Why it matters
Raw model output is rarely safe to send unchanged. Uncapped highs can dominate learning and trigger volatile CPAs. Compressed lows make value optimization resemble conversion maximization. Wrong timing (value before the anchor event, or too late for signal freshness) wastes signal volume.
Transformation encodes business judgment: net vs gross revenue, margin floors, promo skepticism, and how aggressively to rank users before outcomes mature. It is where signal optimization becomes concrete policy, not only data science output.
Teams that skip transformation often conclude "pLTV doesn't work" when the model was fine but the bidding input was not platform-shaped.
Signal transformation
Transformation is a required layer in Churney-style activation:
- Model: Generate user-level pLTV from first-party data in the data warehouse.
- Calibrate: Align rank and scale on mature cohorts (calibration panels).
- Transform: Apply rules for magnitude, caps/floors, event timing, and optional conservative bias for uncertain early events.
- Deliver: Send transformed values server-side with match-ready identifiers.
- Measure: Holdout tests vs BAU on incremental ROAS and cohort deciles; adjust when model drift appears.
Transformation interacts with user value re-ranking: spread must be wide enough to differentiate users, bounded enough to keep signal quality high.
Category variants
| Vertical | Transform focus | Example rule |
|---|---|---|
| Ecommerce | Net revenue, refund haircuts | Cap first-order value pending return window |
| Subscription | Trial vs paid weighting | Lower trial-start values until payment |
| Mobile app | Multi-horizon IAP blend | Floor for non-payer scores on ad-supported users |
Common mistakes
- Sending raw model scores without business margin alignment.
- No caps on outliers, destabilizing platform learning.
- Transforming once at launch without iteration after mix shifts.
- Timing value on the wrong event, breaking optimization window alignment.
- Over-compressing spread, eliminating re-ranking benefit.
Advertiser lens
| Role | Cares about |
|---|---|
| Data science | Calibrated inputs and stable score distributions post-transform |
| UA / performance | Predictable CPA/ROAS response to value changes |
| Finance | Margin alignment of sent values |
| Engineering | Rule versioning and pipeline reproducibility |
FAQ
Is transformation the same as calibration?
Calibration aligns predictions to realized outcomes; transformation applies business and platform rules to calibrated scores.
Who owns transformation rules?
Joint: finance defines margin logic, data science implements, UA validates bidding impact, engineering deploys.
Can transforms fix bad model rank?
Partially via monotonic adjustments, but rank failure usually needs model work, not only caps.
How do you test a transform change?
Run a holdout test or structured experiment; compare cohort LTV by sent-value decile vs BAU.
Do platforms see transform logic?
No. They receive final event values and timestamps you send.
What tools apply transformation?
Typically data warehouse SQL or feature pipelines, then activation layer (Churney) before Meta CAPI / Google Ads Conversion API.
Not the same as
| Term | Difference |
|---|---|
| Feature engineering | Model inputs, not platform-facing values |
| ETL mapping | Schema plumbing, not economic bidding policy |
| Bid strategy | Budget rules in ads UI, not conversion value payload |
