value-based bidding in ad platforms, by sending predicted value back as the conversion value rather than first order value
• Setting a maximum acquisition cost per channel based on predicted rather than historical value
• Deciding who gets proactive service when a delivery goes wrong
• Choosing which lapsing customers are worth a retention discount
• Forecasting revenue from the existing customer base for finance

The retention use overlaps with churn prediction. If that is the sharper problem for you, our post on building a churn prediction model covers the other half. In practice the two scores work best side by side: likely to leave and worth keeping.

When CLV prediction is not worth it

Honest answer: for plenty of businesses.

• One-off purchase businesses, such as a kitchen fitter, where almost nobody buys twice
• Contract businesses where value is already written in the contract
• Businesses with fewer than a few thousand customers, where cohort tables and judgement are enough
• Anyone who cannot act on the result, because the ad account, CRM and retention tools are not connected to anything

The last one is the most common. Before we build a CLV model at SpiderHunts we ask what would change on Monday morning if the prediction existed. If nobody can answer, the money is better spent on the plumbing first.

Building it into the business

A production CLV pipeline is mostly data engineering. Orders, returns, refunds and customer identity need joining reliably, because a customer who checks out as a guest three times looks like three low-value customers. Returns matter more than people think: predicted gross spend without predicted returns overstates the value of exactly the customers who order to try on and send back.

The model then runs on a schedule, writes scores back to the CRM and ad platforms, and gets checked each quarter against what customers really spent. Our data science work usually spends more time on identity matching than on the model itself, and that is the right proportion.

Frequently asked questions