Predicting customer lifetime value: a practical guide for marketing teams
Every marketing team says it wants "high-value customers". Far fewer can tell you, on the day someone signs up, whether that person is likely to be one. That's what lifetime value (LTV) prediction does, and it's more accessible than most teams assume.
What an LTV model actually does
An LTV model looks at what a new customer did in their first few days (how they arrived, what they browsed, how quickly they made a first purchase, how much they spent) and estimates what they're likely to be worth over the next 90 or 180 days.
It doesn't need to be perfect. It needs to be good enough to rank customers: who's likely to be worth a lot, who's likely to churn, and who's in between.
You probably already have the data
For most businesses, the inputs are already sitting in existing systems:
Acquisition data: channel, campaign, device, country.
Early behaviour: sessions, time to first purchase, first-week activity.
Transactions: order or deposit values and frequency.
Outcomes: what customers from 6 to 12 months ago actually ended up spending.
That last part is your training set. Historical customers with known outcomes teach the model what early signals look like for valuable ones.
Keep the first model simple
You don't need a data science department to start. If your data is in BigQuery, BigQuery ML lets you train a regression or boosted-tree model with SQL. A gradient-boosted model on a few dozen well-chosen features will usually beat an elaborate model built on messy data.
The real work is in the features and the data hygiene, not in the algorithm.
Where the value shows up
A prediction is only useful if it changes a decision. The most common wins:
Smarter bidding. Feed predicted value back to ad platforms so they optimise for valuable customers, not just cheap conversions.
Earlier retention. Flag likely high-value customers who go quiet, and reach them before they leave.
Better budget allocation. Compare channels on predicted 90-day value per euro spent, not cost per sign-up.
Personalised offers. Stop giving the same bonus or discount to everyone regardless of how likely they are to stay.
Measure it honestly
Before trusting the model, test it on customers it has never seen and check whether its rankings hold up. Then run a controlled test: use predictions for one segment and business-as-usual for another. If the predicted-value group doesn't perform better, the model isn't ready yet.
LTV prediction isn't magic. It's a disciplined way of spending more on the customers who are worth it, and less on the ones who aren't.
Comments