Engagement snapshot
- Mandate
- Make pricing anomalies explainable fast enough to protect margin in the trading window.
- Timeline
- 10 weeks from data alignment to daily analyst use.
- Team shape
- Pricing director, 2 data engineers, 1 ML engineer, and 3 senior analysts.
The problem
Pricing analysts could see anomalies in dashboards, but still had to manually trace whether they came from feed lag, promotion stacking, or seller competition.
What we built
Built anomaly detection with narrative explainers that stitched together feed changes, promo logic, competitor signals, and margin thresholds.
Operating context
The marketplace already had anomaly alerts, but the analyst time sink came after the alert fired. Teams still had to reconstruct what changed and which lever should be corrected first.
Key constraints
- Explanations had to stay grounded in actual pricing events and margin rules, not generic summaries.
- Seller behavior, promotions, and feed changes needed separate attribution paths.
- Analysts wanted action-oriented explanations, not statistical jargon.
What we built
Event-sequenced anomaly context
Combined price feed updates, promotional rules, seller movements, and margin thresholds into one timeline.
Narrative explanation layer
Generated concise explanations tuned to analyst action, including likely cause and recommended next investigation step.
Decision feedback capture
Stored analyst outcomes so the system could learn which anomaly patterns were truly actionable.
Delivery path
Analyst workflow mapping
Observed how analysts triaged anomalies to distinguish useful evidence from dashboard noise.
Causal explanation pilot
Evaluated explanations on recent anomalies and scored whether analysts could act on them without extra digging.
Daily operations deployment
Integrated the explainer into analyst queues and measured response time and recovered margin by anomaly class.
Why it mattered
The value was not just faster alerts. Analysts could move from alert to action with much less reconstruction work, which meant more anomalies were actually corrected while the commercial window was still open.
Implementation notes
- Analysts needed ranked evidence and recommended next actions more than long-form explanations.
- Anomaly explainers work best when event history is first-class data, not an afterthought.
- Precision increased once the team encoded business-specific margin rules directly into the explanation layer.