Aggregate–Marginal Sign-Divergence Alert¶
Software tool — instantiates Aggregate–Marginal Trajectory Reconciliation
Opens review when linked aggregate and contribution trajectories meet sign, persistence, materiality, uncertainty, and quality conditions.
A standing monitor, not an analysis. The Aggregate–Marginal Sign-Divergence Alert watches one declared aggregate–marginal pair continuously and does exactly one thing: when the two trajectories point in opposite directions and that opposition clears every gate — a minimum persistence, a material magnitude, non-overlapping uncertainty, a minimum sample, and clean data — it opens a governed review. It never estimates the trajectories itself and never decides what to do about them. Its entire contribution is turning a five-part firing condition into an auditable, versioned trigger, so a masked transition cannot slip past unnoticed and cannot be waved through on a single dramatic week. Its defining discipline is restraint: a fired alert is an invitation to look, not a verdict.
Example¶
A mobile carrier's blended ARPU has been flat-to-rising for six quarters, and leadership reads the number as stability. The alert is wired to the pair — blended ARPU as the aggregate, the ARPU of each week's new activations as the entering contribution, linked by the subscriber stock-flow identity. Its rule fires only if new-activation ARPU has a negative slope while blended ARPU's slope is at least flat, the gap exceeds a material threshold, the two bands do not overlap, at least a few thousand activations back the estimate, and the billing feed is complete. For five straight weeks the new-cohort line drifts down while the blended line holds. On week five every gate passes, and the alert opens a review ticket routed to the pricing-and-retention council, tagged with the two trajectories, the persistence count, and a data-quality snapshot. No price change, no reorganization — just a review that now has to happen. (Figures illustrative.)
How it works¶
- Bound to a declared pair. The alert refuses to fire on an unlinked comparison; it requires the contribution identity that ties the entering measure to the future aggregate.
- Five gates, all ANDed. Sign opposition, persistence (consecutive or cumulative periods), material magnitude, non-overlapping uncertainty, and a minimum sample with a fresh, complete data feed.
- Suppression over guessing. If any gate fails — thin sample, stale feed — the alert suspends judgment rather than firing or silently clearing.
- Output is an opened review, not an action. A fired alert routes to a named owner with the evidence attached; every threshold and every change to it is version-stamped; snoozes and overrides are logged for audit.
Tuning parameters¶
- Persistence window — how many periods the opposition must hold before firing; longer cuts false alarms but delays the warning.
- Materiality threshold — the smallest gap that counts; raise it to suppress trivia, at the cost of missing slow bleeds.
- Uncertainty gate — require strictly non-overlapping bands versus a looser posterior-probability trigger; stricter avoids noise-chasing but raises the miss rate.
- Precautionary asymmetry — a safety- or equity-critical pair may fire on shorter persistence or lower materiality; the asymmetry must be declared, never smuggled in.
- Snooze and version control — how long a reviewed alert stays quiet, and who may retune thresholds; loose control invites suppression by the very team the alert would embarrass.
When it helps, and when it misleads¶
Its strength is converting "someone should have caught this" into a preregistered, hard-to-suppress trigger; the ANDed gates are exactly the anti-cherry-pick disciplines the archetype demands, and separating detection from action keeps the tool trusted.
Its failure modes are alert fatigue and gaming: thresholds quietly tuned to keep an inconvenient episode below the line, or so many alerts that reviews become rubber stamps. The classic misuse is firing on a single extreme week with no persistence gate, which produces whiplash. The guarding discipline mirrors statistical process control, where a run of points rather than one outlier signals a genuine shift[n1]: keep thresholds versioned, audit every snooze and override, and give the alert an owner distinct from the metric's owner.
How it implements the components¶
opposite_sign_divergence_test— encodes the sign-opposition-plus-persistence-plus-materiality condition as an automated, versioned firing rule rather than a one-off chart read.uncertainty_and_minimum_sample_guardrail— the non-overlap and minimum-sample gates suspend firing when the evidence is too thin or too noisy to govern.divergence_resolution_and_rebaseline_rule— a fired alert opens the review and routes it into the escalation-and-closure cadence, preserving threshold versions and exception logs.
It opens a review but neither estimates the leading-edge trajectory it watches (marginal_contribution_estimator — that belongs to Rolling Marginal-Contribution Curve) nor chooses the response (current_state_versus_leading_edge_decision_rule, protected_aggregate_obligation_check — those belong to Dual-Metric Decision Memo).
Related¶
- Instantiates: Aggregate–Marginal Trajectory Reconciliation — the alert is the always-on tripwire that starts the governed loop.
- Consumes: Rolling Marginal-Contribution Curve supplies the leading-edge trajectory it watches; Paired Confidence-Band Review supplies the bands its uncertainty gate reads.
- Sibling mechanisms: Cohort or Vintage Analysis · Contribution Waterfall Decomposition · Crossover Scenario Projection · Cumulative-versus-Incremental Dashboard · Dual-Metric Decision Memo · First-Difference or Derivative Estimate · Mix-Shift and Base-Effect Audit · Paired Confidence-Band Review · Rolling Marginal-Contribution Curve
Editorial Notes¶
Form Classification¶
Form family: Monitoring, Sensing & Alerting
Rationale: The mechanism opens review when linked aggregate and contribution trajectories meet sign, persistence, materiality, uncertainty, and quality conditions, so its operative form is repeated observation and operational signaling.
Independent corroboration: The frozen evidence defines Aggregate–Marginal Sign-Divergence Alert as 'Opens review when linked aggregate and contribution trajectories meet sign, persistence, materiality, uncertainty, and quality conditions', so its operative form is Monitoring, Sensing & Alerting.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Specialized
Rationale: Statistical process control uses persistence, run, sample-size, and uncertainty rules to distinguish a real trajectory shift from ordinary variation before opening review.
Related originating lineages:
- Accounting & Auditing — Materiality, data-quality controls, versioned triggers, and governed review contribute assurance discipline.
- Data Science & Analytics — Streaming monitoring implements versioned aggregate-contribution pairs, data-quality gates, alerting, and auditable rebaseline workflows.
- Economics & Finance — Aggregate versus marginal analysis explains how a healthy total can coexist with a deteriorating contribution and mask a transition.
Review resolution: NIST describes statistical control as monitoring changing bias and variability with control charts and remedial procedures, and requires uncertainty to support valid interpretation. Those inferential gates define when this alert may fire, making statistics and experimental design primary; data engineering, materiality assurance, and marginal economics form the Encyclopedia's paired-trajectory implementation.
Attribution caveat: Operational analytics implements the alert, but the persistence, uncertainty, sample, and suppression gates are statistical-process-control logic.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
- NIST/SEMATECH — Statistical Control of a Measurement Process
- NIST — Principles of Quality Assurance of Chemical Measurements
Notes¶
The alert's value is entirely in what it refuses to do. It will not act, and it will not fire on thin evidence. An alert that dictates action becomes a tripwire the metric owner eventually disconnects; one that only opens a review, on preregistered and audited terms, stays trusted long enough to catch the transition it exists for.
[n1] In statistical process control, a Shewhart control chart separates ordinary variation from a genuine shift using run rules — for example several consecutive points trending one way — rather than reacting to a single out-of-band point. That persistence logic is exactly what this alert applies to a sign divergence. ↩