Stack · Published 2026-07-15 · 3 min read
How I work with neural networks. Part 2 of 4: Monday predictions
The full description of the infrastructure is in the previous post.
Every Monday morning all my clients are guaranteed a prediction — a forecast of the future ROAS of the week just gone. It costs me no more than five minutes.
Why a prediction at all
At any moment we can work out day-zero ROAS: LTV on day 0, multiplied by conversions, divided by spend.
But D0 ROAS says almost nothing, because the money from subscriptions and renewals arrives later. The client does not need to know how much it has paid back today, but what it will grow into in a month or in six. So what is needed is closer to: LTV at 180 days, multiplied by conversions, divided by spend.
Which metrics I look at
How to compute LTV in the formula above depends on the monetisation model.
- ARPPU
A direct subscription with no trial.
- ARPAS
A subscription with a trial.
Both parameters are available as they are in Apphud and Adapty.
- A fixed value
Sometimes a project is one or two days old and carries a pile of campaigns that make no sense yet, and you still want to judge whether they work. Even without digging into revenue analytics you can build a prediction by taking some fixed LTV. For example: "we see LTV around $30, we assume it will grow to $50 because that is what competitors get, so we take $50 as the final LTV".
That is not accurate, but a poor prediction beats no prediction at all.
How the prediction is calculated
It depends on the age of the cohort.
- A young cohort
We look at how the LTV curve grew over the past month and build a logarithmic interpolation. The task is to fit a logarithm that passes through every known point and lets the data be extrapolated. This is where neural networks show their worth: you can use elaborate algorithms that you would be too lazy to compute by hand.
- An old cohort
If the cohort is old enough, the prediction is a single coefficient. For example: ARPAS on day 3 is 5 and on day 180 it is 8. Divide 8 by 5 and you get 1.6. Multiply the day-3 ARPAS of the new cohort by 1.6 and you have its LTV on day 180.
Two modes: automatic and manual
- Full automation
The client has a Meta account and Adjust, and MCP is connected to both. The system fetches spend and cohorts itself and computes the prediction without me.
- Manual mode
If the client's analytics cannot be connected over API — they are on Apphud, say, or there is no MCP access — I export a single CSV, hand it to the bot, and it builds the prediction.
The takeaway
AI removes the routine: gathering the data, the arithmetic, the formatting, the sending. Monday, morning, report with the client. Five minutes of my time.
If you want to assemble the same workflow yourself, setting it up from scratch is exactly what I cover in the AI module of User Acquisition — from getting the tokens to the first campaign launched through Claude Code. The module can be bought on its own.
Questions by direct message: @SiberianPaaanda