Growth Systems Library
Predictive LTV
Predictive LTV uses machine learning to forecast a customer's lifetime value at or near the point of acquisition — enabling growth teams to make bidding and budget decisions based on the expected value of a customer rather than their first-order revenue.
Predictive LTV (lifetime value) is a machine learning approach that estimates the expected future revenue or contribution margin from a customer, typically within the first 7–30 days of acquisition, based on early behavioural signals. Unlike historical LTV (which is calculated retrospectively once a customer's purchase history is complete), predictive LTV generates a forward-looking score that can be used immediately to inform acquisition spending and audience targeting decisions.
Exactius builds predictive LTV models as a core component of the data infrastructure layer in the Growth Operating System — enabling the Capital Allocation Loop to use expected customer value rather than observed first-order metrics as its primary input.
The fundamental problem with optimising acquisition on first-order metrics (CPA, ROAS) is that the customers who cost the least to acquire are not always the customers with the highest lifetime value. A subscription business that bids to minimise trial-to-paid conversion cost may acquire a large number of customers who cancel in month one. A DTC brand optimising for lowest cost per first purchase may fill its customer base with discount-seeking one-time buyers. Predictive LTV solves this by enabling the system to bid more for customers who are likely to be high-LTV, even if they cost more to acquire initially.
The LTV:CAC implication is direct. If predictive LTV can identify, at acquisition, which customers will generate 3x the revenue of the average customer, the business can afford to pay up to 3x the average CAC for those customers and maintain the same LTV:CAC ratio. This is why predictive LTV is one of the highest-leverage analytical investments for subscription and repeat-purchase businesses.
Predictive LTV model inputs (early behavioural signals with highest predictive power): product category of first purchase (some categories have structurally higher repeat purchase rates); number of items in first order; time between first and second purchase; email engagement rate in first 14 days; on-site session frequency in first 30 days; discount usage on first purchase (heavy discounters tend to have lower LTV); referral source and acquisition channel.
Model validation: compare predicted LTV scores at day 30 against actual LTV at day 180 and day 365 for historical cohorts. A well-calibrated model should rank customers correctly in terms of LTV quintile even if the absolute prediction is imprecise. The goal is not a perfectly accurate point estimate — it is a reliable rank ordering that enables differential bidding and audience targeting. Model accuracy benchmarks: Spearman rank correlation between predicted and actual LTV above 0.5 indicates useful predictive power; above 0.7 is strong.
Most businesses optimise acquisition on first-order metrics by default, not because they believe it is the right measure, but because it is the only metric available at acquisition time. Predictive LTV changes what is available. The Growth Operating System, developed by David Manela, uses predicted LTV as an input into the Capital Allocation Loop where data maturity supports it — enabling bid strategies on Meta and Google that weight high-predicted-LTV customer segments more heavily, and excluding audience segments with systematically low predicted LTV from prospecting budgets.
The data requirement for predictive LTV is significant: minimum 2–3 years of customer purchase history with sufficient cohort volume to train and validate a model. Businesses with less data history can use simpler heuristics (e.g. acquisition channel as a proxy for expected LTV) rather than a full machine learning model. Exactius assesses data maturity in the first 30 days of an engagement and builds the most sophisticated LTV-based bidding approach the data supports.
→ Learn more about the Growth Operating System at davidmanela.com/frameworks/growth-operating-system
What data do you need to build a predictive LTV model?
Building a reliable predictive LTV model requires: at minimum 2 years of customer purchase history (to observe LTV outcomes across enough time periods to train on); a minimum of 10,000–20,000 customer records with complete behavioural data across multiple touchpoints (first purchase category, session frequency, email engagement); and a persistent customer identifier that links acquisition source data to downstream purchase and retention data. Without the data volume and history, even a sophisticated model will produce unreliable predictions. Businesses with less than 2 years of data or under 10,000 customers should use cohort-based heuristics — such as acquisition channel LTV averages — as a simpler and more reliable approach.
How does predictive LTV improve ad bidding?
Predictive LTV improves ad bidding by enabling the platform's algorithm to bid differently for different audience segments based on their expected value, not just their probability of making a first purchase. Google's Target ROAS and Meta's Value Optimisation bidding strategies both accept customer value signals — if you feed predicted LTV scores as the conversion value rather than order revenue, the algorithm optimises for the audiences and creatives that attract high-LTV customers. The result is a bid strategy that spends more to acquire high-LTV customers and less on low-LTV customers, improving overall LTV:CAC even if the average CPA increases.
What is the difference between predicted LTV and historical LTV?
Historical LTV is a backward-looking calculation: the sum of revenue or contribution margin a customer has already generated over their full customer lifetime. It is precise but delayed — you cannot know a customer's historical LTV until their relationship with your business is far enough along to observe meaningful retention behaviour. Predicted LTV is a forward-looking estimate generated at or near the point of acquisition, based on early signals. It is less precise than historical LTV but immediately actionable. The value of predicted LTV is that it enables acquisition decisions to be made based on expected future value rather than waiting months or years for the historical LTV to materialise.
Exactius Growth
Your growth system is either compounding or degrading.
Book a diagnostic call. We'll identify where your growth system is breaking.
Book a call