• ▸4 prediction types · purchase · repeat · high_ltv · churn
• ▸Hourly retraining · per-merchant LR/RFM, ≥100 historical orders
• ▸≥ 70% probability gate · below threshold = nothing leaves your edge
• ▸3 destinations · Meta CAPI · GA4 · Klaviyo (per-platform toggles)
• ▸7-day forecast · projected predicted-conversions per platform

{
  "event_name": "predicted_purchase",
  "event_id":   "predicted_b1c4e3d2-...",
  "event_time": 1714137600,
  "action_source": "website",
  "user_data": {
    "em":  ["a8f5f167f44f4964e6c998dee827110c"],
    "fbp": "fb.1.1714137592.987654321",
    "external_id": ["c3f7e92d..."]
  },
  "custom_data": {
    "predicted_value":       87.50,
    "predicted_currency":    "EUR",
    "predicted_probability": 0.72,
    "model_version":         "rfm-lr-v1",
    "merchant_id":           "glass-house"
  }
}

Meta's delivery model is a hungry one. It needs roughly 50 attributed conversions per ad set per week to exit the learning phase, and it weights server-side purchase signal far more heavily than view content or add to cart. The problem is timing: most paid traffic browses for two to fourteen days before converting, so the algorithm spends your budget on prospecting impressions while waiting for a signal that arrives too late to course-correct delivery.

The standard fix is to widen your attribution window or dump optimisation events from purchase down to add to cart — both of which work and both of which dilute the signal Meta is actually trying to learn from. You end up with a model trained on intent, not outcome, and CPA drifts upward as the auction gets more expensive.

Boosted Events fills the signal gap a different way. Instead of changing the optimisation event, we generate a second, predictive event next to your real one. When a user's recency, frequency and intent score crosses the 70% probability threshold, we ship a predicted_purchase with a deduplicable event_id and an estimated value to Meta CAPI — within minutes, not days.

Your real purchase still fires when it lands. Meta's dedupe key matches, the prediction is reconciled, and the algorithm gets a denser stream of high-quality signal. You keep the optimisation event you actually want, and the learning phase finishes in days instead of weeks.

User is likely to purchase in the next 24h.

RFM features + session intent (view_item, add_to_cart, begin_checkout). Gated at 0.7 probability. Sent with predicted_value (decayed AOV) and a deduplicable event_id prefixed predicted_ — Meta reconciles when the real purchase fires.

First-time buyer is likely to repeat within 30 days.

Triggered when an existing customer's session looks like their pre-second-purchase pattern. Klaviyo flag predicted_repeat_buyer powers replenishment and upsell flows. Meta CAPI receives it as a custom event for prospecting suppression.

User profile matches your top-decile LTV cohort.

Built from monetary feature + product mix. Sent to Meta CAPI as predicted_high_ltv with custom_data.predicted_value so value-based bidding can lean in. GA4 receives it as a user property for audience export.

Active customer is likely to lapse in the next 30 days.

Inverse RFM signal: high frequency, dropping recency. Klaviyo gets a custom_property to trigger win-back. Meta CAPI does not receive churn predictions (no upside for delivery, risk for spend).

-- runs hourly per merchant in
-- workers/cron/src/boosted-event-predictor.ts
select
  e.event_id, e.user_email_hash,
  -- recency · frequency · monetary, last 30d
  exp(-extract(epoch from now() - max(p.event_time))
      / 86400 / 14)                 as recency,
  count(p.id)                       as frequency,
  coalesce(sum(p.value), 0)         as monetary
from events e
left join events p
  on  p.user_email_hash = e.user_email_hash
  and p.event_name      = 'purchase'
  and p.event_time     >= now() - interval '30 days'
where e.event_name in
        ('view_item','add_to_cart','begin_checkout')
  and e.merchant_id   = $1
  and e.event_time   >= now() - interval '24 hours'
group by e.event_id, e.user_email_hash;

DTC home goods brand. Enabled predicted_purchase only. Meta delivery exited learning phase 8 days into the test window vs. 21 days on the prior identical campaign — CPA dropped from €38.10 to €33.45 across the same creative.

Wired predicted_repeat into the Klaviyo replenishment flow. Win-back emails fire two days before the model expects re-purchase intent rather than three weeks after the last order. AOV held flat; second-order rate climbed.

Used predicted_high_ltv to seed a Meta value-based audience. predicted_value is sent in EUR and clamped at top-decile AOV. Spend is identical; iROAS on the predicted-LTV ad set tracks 1.6× the prospecting baseline.