← Archimedes04 / 04
Proprietary engine — module 04

Signal monitoring

Most accounts do not fail suddenly. They drift for six weeks while everyone looks at the monthly report.
Signal monitoring — Archimedes module at Atlas Marketing
Monitoring compares every tracked metric against the forecast produced by the modelling layer, not against last week. That matters: a 12% drop in December may be entirely expected, while a flat week in peak season is a genuine problem. Anomalies are scored against the credible interval of the forecast and only breach when they fall outside it for a sustained window.
Alerts are routed to the responsible person with the likely cause attached — creative fatigue, a pricing change by a competitor, a tracking break, a landing-page regression — because an alert without a diagnosis just moves the work.
The same layer powers the client dashboard: one live view of spend, demand, share of search and forecast variance, updated daily and reviewed with the client every month.

Specification

Data latency
Daily, platform-dependent
Alert window
3–7 days persistence, metric-dependent
Coverage
Media, site, CRM and category signals
Review
Monthly, with the account lead

Data inputs

  • Ad platform and analytics APIs
  • CRM and revenue data
  • Forecast intervals from the modelling layer
  • Competitor activity from the market panel
  • Site performance and tracking health checks

How it runs

4 stages
  • 01

    Collect

    Daily automated pulls from ad platforms, GA4, CRM and the market panel into one warehouse.

  • 02

    Score

    Each metric compared against its forecast interval; deviations scored by magnitude and persistence.

  • 03

    Diagnose

    Correlated signals checked automatically — frequency, CPM, competitor spend, page speed, tracking health.

  • 04

    Alert

    Routed to the account lead with severity, probable cause and recommended action.

Methods inside the model

4 methods
  • 01

    Forecast-relative anomaly detection

    Deviation measured against expected value and interval, so seasonality never triggers a false alarm.

  • 02

    Persistence filtering

    A signal must hold for a defined window before alerting — noise does not wake anyone up.

  • 03

    Cause correlation

    Automatic checks across frequency, CPM, competitor spend and tracking to attach a probable cause to each alert.

  • 04

    Tracking integrity

    Daily validation of tags, events and attribution so a data break is never mistaken for a demand drop.

What it produces

  • /Live client dashboard, updated daily
  • /Alert log with cause and resolution
  • /Weekly variance-to-forecast summary
  • /Monthly performance review pack

Honest limits

  • Alerts are only as good as tracking. Broken or consent-limited measurement is flagged, not silently modelled around.
  • Platform data is delayed and occasionally restated; same-day conclusions are avoided by design.
Next moduleMarket analysis →
Talk to us about Archimedes →