Predictive LTV
Why this matters for your business
Customer lifetime value, until recently, was a historical metric. "This customer has spent $480 with us." Useful for analysis, useless for spending decisions today — by the time you know a customer was worth $480, you've already spent the same on every other customer in their cohort, including the ones who turned out to be worth $40.
Predictive LTV flips this. Using each customer's purchase behaviour, engagement signals, and cohort patterns, the model forecasts what they'll spend in the next 90 days — and in some segments, the next 12 months. That number is available today, on every customer, to drive every spending decision: ad bid caps, loyalty thresholds, win-back budgets, VIP tiering, and even when to send a hand-written follow-up from the founder.
The shift in mental model is what matters. You stop treating customers as a uniform list and start treating them as a portfolio with different expected returns. Top-decile customers might be worth 40× the bottom-decile. Acting on that asymmetry — giving the top-decile more attention, more budget, more care, and the bottom-decile less — is the highest-leverage decision in e-commerce marketing.
What this typically unlocks
| Outcome | Typical lift |
|---|---|
| ROAS on retargeting | +22% — bid more on top-LTV cohort, less on bottom |
| Customer-acquisition payback period | −35% — focus acquisition on look-alikes of highest-LTV cohort |
| Win-back budget efficiency | +58% — only spend on customers worth winning back |
| Loyalty-tier eligibility accuracy | replaces "≥ 3 orders" with predicted future value |
| VIP outreach hit-rate | +3.4× — who you call matters more than what you say |
| Subscription churn prediction | +0.32 AUC vs. naive recency-only models |
What you actually get
Three forecasts per customer, refreshed weekly:
| Metric | What it tells you | Refresh |
|---|---|---|
| Predicted LTV (next 90d) | Expected revenue from this customer in the next 90 days | Weekly |
| Predicted LTV (next 12mo) | Long-tail expectation; useful for high-AOV / subscription | Weekly |
| Confidence interval | How tight the prediction is (e.g. "$120 ± $30") | Per prediction |
Plus segment-level rollups:
- Top-decile / quartile / median LTV by cohort
- LTV by acquisition channel — which channels deliver high-LTV customers
- LTV by first product — which products predict high-LTV repeat purchases
- LTV trend over time — is your customer base getting more or less valuable?
How it powers every part of your store
| Where predictive LTV moves the needle | Mechanism |
|---|---|
| Ad bid caps | Pay more to retarget top-decile LTV; pay less for bottom |
| Lookalike audiences | Build Meta/Google lookalikes seeded only on top-LTV customers |
| Loyalty tier thresholds | "VIP = predicted 12mo LTV > $500" beats "≥ 3 orders" |
| Win-back budgets | Only spend acquisition budget on win-back targets worth > $100 predicted |
| VIP human outreach | Trigger founder/CSM email to customers with predicted 12mo LTV > $1K |
| Subscription save-offers | Offer larger discounts to higher-LTV customers about to cancel |
| Inventory planning | Forecast cohort-level demand ahead of stocking |
| Investor reporting | Cohort LTV is the cleanest growth signal for diligence |
How it works (without the technical bits)
The model considers four families of signals per customer:
What's in each signal family
- Purchase history. How recently, how often, and how much the customer has bought. Plus the gap pattern between orders — a customer who buys reliably every 30 days is more predictable than one with random gaps.
- Engagement signals. Email opens, WhatsApp replies, site visits, support touches. Customers who engage between purchases predict much higher LTV than dormant ones with identical purchase history.
- Cohort patterns. "Customers like this one" — same acquisition channel, same first product, same locale — historically had a particular revenue trajectory. The model uses that as a Bayesian prior.
- Seasonality + trend. Your shop's natural cyclicality (Q4 spike, summer slump, etc.) shapes the prediction. A customer's predicted 90d in October is computed differently from the same customer's predicted 90d in February.
The math, lightly
Two models combine: BG/NBD (predicts how many orders the customer will place) and Gamma–Gamma (predicts the average value of those orders given the count). Multiply for total predicted revenue. Both are well-established in academic e-commerce literature; we tune them on each shop's data weekly so the predictions reflect your specific business, not a generic model.
You don't need to know any of that to use the output — just "this customer has a predicted next-90d revenue of $148 ± $32." The bracket is the model's confidence interval, which is itself useful: customers with tight intervals are predictable; wide- interval customers are noisy and warrant less aggressive spending.
Calibration — why you can trust the number
Each week we measure: customers we predicted would spend $X in the next 90 days — what did they actually spend? We track calibration metrics:
| Metric | What it means |
|---|---|
| Mean absolute error (MAE) | Average $ deviation between predicted and actual |
| Calibration plot | Predicted vs. actual binned at deciles — should fall on the 45° line |
| Top-decile capture | Of the top 10% predicted, what fraction were actually top 10% by spend? |
| Bottom-decile capture | Of the bottom 10% predicted, what fraction were actually bottom 10%? |
You see these in the LTV admin tab. If calibration drifts (e.g. the model is 30% too optimistic for a quarter), the system flags it and re-trains. Predictions you can't trust are worse than no predictions at all.
When the model isn't confident
Three cases where the model abstains:
- New customers with < 2 orders. Returns "insufficient data" — these customers get a cohort-level fallback estimate based on their first product + acquisition channel.
- Outlier customers whose pattern doesn't match anyone else's. Returns a wide confidence interval and flags for review.
- Shop-level data shortages (< 1,000 orders/month). Returns shop-cohort estimates rather than per-customer; you'll see "predictions enable at 1,000+ orders/month" in the UI.
Knowing when not to trust a prediction is half the value.
Real merchant scenarios
Scenario A — DTC brand reallocates ad budget
Setup. $6M/year skincare brand. Spending $80K/month on Meta retargeting at a flat bid cap. Looking for ROAS improvement without increasing budget.
Decision. Used predictive LTV to bucket the retargeting audience into 4 deciles. Set tiered bid caps:
| Decile | Predicted 12mo LTV | Bid cap | Audience size |
|---|---|---|---|
| Top 10% | $480+ | $14 CPM | 4,200 |
| Next 20% | $220–$480 | $9 CPM | 8,400 |
| Middle 50% | $80–$220 | $5 CPM | 21,000 |
| Bottom 20% | < $80 | $0 (excluded) | 8,400 |
90-day result.
| Metric | Before | After |
|---|---|---|
| Total Meta spend | $80K/month | $80K/month (flat) |
| Customers reached | 42K | 33.6K (focused) |
| Attributed revenue | $312K | $409K |
| ROAS | 3.9× | 5.1× (+31%) |
The bottom 20% being excluded was where the win came from — those customers were costing money to retarget and almost never buying again. Predictive LTV identified them.
Scenario B — Subscription brand redesigns save-offers
Setup. Subscription box, 24K active subs. Every cancel attempt prompted a "10% off your next box" offer. Boring, expensive, low-conversion.
Decision. Tier the save-offer by predicted 12mo LTV:
| LTV tier | Save-offer |
|---|---|
| $300+ predicted | "Pause for 1 month, free; no charge until you choose" |
| $100–$300 predicted | "10% off next 3 boxes" |
| < $100 predicted | "Sorry to see you go" — no offer |
Outcome.
| Metric | Result |
|---|---|
| Save-offer redemption rate | 22% (was 7%) |
| Avg discount given per save | $4 (was $11) |
| Subscriptions saved/month | +180 |
| Revenue impact (12mo) | +$420K |
| Margin impact (vs. flat 10% off everyone) | +$95K saved |
The "no offer for low-LTV" tier was uncomfortable but right — those customers had low future value and the brand was spending margin to keep them.
Scenario C — Loyalty redesign
Setup. Apparel brand, traditional loyalty program: 3 orders = "Silver", 6 orders = "Gold", 12 = "Platinum". Tiers based on historical order count.
Problem. Silver customers and Gold customers had nearly identical predicted future value — the order-count threshold wasn't capturing actual customer value. Meanwhile, some 2-order customers had massive AOVs and high engagement (predicted to be huge in the future) but were stuck at no-tier.
Decision. Switch loyalty thresholds to predicted 12mo LTV:
| Tier | Old rule | New rule | Result |
|---|---|---|---|
| Silver | 3+ orders | Predicted LTV $200+ | -12% members, +18% avg per-member spend |
| Gold | 6+ orders | Predicted LTV $500+ | -18% members, +26% avg |
| Platinum | 12+ orders | Predicted LTV $1,200+ | -22% members, +44% avg |
90-day result.
- Total members in tiers down 16%.
- Total revenue from tiered customers up 24%.
- Cost of loyalty rewards (free shipping, exclusive access) up only 9% — concentrated on customers who actually warranted it.
- Brand-positioning win: Platinum became genuinely exclusive (~0.3% of base rather than 4%).
Scenario D — Founder-led VIP outreach
Setup. Bootstrapped DTC brand, founder still personally emails certain customers. Wants to know which 100 customers/ month to email for highest impact.
Old method. Top 100 by historical spend. Took 8 hours/month; half the recipients hadn't bought in months and didn't reply.
New method. Top 100 by predicted 12mo LTV minus customers who'd received a personal email in the last 60 days.
90-day result. The same 8 hours/month produces:
| Metric | Old | New |
|---|---|---|
| Reply rate | 14% | 47% |
| 60-day re-purchase | 28% | 61% |
| Avg revenue per recipient | $84 | $238 |
Same effort, ~3× the impact. The founder still picks the copy and the personal touch — predictive LTV picks who.
Scenario E — Acquisition channel optimization
Setup. $12M brand running 6 acquisition channels (Google Ads, Meta, TikTok, influencer, podcast, organic search).
Question. Which channel to scale, which to cut?
LTV-by-first-touch analysis (12-month look-ahead):
| First touch | Customers | Avg predicted 12mo LTV | CAC | LTV:CAC |
|---|---|---|---|---|
| Google Ads | 12,400 | $186 | $42 | 4.4× |
| Meta acquisition | 14,200 | $98 | $48 | 2.0× |
| TikTok | 6,800 | $74 | $35 | 2.1× |
| Influencer | 3,100 | $244 | $62 | 3.9× |
| Podcast | 1,200 | $312 | $89 | 3.5× |
| Organic search | 4,400 | $208 | $0 | ∞ |
Decision. Doubled influencer + podcast budgets. Cut Meta acquisition by 30%. Maintained Google. Result: total customer count down 8% the next quarter; revenue up 14%, profitability up 22%.
The lesson: getting more customers is easy; getting more high-LTV customers is hard. Predictive LTV makes the second question answerable.
Scenario F — Catching a churn cohort early
Setup. Mid-market brand noticed predicted 90d LTV declining across one specific segment ("repeat buyers, ages 25-34, mobile- first") quarter over quarter — even though current revenue looked fine.
The investigation. Engagement signals (email opens, site visits) had collapsed in this segment despite stable purchase counts. The model picked up the behavioural shift before the revenue shift would have shown up.
The cause. A site redesign 3 months earlier had broken the mobile checkout for this cohort's primary device (a specific older Android browser). Customers were buying once out of habit, then quietly leaving.
The fix. Mobile-checkout fix shipped in 2 weeks. Predicted LTV for the cohort recovered over the next 8 weeks. Revenue didn't drop because the brand caught it 3 months early — a problem that would have surfaced as a Q3 revenue miss instead showed up as a Q1 prediction-decline.
This is the pattern: predictions are early-warning systems. By the time historical revenue tells you something is wrong, you've already lost the quarter.
Best practices
✅ Use predicted LTV in deciles, not absolute thresholds. "Top 10% by predicted LTV" is robust to the model recalibrating; "customers with predicted LTV > $250" can drift if the model shifts.
✅ Pair with confidence intervals. A high prediction with wide CI is much less actionable than a moderate prediction with tight CI. Use both numbers.
✅ Recompute weekly, act monthly. The data refreshes weekly but you should set the bid cap / loyalty threshold / budget allocation monthly — week-to-week noise is real.
✅ Watch the calibration metrics. If MAE drifts >20% or calibration plot moves off the 45° line, surface to support. The model is updated continuously but unusual shop changes (major product line shifts, big M&A) can throw it off.
✅ Combine with churn risk. A customer with high LTV and high churn risk is the highest-priority intervention target. A customer with high LTV and low churn risk is on autopilot.
❌ Don't use predicted LTV for individual customer treatment in fully-automated pipelines. Use deciles or tiers. A customer the model thinks is worth $1,200 but is really worth $50 will get a wildly inappropriate offer if you act on the single number.
❌ Don't use predicted LTV for new customers (< 2 orders). The model abstains here; respect it.
❌ Don't aggregate predicted LTV without weighting by confidence. "Total predicted revenue across customers" is meaningful; "average predicted LTV" can be skewed by a few high-uncertainty outliers.
Plan tiers
| Capability | Free | Starter | Pro | Agency | Enterprise |
|---|---|---|---|---|---|
| Predicted 90d LTV | — | — | ✓ | ✓ | ✓ |
| Predicted 12mo LTV | — | — | ✓ | ✓ | ✓ |
| Confidence intervals | — | — | ✓ | ✓ | ✓ |
| LTV-by-channel rollup | — | — | ✓ | ✓ | ✓ |
| LTV-by-product rollup | — | — | ✓ | ✓ | ✓ |
| Cohort LTV trends | — | — | ✓ | ✓ | ✓ |
| Calibration dashboard | — | — | ✓ | ✓ | ✓ |
| Segment-level forecasts | — | — | ✓ | ✓ | ✓ |
| Programmatic API access | — | — | ✓ | ✓ | ✓ |
| Multi-shop LTV roll-up | — | — | — | ✓ | ✓ |
| Custom model training | — | — | — | — | ✓ |
Available from Pro tier and above. The model needs ~1,000+ historical orders to produce per-customer predictions; below that, the system falls back to cohort-level estimates and shows a clear "predictions improve at order volume X" prompt.
Frequently asked
How accurate is the prediction? Calibration metrics are visible in the LTV tab. For a typical shop with > 5K orders/month, MAE is around 15–25% of the predicted value — meaning a customer predicted to spend $200 in the next 90 days actually spends $200 ± ~$40 on average. Customers with longer history get tighter intervals.
What if my shop has seasonal spikes? The model includes seasonality. Predictions in October vs. February explicitly account for your shop's holiday cycle.
Can I predict LTV for new customers? For customers with 0–1 orders, the model returns a cohort-level estimate (based on first product, acquisition channel, locale) rather than a per-customer prediction. Better than nothing, but treat as directional.
Can I see why a customer's prediction is what it is? Yes — the customer detail page has a "Why this prediction?" panel showing the top contributing signals (e.g. "high opens in last 30d", "subscription cadence stable", "cohort historically 12mo LTV $480").
How do I export predictions for use in Meta/Google?
Pro+ plans include daily push to Meta/Google Custom Audiences
keyed on predicted LTV decile. You can also pull via the
v1 admin API at
/api/v1/customers?include=predicted_ltv.
What happens when I uninstall the app? Predictions stop refreshing. Stored predictions remain queryable via the GDPR export endpoint for 30 days, then deleted along with the rest of the customer profile.
Does the model use any cross-shop data? No. Each shop's model is trained on that shop's data only. This is a deliberate trade-off (small shops get less data) in favour of (a) privacy and (b) avoiding cross-shop signal contamination from very different categories.
See also
- Customer 360 — the profile predictions attach to
- Segments & cohorts — segment by predicted LTV decile
- Campaigns — bid + budget by LTV cohort
- Journeys & automations — VIP upgrade triggers on LTV thresholds
- Attribution & revenue — historical revenue (predictive LTV's input)
- Sales engine overview