Why it matters
Platforms need signals at the anchor event, not six months later. Leading indicators bridge the gap between what is observable now and what finance cares about at cohort maturity. Without them, models default to demographics or shallow first-purchase proxies that break under promo mix or feedback loops.
Not every early action is a good indicator. Some spike conversion volume but anti-correlate with margin (heavy discount redemption, bracketing patterns). Signal optimization tests whether indicator-driven scores improve incremental ROAS in holdout tests, not only offline correlation.
Strong indicator sets are vertical-specific and must be monitored for model drift when product or acquisition strategy changes.
Leading indicators
Indicators feed the early score pipeline:
- Discover: Mine first-party data in the data warehouse for behaviors that predict the chosen prediction horizon.
- Model: Train user-level pLTV using leading indicators available at score time (no label leakage).
- Calibrate: Validate rank and scale on mature cohorts; compare to proxy metrics alone.
- Transform: Apply signal transformation so indicator-driven values are conservative and platform-ready.
- Activate and prove: Send via Meta CAPI, Google Ads Conversion API, or app paths; read out vs BAU with agreed maturity.
Churney uses indicator-rich models so bidding can learn on early value users without waiting for repeat purchases or renewals.
Category variants
| Vertical | Example indicators | Caution |
|---|---|---|
| Ecommerce | Category breadth, second order within 14d | Promo redemptions that attract one-and-done buyers |
| Subscription | Activation steps, trial engagement depth | Trial starts without payment intent |
| Mobile app | Tutorial completion, D1 retention, first IAP timing | Whale outliers distort calibration |
Common mistakes
- Treating any early conversion as a leading indicator without LTV validation.
- Using indicators unavailable at score time (leakage from future events).
- Overfitting promo periods that do not generalize.
- No drift monitoring when product onboarding changes.
- Replacing holdouts with offline correlation slides.
Advertiser lens
| Role | Cares about |
|---|---|
| Data science | Predictive power, leakage checks, feature stability |
| Product / growth | Which behaviors are worth incentivizing |
| UA / performance | Whether early scores move bidding outcomes |
| Analytics | Holdout design linking indicators to realized LTV |
FAQ
What makes a good leading indicator?
Available early, stable across seasons, and validated against realized LTV at the chosen horizon.
Are leading indicators the same as proxy metrics?
Proxies are often single events used for bidding; leading indicators are model features that may combine many behaviors.
How many indicators do you need?
Enough signal for rank, not so many that sparse data breaks scores; vertical and volume dependent.
Can indicators change after launch?
Yes, when product flows or acquisition mix shifts; revalidate calibration and run holdouts.
Do platforms see individual indicators?
Typically no; they receive aggregated value on conversion events you define.
How does Churney use leading indicators?
Data warehouse feature engineering feeds per-user pLTV, then calibrated values go to ad platforms server-side.
Not the same as
| Term | Difference |
|---|---|
| Vanity metric | Activity without proven LTV linkage |
| North star metric | Company-level success metric, not per-user bidding feature |
| Realized LTV | Outcome after maturity, not early predictor |
