A distribution network that couldn’t see its own blind spots.

02 · The Raw Data

Real operational data never arrives clean.

These are raw shipment records, straight from the source — the kind of messy operational data an analyst actually inherits. Of 13,715 rows, 2,583 (18.8%) carried at least one quality issue. A representative sample, every problem flagged:

Note that one shipment ID can legitimately repeat across its SKUs — which is exactly why duplicates are detected on shipment + SKU, not the ID alone. Nothing is discarded yet: next, how each defect is detected and resolved — recovering what we can, dropping only the unusable.

03 · The Cleaning

Most problems get fixed, not thrown away.

Every injected defect was detected, then resolved by the cheapest honest means: recover in place where the data allows, flag and keep the real service events, and drop only what is genuinely unusable. 456 issues were repaired without losing a row; just 279 rows were dropped.

✓ Detection check: every injected defect was found — 738/738 matched (typos, duplicates, phantom stores, missing/negative quantities).

04 · The Findings

Where the sales leaked — and why.

An estimated $1.53M (est.) in sales never converted — demand that arrived to empty shelves. Three questions: where did the loss concentrate, when do deliveries fail, and how good is the forecast?

05 · The Control Tower

Everything above — live.

Filter by region, brand, and date. Every KPI and chart recomputes from the same daily aggregate. With no filter applied, the numbers equal the analysis exactly — the same data, made interactive.

06 · The Prediction

Now — see it coming.

Everything so far looked backward. The last move is forward: a stock-out risk score for every SKU at every CEDIS — and, deliberately, one you can read. Not a black box; a transparent scoring heuristic where every point traces to a cause a planner can act on. That legibility is the point, not a limitation.

The 7-day risk view

Of 120 SKU×CEDIS positions, the model flags a deliberately tight critical list — 13 High, 27 Medium, 80 Low. Most of the network is fine; attention goes where it’s earned.

07 · The Recommendation

What I’d do Monday morning.

Three moves, in priority order. Each ties back to a finding above — none is generic, and the numbers are the analysis’s, not a wishlist.

1. 01

   Concentrate replenishment where the loss is.

   ↓ Findings — Three CEDIS — Mexico City, Guadalajara, Monterrey — drive 62% of an estimated $1.53M in lost sales.

   Review safety stock, reorder points, and inbound lead times for those three centers first. The loss is concentrated, so the earliest fixes capture most of the recoverable upside before touching the long tail.

2. 02

   Move the weekend dispatch cutoff earlier.

   ↓ Findings — Friday orders run late 12.4% of the time vs 6.2% on Tuesday — a weekend dispatch backlog.

   Bring the Thursday/Friday cutoff forward, or add weekend dispatch capacity, so late-week orders clear before the backlog builds. A scheduling change, not a capital one — cheap to pilot.

3. 03

   Make the forecast event-aware, and run the risk watchlist weekly.

   ↓ Findings + Prediction — Forecast error of 18.9% that misses the event surge, and 13 positions already flagged High-risk.

   Add an event- and payday-aware uplift to the forecast for beverages & snacks, and stand up the High-risk list as a standing weekly report — so planners pre-position stock instead of reacting after the shelf is empty.