Redesigning the SKAN conversion-value schema lifted ROAS prediction accuracy 2.4x

Redesigning the SKAN conversion-value schema lifted ROAS prediction accuracy 2.4x
An app encoded raw revenue buckets into SKAN's 6-bit value, but most users generate $0 by the postback timer, so the signal was almost all zeros.

They switched to encoding predicted-LTV tiers from a 0–24h behavioral model (sessions, tutorial, first purchase intent) instead of realized revenue.

Result: correlation between SKAN value and actual D30 ROAS rose 2.4x; UA bidding got materially sharper.

✓ Predicted-LTV encoding beats realized-revenue inside SKAN's tiny window
✓ Turned a mostly-zero signal into a useful one
✗ Model needs retraining as the game economy shifts
✗ Mis-specified buckets are hard to debug post-hoc

Verdict: encode predicted LTV, not realized revenue, into SKAN values.
Best for: apps whose monetization happens after the postback timer.