Rebuilt a returns dashboard nobody trusted — and found the numbers were double-counted

My intern project was to 'refresh' a returns dashboard for one category team. Pretty quickly a manager told me nobody looked at it because the return rate seemed too high. So instead of just restyling it, I traced the SQL behind the top tile. The return rate was computed as returns divided by orders in the same window, but returns lag orders by weeks, so December returns were being divided by January orders. On top of that, one join fanned out when an order had multiple line items, so multi-item orders counted their returns twice. I rewrote it to cohort returns back to the original order date and de-duplicated at the order level. The 'scary' 11% dropped to about 6.5%, which matched what the finance team saw in their own numbers. I built the new version in QuickSight, wrote a short doc on the cohorting logic, and added a data-quality check that flagged if returns ever exceeded orders for a cohort. My manager had the category team start using it again in their weekly review. It was not glamorous, but it taught me that a wrong number is worse than no number.