Solutions / LTV prediction

Customer LTV prediction

Every customer gets a predicted lifetime value in their first days, so acquisition bids on what a customer will be worth rather than on what they spent on the first order.

Marketing Analytics ML
100% of the base
scored daily, not monthly
Day-one forecast
instead of a 12-month wait
Backtested
on your own cohorts
Who it is for

Subscription, marketplace and e-commerce teams with at least a year of purchase history and enough volume for cohorts to mean something.

Short answer

A model trained on your own purchase history predicts what each customer will be worth over the next 12 or 24 months, from what they did in their first days — the products they chose, the channel they came from, the discount they needed, how they paid. The prediction is pushed back into the ad platforms as a conversion value, so bidding optimises toward customers who stay instead of customers who convert cheaply and leave.

The problem

Ad platforms optimise for the first order, and the first order is the weakest predictor you own.

01

Bidding treats every customer as equal

A discount hunter and a customer who will reorder for three years both send back the same conversion value. The algorithm faithfully buys more of the cheap one, because that is what you asked it for.

02

LTV is one blended number

A single average across the whole base hides that one channel brings customers worth four times the other. Blended LTV is a comforting number that supports no decision.

03

The answer arrives a year late

Real LTV is known when the cohort matures — which is exactly when the budget that created it has already been spent.

How it works

From trigger to result, step by step.

01

Define what value means here

Twelve or twenty-four months, revenue or contribution margin, returns and refunds netted or not. We settle this with your finance team first, because a model that predicts the wrong quantity is worse than no model.

02

Build features from the first days

First basket contents, category, discount depth, acquisition channel and campaign, device, payment method, delivery outcome, early support contact. The point is to predict from what you know at day one, not day ninety.

03

Train and backtest on real cohorts

The model is trained on older cohorts and tested on ones it never saw, then compared against the naive baseline — first-order value. If it does not beat the baseline meaningfully, we say so and stop.

04

Push the value where bidding happens

Predicted value goes back to Meta and Google as a conversion value through their APIs, and into your CRM and BI. Bidding then optimises toward long-term value with the same machinery it already uses.

05

Monitor drift and refit

Cohorts age, the mix changes, a new channel behaves differently. Accuracy is tracked monthly against realised value, and the model is refit on a schedule rather than when someone loses trust in it.

Before / after

What changes on the ground.

Today, by hand
×Ads optimise on first-order value
×One blended LTV for the whole base
×Channel quality is known a year later
×Budget follows cheap conversions
With the automation running
Predicted value on 100% of customers, daily
LTV by channel, campaign and cohort
A usable estimate within the first days
Bidding optimised toward customers who stay
What you get

Delivered, not demoed.

An LTV definition agreed with finance, in margin or revenue
A model trained on your history, backtested against a naive baseline
Daily scoring of the whole base into your warehouse and CRM
Conversion-value feeds into Meta and Google Ads
Monthly accuracy monitoring and a refit schedule
Built with

We build in your stack rather than moving you onto ours. The list below is what this solution most often connects to — other systems are a scoping question, not a blocker.

Python scikit-learn XGBoost BigQuery dbt Airflow Meta Conversions API Google Ads API Shopify
Time to production6–8 weeks
Build priceFixed quote
First stepFree mini-audit
Honest limits

When this is not the right solution.

·If you have under a year of history, or a few hundred customers, the cohorts are too thin to learn from and any number we print would be theatre.
·If your product is bought once in a lifetime, there is no lifetime value to predict. Spend the effort on conversion rate and referral instead.
·If you do not know margin per product, an LTV in revenue will point you at your least profitable customers with great confidence. We would rather fix the margin data first.

Questions we get about this one

Accurate enough to rank customers, rarely accurate per individual. We report how much better it is than first-order value on a held-out cohort, and that comparison is what matters for bidding.

Bring us the process that hurts.

The mini-audit is free: we take your version of this process apart and tell you plainly whether automating it pays. If it does, you get a scope and a fixed price.