Supply chains fail customers most visibly when deliveries arrive late. This project builds a control tower dashboard that turns raw order and shipment data into a single, clear view of delivery performance — using the DataCo Smart Supply Chain dataset (180,519 orders, 5 global markets).
This project is a complete, working research framework — not yet a full-dataset result. Here's exactly what that means:
| Component | Status |
|---|---|
| Research design, methodology, KPI definitions | ✅ Complete |
| Literature review (22 PRISMA-tracked sources, all verified; 69 total citations in the full paper) | ✅ Complete |
| Excel formula engine (KPIs, Pareto, control limits) | ✅ Built and tested |
| Formula engine validated on illustrative sample (24 orders) | ✅ Done |
| Full DataCo dataset (180,519 orders) run | ⏳ Not yet executed |
| Predictive model (logistic regression) evaluated on real data | ⏳ Not yet executed |
The companion workbook and report are explicit about this throughout — KPI cells are formula-linked so that running the full DataCo extract through the same pipeline updates every result automatically. The 24-row sample exists to prove the formula logic works, not to claim a finished result. The honest next step is documented in the report's own "Further Work" section: pull the full Mendeley/Kaggle extract, run the cleaning protocol, and let the engine recompute on real numbers.
The project asks: what drives late delivery, and can high-risk orders be flagged before dispatch? It works through four analytical layers — descriptive KPIs, diagnostic segmentation (Pareto + control limits), a baseline predictive classifier, and a five-view Power BI/Tableau dashboard — all grounded in a literature review spanning the bullwhip effect, supply chain visibility, control tower architecture, and India's national logistics policy context.
A key methodological decision worth highlighting: the DataCo dataset ships with a pre-computed "late delivery risk" flag. This project deliberately excludes that flag as a model feature, since it appears to be derived from the same fields used to define the prediction target — using it would cause target leakage and produce inflated, meaningless accuracy. This kind of leakage check is exactly the sort of thing that separates a real analysis from a naively-built one.
1. Data Cleaning & Star Schema
└── Parse dates, derive delay (actual − scheduled), build OTIF flag, separate cancellations
2. Descriptive Layer
├── Headline KPIs: OTIF rate, late delivery rate, order cycle time, fill rate, perfect order proxy
└── Breakdown by region, category, shipping mode, customer segment
3. Diagnostic Layer
├── Pareto analysis — isolates the few drivers behind most late orders
└── Statistical process control limits — separates true exceptions from normal variation
4. Predictive Layer
├── Baseline logistic regression: shipping mode, market, category, scheduled days, order value
├── Provided "late delivery risk" flag explicitly excluded (target leakage control)
└── Target: precision > 0.70, recall > 0.65 on held-out test set (not yet evaluated on full data)
5. Dashboard (Power BI / Tableau)
└── Five linked views: Executive Overview, Regional Performance, Category Performance,
Exception Monitor, Predictive Risk
supply-chain-control-tower/
│
├── README.md
├── LICENSE
│
├── workbook/
│ └── Project1_Research_Workbook.xlsx # 12-sheet workbook: search log, PRISMA, formula
│ # engine (illustrative sample), references
│
├── report/
│ ├── Supply_Chain_Control_Tower_Research_Package.docx # Full paper (~11,800 words):
│ │ # abstract, 10-part lit review,
│ │ # methodology, results, limitations,
│ │ # 69-source APA 7 reference list,
│ │ # publication upgrade plan
│ └── Project1_Supply_Chain_Control_Tower.docx # Condensed interview-prep version
│
└── data/raw/
└── README.md # Dataset sourcing instructions
This project uses the DataCo Smart Supply Chain Dataset (Constante, Silva & Pereira, 2019):
🔗 Kaggle: DataCo SMART Supply Chain for Big Data Analysis 🔗 Mendeley Data (original, DOI-citable, no login required): 10.17632/8gx2fvg2k6.5
180,519 order/shipment records, 53 fields, covering 5 global markets (LATAM, Europe, Pacific Asia, USCA, Africa).
A sample of the Excel formula engine already built and tested in the workbook:
| KPI | Formula Logic |
|---|---|
| Late delivery rate | Late orders / Delivered base |
| OTIF rate | Orders meeting both on-time AND in-full / Total delivered |
| Average delay (days) | AVERAGE(real days − scheduled days) across delivered orders |
| Fill rate | Quantity delivered / Quantity ordered |
| Sales at risk | SUMIF(status = "Late delivery", sales) |
The review spans global, national (India), state, and local evidence levels:
- Established evidence: information distortion produces the bullwhip effect (Lee, Padmanabhan & Whang, 1997); visibility supports supply chain agility (Barratt & Oke, 2007; Brusset, 2016)
- National context: India's National Logistics Policy targets cutting logistics cost from ~13–14% to 7.5–8% of GDP (DPIIT, 2022); India rose from rank 54 to rank 38 on the World Bank Logistics Performance Index between 2014 and 2023
- Methodology grounding: control tower architecture (Trzuskawska-Grzesinska, 2017), dashboard design principles (Few, 2013), target leakage avoidance (Hazen, Boone, Ezell & Jones-Farmer, 2014)
Full reference list (69 sources, APA 7) in report/Supply_Chain_Control_Tower_Research_Package.docx.
- Historic data — DataCo covers 2015–2018; any "real-time" framing should be understood as near-real-time on a refreshed extract, not a live feed
- Provided risk label excluded from modelling to avoid target leakage — this is a deliberate methodological choice, not an oversight
- Illustrative sample only — current workbook results (41% late delivery rate, 54.5% OTIF on a 24-order sample) are formula demonstrations, not full-dataset findings
- Cost proxies — "sales at risk" approximates but does not equal actual financial cost of delay
- Pull the full 180,519-row DataCo extract from Mendeley
- Run the 12-step cleaning protocol already logged in the workbook
- Let the formula engine recompute all KPIs automatically on real data
- Train and evaluate the logistic regression classifier on a proper train/test split
- Build the Power BI dashboard on the real, cleaned dataset
Project 1 of 5 in a data analytics portfolio built during BSc Analytics & Sustainability Studies at TISS Mumbai. Methodology mirrors the MIS dashboard and data-cleaning work done during the Aajeevika Bureau internship.