Predictive LTV vs Historical LTV
Historical LTV tells you what customers were worth; predictive LTV estimates what they will be worth. Mixing the two is the most common LTV reporting error — here's how to keep them in their lanes.
Predictive LTV vs Historical LTV
Two LTV methods: historical measures completed cohort revenue; predictive projects future revenue from early purchase signals.
Historical LTV is a backward-looking measurement. You take a cohort of customers who first ordered N months ago, sum what they actually spent, and divide by cohort size. The number is real, but it lags — a 24-month historical LTV describes customers acquired two years ago, under different products, prices, and ad mix.
Predictive LTV is a forward-looking estimate. From a customer's first 30-90 days of behaviour — order count, AOV, category, channel, return rate — a model projects what they will spend over a defined horizon (typically 12, 24, or 36 months). It's faster and decision-useful, but it's a probability, not a receipt. The two methods answer different questions and should never be plotted on the same axis without labels.
The choice between predictive and historical LTV isn't ideological — it's situational. Finance closing the year needs realised numbers. A growth team deciding next week's bid caps needs an estimate of what the customers they just acquired will be worth. Both are valid; conflating them is what causes payback periods to drift and channel mixes to misallocate.
Most LTV reporting errors trace to one of three confusions: comparing a 12-month historical figure to a 24-month predictive one, treating a predictive model's point estimate as a fact, or refreshing the historical number so rarely that it describes a business that no longer exists. The fix is structural — label every LTV figure with its method and horizon, and never average across them.
Predictive LTV vs historical LTV at a glance
| Dimension | Historical LTV | Predictive LTV |
|---|---|---|
| Direction | Backward-looking | Forward-looking |
| Data source | Completed transactions | Early signals + model |
| Time to first usable number | 12-24 months after acquisition | 30-90 days after first order |
| Accuracy | Exact for the window measured | Estimate with confidence interval |
| Decision use | Finance, year-end reporting, valuation | Bid caps, channel mix, cohort triage |
| Refresh cadence | Quarterly | Weekly or per-cohort |
| Failure mode | Describes a business that no longer exists | Confident wrong answer if model drifts |
| Audit trail | Reproducible from order data | Requires model versioning |
Read the table as a division of labour, not a competition. A mature LTV measurement practice runs both in parallel — historical as the ground truth that calibrates the model, predictive as the operational signal the growth team actually uses.
When to use historical LTV
Historical LTV is the right answer whenever the cost of being wrong is high and you have time to wait. Board decks, investor updates, year-end CAC payback calculations, and brand valuation all need realised numbers. If your apparel store is reporting LTV to a potential acquirer, no one wants to hear about an 87%-confidence projection.
It's also the only honest way to validate a predictive model. You run the predictions against a held-out cohort, wait until that cohort matures, then compare the predicted distribution to what actually happened. Without this calibration loop, predictive LTV becomes a story rather than a measurement.
The most common LTV reporting error
Comparing this quarter's predictive LTV against last year's historical LTV and concluding that LTV is rising. You're comparing a 24-month projection to a 12-month receipt. The number always looks better — because it's a different number. Always label the method and horizon; never put both on the same trend line without separating the series.
When to use predictive LTV
Predictive LTV earns its place in any decision that can't wait two years. Meta and Google bidding, Klaviyo flow prioritisation, VIP-tier qualification, and channel budget reallocation all benefit from an estimate available 30-90 days after acquisition. A beauty brand routing high-pLTV first-buyers into a richer welcome series doesn't need exact figures — directional rank is enough.
The trade-off is model risk. A predictive LTV model trained on 2022 cohorts may quietly drift as iOS attribution changes, promotional intensity shifts, or your product mix evolves. The mitigation is unglamorous: version every model, hold out a calibration cohort, and re-fit at least quarterly against fresh historical data.
Predictive LTV converges toward historical LTV as cohort matures
Predictive LTV (24-month projection)
Historical LTV (realised)
Predictive LTV vs historical LTV FAQ
Historical LTV is exact for the window it measures — there's no estimation error. Predictive LTV is an estimate with a confidence interval. But 'accurate' is the wrong frame: a perfectly accurate historical number that's 18 months stale can be more misleading for a current bidding decision than a predictive number with a ±15% confidence band.
Most predictive LTV models need at least 30 days of behaviour for a stable estimate, and 60-90 days for one that's tight enough to drive bidding. Before day 30 you're mostly projecting from the first order and acquisition channel, which is noisy. After day 90 the marginal accuracy gain from waiting drops sharply.
Yes, and you should — but label them explicitly: 'Historical 12-month LTV: €158. Predictive 24-month LTV for the most recent cohort: €182.' Putting them on the same line item without method and horizon labels is how finance ends up with two different LTV numbers in the same deck and nobody trusting either.
Order count and AOV in the first 30-90 days, acquisition channel, first-product category, discount depth on first order, geography, return rate, and email/SMS engagement. More sophisticated models add browse behaviour and category breadth. The first three inputs usually capture 70-80% of the explanatory power.
Shopify shows a 'predicted spend tier' on customer profiles, but it's a coarse three-bucket classification, not a euro value. For modelled pLTV with a confidence interval, you need a dedicated analytics layer that ingests order history and exposes the prediction as a customer attribute you can sync to Klaviyo or Meta.
At minimum quarterly, ideally monthly. The 24-month historical LTV from a year ago describes customers acquired three years ago — different products, different ad mix, different prices. If you only refresh annually, you're effectively making decisions on a number that's two product cycles old.
A well-calibrated model on a stable DTC business typically lands within ±10-20% of the eventual realised LTV at the cohort level, with wider error at the individual customer level. Anything claiming better than ±10% is either overfit or measuring a very predictable subscription-style business.
For operational payback monitoring — yes, with the caveat that you state the method. For contractual or board-reported payback — no, use historical. The risk of using predictive in board reporting is that a model recalibration silently changes the payback figure, and you can't reconstruct why.
Hold out a cohort, store the day-90 predictions, then wait until that cohort hits the prediction horizon and compare predicted vs realised at the cohort level. Track the mean error and the calibration curve quarterly. If mean error drifts above 15-20% or the calibration curve bends, retrain.
Only when the predictive model has been calibrated against the exact same horizon as the historical figure for a fully-matured cohort — and even then they'll differ by the model's residual error. If your pLTV and hLTV are reported as identical, someone has confused them. They should converge, not match.
Track CAC, channels, and funnel conversion in one place
Metricuno connects ad spend, funnel events, and revenue so you can see CAC by channel, cohort, and campaign — without stitching together five tools.