How Torix Predicts Subscriber Value (pLTV)

Torix Growth · Internal explainer · July 2026

How we predict what a new subscriber is worth

When our new ad-reporting pipeline launches, every subscription will be reported to Meta with a dollar value attached. This page explains where that dollar value comes from, how the predicted lifetime value (pLTV) calculation works, and when it switches on.

Built & validated in testing — not yet live in production Day one reports actual prices; predictions switch on later

All figures on this page are computed from real Torix payment records as of 12 July 2026.

Average revenue per new subscriber, first 90 days

$26.15

across all plans and countries

Subscribers with 90+ days of payment history

312

the learning base grows every day

Value window the model predicts

180 days

expected revenue in a subscriber's first 6 months

Part 1

The problem, in one chart

On the day someone subscribes, every weekly subscriber pays the same $10.99 — so to Meta they all look identical, even though they are worth wildly different amounts:

Two real weekly subscribers, same first payment

Actual revenue collected over their first 90 days

Cancelled after week 1the most common outcome
$10.99
Stayed 13 weeksthe best outcome so far
$142.87
Both look identical to Meta on day one — a $10.99 purchase. One turned out to be worth 13× the other. If Meta never learns the difference, it can't find us more of the good ones.

Meta learns at the moment of purchase — waiting months to find out who was valuable is too late to be useful. So pLTV does one thing: at the moment someone subscribes, we report our best prediction of what they'll be worth, instead of just their first payment. (At launch the pipeline starts by reporting actual prices — accurate and fast, delivered within minutes of the purchase — and the prediction switch flips later; see Part 5.)

Part 2

How the prediction is calculated

There's no black box here. The prediction is a track-record average: we look at what similar subscribers actually went on to pay us, and use that as the expectation for each new one. Four steps:

Group past subscribers by two traits

Which plan they bought (weekly / monthly / yearly) and which country they signed up from. Two traits, because they're what we reliably know within seconds of the purchase — and they already separate value well.

Measure what each group actually paid us

For every subscriber in a group, add up all the real, non-refunded payments they made in their first 180 days — the first payment plus every renewal. Only subscribers whose full 180 days have already passed are counted, so the answer is never a guess about the future; it's a record of the past.

The group's average becomes the prediction

A new subscriber is predicted to be worth the average 180-day revenue of the group they match. Individually some will beat it and some will fall short — that's expected. Meta learns from thousands of reports, so what matters is that the average is honest.

Refresh monthly, score instantly

The group averages are recomputed on the 1st of every month from the latest history. When someone new subscribes, they're matched to their group and get their predicted value within a few minutes of paying.

The safety rules

Averages are only trustworthy with enough people behind them, so the model refuses to publish thin data:

What the payment history says right now

Average revenue per new subscriber over their first 90 days, by plan

Weekly282 subscribers
$20.44
Monthly8 subscribers*
$48.15
Yearly22 subscribers*
$91.34
Real production data, 12 July 2026, 90-day window. *Lighter bars mark groups still below the 25-subscriber minimum — today the model would treat them with its fallback rules rather than trust the raw average.

Broken into the full groups the model uses, the picture today looks like this:

Group (plan · country) Subscribers Avg revenue, first 90 days
Groups with enough people (25+)
Weekly · United States 228 $21.20
Weekly · Rest of world 54 $17.26
Still too small — safety rules apply (shown for transparency)
Yearly · United States 19 $89.99
Monthly · United States 5 $47.99
Monthly · Rest of world 3 $48.41
Yearly · Rest of world 3 $99.87
Safety-net averages (always available)
Any weekly subscriber 282 $20.44
Any monthly subscriber 8 $48.15
Any yearly subscriber 22 $91.34
Any subscriber at all 312 $26.15

Part 3

A worked example

Once predictions are switched on

A new customer in Texas subscribes to the weekly plan. Within minutes, here's what happens:

$10.99 what they actually paid today
Weekly · US the group they match; 228 previous subscribers averaged $21.20
$21.20 the value Meta is told this purchase is worth

Why is an average of very different outcomes fair? Because it is the honest expected value. Here's the actual spread behind that $21.20 — all 228 US weekly subscribers with 90 days of history:

One group, up close: Weekly · US

228 subscribers, by how many weekly payments they made in their first 90 days

0
50
100
150
155
5
012345678910111213

Weekly payments made in the first 90 days  ·  vertical scale: number of subscribers

Most stop after one payment ($10.99); a long tail keeps paying — five subscribers made all 13 possible payments (≈$143 each). The honest expectation across all 228 is $21.20. Hover or tab onto any column for its exact numbers.
View as a table
Payments madeSubscribersRevenue each
0 (refunded)4$0.00
1155$10.99
233$21.98
310$32.67
47$43.96
55$54.95
63$65.94
82$87.92
93$98.91
101$109.90
135$142.87

Part 4

One rule: never send Meta the same number for everyone

Meta learns by comparing purchases. If every subscriber were reported at the overall average, Meta would see “values” but have no way to tell a great subscriber from a poor one — and value-based targeting would learn nothing:

✕  Useless to Meta — one flat average

$26$26$26$26$26

every subscriber looks identical — nothing to compare

✓  Useful — each group's own number

$17$20$21$48$91

these five are real numbers from today's group table — Meta can now tell who's worth chasing

Who gets which number?

Assignment is mechanical — no judgment calls. Two facts, known within seconds of the purchase, decide it: the plan they bought and the country they bought it from. The model then walks a short checklist: use their exact group if it has enough history behind it; otherwise, step to the nearest safety net. Five new subscribers, minutes after paying:

A new subscriber… How the model decides Meta hears
Weekly · Dallas, US “Weekly · US” has 228 people behind it — plenty. Their average applies. $21.20
Weekly · São Paulo, Brazil Brazil doesn't have 50 subscribers of its own yet, so it counts as “Rest of world” (54 people). $17.26
Monthly · New York, US “Monthly · US” exists but has only 5 people — below the 25-person minimum, so it isn't trusted. Falls back to the average of all monthly subscribers. $48.15
Yearly · Berlin, Germany Yearly groups are still small everywhere, so the all-yearly average applies. $91.34
Anything brand new A subscription with no history behind it at all (say, a just-launched plan) gets the all-subscriber average — there is always an answer, never a blank. $26.15

Note what the fallbacks do to the numbers: they get less specific, not less honest — each step widens the circle of comparable subscribers until there are enough of them to trust. As history accumulates, more subscribers land in their exact group instead of a safety net, and the predictions sharpen on their own. And a built-in check requires at least 5 different prediction levels in play (Meta's own minimum is 2) — if the data ever collapses toward one number, the predictions can't be switched on, and would be switched back off.

Part 5

Where this stands, and what happens next

The entire machinery — reporting purchases to Meta, building the group averages monthly, scoring each new subscriber within minutes — is built and validated end-to-end in our testing environment, including against Meta's own test tools. It has not launched in production yet. When it does, what Meta receives is controlled by a single switch: launch sends the actual price, and flipping it to predicted value is a one-line, instantly reversible change.

Beyond the production launch itself, two things gate that flip:

1 · The data needs to finish maturing

Our clean payment history begins 15 February 2026. The model only trusts subscribers whose full 180-day window has already played out — and the very first payers only reach that mark in mid-August. Every month after that, the learning base grows and the groups get sharper.

15 Feb 2026

Clean payment history begins

Today · 12 Jul

Pipeline validated in testing; production launch ahead

~14 Aug 2026

First subscribers complete a full 180-day window

1 Sep 2026

First monthly model build with real 180-day groups

2 · Meta needs to confirm the fine print

Torix is in Meta's pLTV beta program. Before we send predictions, Meta's partner team needs to confirm exactly how they want predicted values delivered and revised. Until they do, we deliberately hold on real prices — sending predictions under the wrong rules could hurt campaign learning rather than help it.

The safety net, restated

When the switch does flip: the change is rehearsed in the testing environment first, results are watched for a week on Meta's own quality dashboards, and rolling back means changing one line back — Meta simply starts receiving actual prices again, exactly as at launch. No customer ever pays a different amount because of any of this; it only changes what we report about ad performance.

Questions we expect

FAQ

Does this change what customers are charged?

No. Billing is completely untouched. This only changes the value we report to Meta when telling it an ad led to a purchase — which in turn shapes who Meta shows our ads to.

What if the prediction is wrong for a specific person?

Individually it often will be — someone predicted at $21.20 might cancel after one week or stay six months. That's fine: Meta learns from thousands of reports, not one, so what matters is that the group averages are honest. That's also why groups need at least 25 subscribers before their average is used.

Why not just keep reporting the actual price?

Because within a plan, the actual price makes everyone identical — every weekly subscriber is a $10.99 purchase, whether they turn out to be worth $10.99 or $142.87. Predictions restore that difference at the moment Meta actually learns from it, so its targeting can chase high-value subscribers instead of just any subscribers.

Why 180 days and not lifetime?

A window has to be long enough to capture real renewal behavior but short enough that history “completes” quickly — with a true lifetime measure we'd wait years before having a single trustworthy number. Six months captures the bulk of the value differences between subscribers while letting the model start learning this year.

What about coin purchases?

Coin packs are one-off purchases, not subscriptions, so there's nothing to predict — they're always reported at the actual price paid. Predictions apply to subscription starts only, and each subscriber is scored once, at the moment they first pay (renewals don't re-trigger it).

Why do the numbers on this page use 90 days, not 180?

Honesty. No subscriber has 180 days of history yet (payment records begin 15 February 2026), so a 180-day table would be empty today. The 90-day figures shown run the exact same calculation on the exact same real data — just with the shorter window the data can already support. The live model will use 180 days.

Sources: all dollar figures, subscriber counts, and distributions computed on 12 July 2026 directly from Torix production payment records (RevenueCat transaction history), using the same rules as the live model — refunds excluded, test purchases excluded, one lineage per subscription. The 90-day window is illustrative; the production model predicts a 180-day window. Group thresholds (50 per country, 25 per group), monthly rebuild schedule, tier design, and rollback behavior are as implemented in the ads-egress service.