# Scoring data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète → across supplier and store feeds
At 6:45 AM on a Monday, a category manager at a mid-size grocery chain pulls up the weekly replenishment report and finds that 340 SKUs (stock-keeping units, the unique codes identifying each product variant) show zero sales at 12 stores. The stores didn't have a bad week. Their point-of-sale (POS) uploads arrived four hours late, missed the nightly batch job, and got silently excluded from the report. Nobody flagged it. This happens most weeks, somewhere in the network, and it quietly corrupts forecasts, restocking, and vendor scorecards. The fix isn't more dashboards. It's a data-quality scorecard that catches this before it reaches decision-makers.
Retail data pipelines are unusually fragile because they combine many independent, loosely governed sources:
Each source has its own update cadence, format, and owner. A supplier might change a UPC without notice. A store might batch-upload sales at midnight instead of hourly. None of this is malicious, but it accumulates into reporting that looks precise and is quietly wrong.
A practical scorecard scores each feed (by store, by supplier, by data domain) on four dimensions. Each should produce a simple percentage so scores are comparable across hundreds of stores or vendors.
Are all expected records present?
Metric: (records received / records expected) × 100
Example: a chain expects nightly POS uploads from 500 stores. On a given night, 486 arrive on time. Completeness = 486/500 = 97.2%. Retailers commonly target 98 to 99.5% completeness for core sales feeds (industry estimate, varies by chain size and IT maturity).
Do the values match reality or a trusted reference?
Metric: (correct records / total records checked) × 100
Classic accuracy failure: a supplier feed lists a SKU as "12-pack" when the physical case is actually 24 units. This silently doubles apparent inventory. Accuracy is usually tested by sampling records against a physical audit, a supplier's own master catalog, or GS1 standards (the global body that issues UPC/EAN barcodes and product data standards; see GS1's data quality guidance).
Did the data arrive within the window needed for decisions to be made on it?
Metric: (records received within SLA / total records) × 100
SLA = service-level agreement, the agreed maximum delay. If replenishment decisions run at 7 AM and a store's sales file lands at 9 AM, it's complete and accurate but useless for that cycle. Timeliness failures are the most common cause of phantom stockouts in fast-moving consumer goods (FMCG).
Do the same entities look the same across systems?
Metric: (matching records across systems / total compared) × 100
The classic case: SKU "884-2201" in the WMS, "SKU884-2201-A" in the POS, and "8842201" in the supplier EDI feed, all referring to the same product. Without a unified product ID strategy, joins across systems silently drop or duplicate records.
Say a retailer receives a weekly inventory feed from a packaged snacks supplier covering 1,200 SKUs across 150 stores (expected 1,200 records per store, 180,000 total).
| Dimension | Measured | Expected | Score |
|---|---|---|---|
| Completeness | 176,400 records received | 180,000 | 98.0% |
| Accuracy | 174,000 records match physical/reference check (sample-based) | 176,400 checked | 98.6% |
| Timeliness | 168,000 records arrived within the 6-hour SLA | 176,400 | 95.2% |
| Consistency | 171,000 SKU codes match master data exactly | 176,400 | 96.9% |
A simple composite score, weighting each dimension equally:
composite = (98.0 + 98.6 + 95.2 + 96.9) / 4 = 97.2%Retailers often set an action threshold around 95%: below that, the feed triggers a review with the supplier or store IT team before it's trusted in automated reordering. This 95% line is a common industry rule of thumb, not a regulatory standard, and should be calibrated to how much downstream risk a given feed carries.
For teams with even light SQLSQLSales Qualified Lead: a prospect the sales team has validated as ready for direct outreach and a proposal, having passed clear qualification criteria.Voir la définition complète → or Python access, this pattern generalizes across feeds:
def completeness(received, expected):
return round(received / expected * 100, 1)
def composite_score(completeness, accuracy, timeliness, consistency, weights=None):
weights = weights or [0.25, 0.25, 0.25, 0.25]
scores = [completeness, accuracy, timeliness, consistency]
return round(sum(s * w for s, w in zip(scores, weights)), 1)
# Snacks supplier example
score = composite_score(98.0, 98.6, 95.2, 96.9)
print(score) # 97.2Weights matter: a retailer running automated (algorithmic) replenishment might weight timeliness at 40% because stale data breaks the reorder logic immediately, while a slower quarterly assortment review can tolerate more lag.
Vérification des acquis
1. In the opening scenario, why did the missing sales data go unnoticed until the weekly report?
2. According to the lesson, what is the fundamental reason retail data pipelines are especially prone to quality problems?
3. Why does the lesson emphasize scoring each feed as a simple percentage rather than using more complex or varied metrics?
4. Select ALL correct answers about why the missing-data problem in the scenario is described as 'quietly' corrupting downstream processes.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers about the different data sources feeding into retail reporting described in the lesson.
Sélectionnez toutes les réponses correctes.
Scoring is diagnostic, not corrective. Governance assigns accountability:
In Europe, GDPR (General Data Protection Regulation) governs customer-linked data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète → obligations, particularly around accuracy and the right to rectification, when loyalty card or e-commerce data is involved. This is a legal obligation, not just an operational best practice, for any personally identifiable data in the feed.
As-of-2025 industry estimates (treat as directional, not audited figures):
Treat all of these as benchmarks to calibrate against, not universal targets. A regional discount chain with manual store uploads may reasonably run leaner SLAs than a global grocer with real-time APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète → feeds.