V1 Are we getting better? Late-delivery rate by month
V2 Where is lateness concentrated? (click a bar to focus)
V3 Promise vs reality: scheduled vs actual shipping days
V4 Which market x service-level cells are worst?
V5 Where do profits leak? Profit earned vs lost on loss-making order lines
How to use this dashboard
Purpose. Built for the Regional Head of E-Commerce Operations to decide which delivery promises, markets and products to act on first. The three tabs follow the decision flow: 1 Service health (is there a problem?) → 2 Root cause (why?) → 3 Commercial exposure (what does it cost?).
Interactions. (1) Global filters (date, market, mode, segment, department) apply everywhere. (2) Click any bar in V2 to focus the whole dashboard on that group; clear it in the sidebar. (3) Breakdown switches in V1, V3, V4, V5. (4) The what-if buffer slider in V3 re-computes lateness if promised dates were padded. (5) Hover for exact values and counts; the KPI strip always reflects the current filters and focus.
Definitions. Unit of analysis = order line (180,519 lines; 65,752 distinct orders). Late = Late_delivery_risk = 1 (actual shipping days exceed scheduled days; cancelled shipments are never flagged). Margin = sum(profit) / sum(sales). Loss-making line = Order Profit Per Order < 0. 90-day re-order = the same customer places another order 1–90 days later; only orders with 90 days of follow-up are counted.
Data notes. Markets are not observed in every month: in 26 of the 37 months only one market appears, and each market is recorded in separate windows (e.g. Africa in only 6 months). Cross-market comparisons therefore pool different periods and monthly trends are deliberately not split by market. Monthly volume also falls from about 5,200 to about 2,100 order lines after September 2017 (months with a single market); treat late-2017/2018 monthly rates with extra caution. No personal identifiers (names, e-mail, password, street) are loaded. The what-if view is a mechanical re-labelling of existing delivery times; it does not model any change in carrier behaviour.