Spotting patterns & outliers
Segmentation, trend breaks, and concentration — the three pattern families that answer 'why did the number move', and the correlation discipline that keeps them honest.
'Refunds are up 40%' is an aggregate, and aggregates hide everything interesting. Three framings do most of the work here: segment (slice by a dimension), trend-break (the point where the line bends), and concentration (how few things drive most of the total). Nearly every 'why did X move' investigation resolves through the same three moves, in roughly this order:
- Segment until it localizes. Cut the metric by every dimension you have — category, channel, store, reason, time. A genuine cause is almost never uniform: it concentrates somewhere. 'Refunds up 40% overall' usually decomposes into 'refunds up 200% in one segment, flat elsewhere' — and now you have a suspect instead of a statistic. Ask AI: 'given these dimensions, what's the segmentation order that isolates a cause fastest?'
- Find the trend break. Plot the metric over time within the suspect segment and look for the month/week the line bends. A bend date is investigative gold: it converts 'why is this high?' into 'what changed around April?' — a question interviews (Module 1) can answer. Gradual drift and step changes have different cause families; note which you're looking at.
- Measure concentration. Is the effect spread across hundreds of SKUs or concentrated in twelve? Ten stores or one? Concentrated effects have specific causes (a vendor, a flag, a process); diffuse effects have systemic ones (pricing, season, mix shift). Concentration also sizes the fix: twelve SKUs is a Tuesday afternoon; everything is a program.
The discipline: correlation is a lead, not a verdict
Your investigation will find that the refund spike coincides with the April mailer switch. Coincidence in time is evidence, not proof — the spring also brought a promo, seasonal gift purchases, and a new store manager. Before promoting a correlation to a cause, run the checks a skeptical CFO would: does the effect appear only where the mechanism applies (shipping damage should hit web orders, not store returns)? Does the reason data match the mechanism ('damaged in transit', not 'changed mind')? Is there a counter-segment the cause should spare, and does it? Each passed check earns confidence; AI is excellent at generating the checks ('list 5 alternative explanations and the cut that would distinguish each') and terrible at caring whether you run them.
Somewhere in the data will be a $1,900 refund on a $19 order, or a store with a 0% refund rate. The amateur move is dropping outliers to clean the chart. The professional move is a two-minute look first — outliers are how you find the data-entry bug, the fraud case, or the store that isn't logging returns at all. (And if data quality issues keep appearing, that's not your analysis failing — that's the warehouse needing the gates Data Foundations builds, or a tool like Vigil watching it.)