How Opal achieved 33% higher M0 ROAS in a Meta test using predictive signals
Day30 used machine learning to score users from onboarding responses and their first six hours of in-app behaviour, helping Meta identify likely paying subscribers while keeping cost per trial comparable.
+33%
First-month net ROAS (M0)
+32%
Trial-start to paid conversion rate
Comparable
Cost per trial vs. control
Measurement note: All figures are relative differences between the test and budget-matched control during this pilot. M0 measures net revenue generated during the first 31 days after trial start, after app-store fees and taxes.
Opal
The acquisition probleM
Trial starts offered volume without sufficient information about customer quality. Purchases offered quality without enough event volume.
Opal's challenge was not generating trials. It was helping Meta distinguish a trial starter from a user likely to pay. As spend scaled, trial volume rose while trial-to-paid conversion fell, making trial starts a less reliable proxy for customer value.
Trial starts gave Meta enough volume to learn, but not enough information about who would ultimately pay. Some subscription routes did not produce the same trial-start event at all.
Opal had also tested a paid-trial optimised campaign. The event fired correctly, but conversions were too sparse for the campaign to exit Meta's learning phase reliably.
Opal needed something between the two: an early event frequent enough for Meta to use, but selective enough to identify users with a materially greater likelihood of becoming paying subscribers.
THE DAY 30 APPROACH
Day30 first strengthened Opal's measurement foundation, then used machine learning to turn six hours of user data into a tunable optimisation signal.
Before building the predictive signal, Day30 reviewed Opal's measurement foundations across Adjust and its advertising-platform integrations. The work covered attribution settings, event mapping and reporting consistency, creating a clear baseline for net M0 ROAS using Adjust last-touch attribution. Only then did predictive modelling and live testing begin.
Day30 trained a machine-learning model to estimate each new user's probability of converting from trial to paid. The model scored users individually, combining their onboarding responses with behaviour observed during the first six hours and accounting for Opal's different trial lengths and pricing structures.
Those probabilities were converted into an optimisation event using a tunable threshold. Raising the threshold produced a smaller, more precise signal population; lowering it generated more events for Meta to learn from. Day30 calibrated this dial to create an event that was materially more selective than a trial start while still occurring frequently enough to support campaign learning.
The signal was routed through Opal's existing Adjust integration. No new Day30 SDK was required.
THE PILOT
Audience, budget, creative, dates and cohort horizons were held constant; the optimisation signal was the deliberate difference.
Day30 and Opal evaluated the new signal through a budget-matched A/B test across three comparable markets.
Both arms ran over the same dates with the same audience, daily budget and creative. Creative winners graduated into both arms at the same time. The control continued optimising towards Opal's existing event; the test arm used Day30's engineered signal.
The acquisition window ran for approximately 30 days. Both groups then completed the same 31-day M0 revenue horizon before the result was evaluated.
The PILOT RESULT
The test did not win by finding cheaper installers. It won by finding users who were substantially more likely to become paying subscribers.
Cost per install was 20% higher in the test arm. Looking only at installs would therefore have suggested that the new signal was less efficient. The downstream results told a different story.
Measure
Test vs. control
What it shows
Net M0 ROAS
+33%
More net revenue during the first 31 days after trial start.
Trial-to-paid conversion
+32%
The acquired cohort was materially more likely to become paying subscribers.
Cost per trial
Comparable
The improvement was achieved without paying materially more to generate a trial.
Cost per install
+20%
The test paid more per installer but acquired a higher-quality customer cohort.
Read together, the metrics explain the pilot results. Day30's signal did not make each install cheaper; it helped Meta acquire a cohort that converted to paid at a substantially higher rate. That improvement was large enough to produce a 33% increase in net first-month ROAS within the pilot.
The 73% increase in trial-to-paid conversion supports the ROAS result: the test was acquiring more valuable customers, not simply assigning more value to the same users.
“These results came from the model I build and monitor for Opal's own internal reporting - not a cherry-picked data source. The pilot result was also visible in our own analytics.”
Yunshu Yang, OpaL
THE FOUR pillars
Why it worked
"I saw the pilot results reflected in our internal data. That gave me confidence to test Day30's predictive signal across more networks and geographies."
Yunshu Yang, OpaL
