Supply Chain Performance and Risk analysis

Project Files & Repository Structure

This repository contains a complete end-to-end supply chain performance and risk analysis using the DataCo Smart Supply Chain dataset. The key files are described below:

• Dataset The original supply chain dataset used for analysis is provided as DataCoSupplyChainDataset.csv It contains order-item–level transactional data spanning 2015–2018.

• Analysis Notebook All data cleaning, KPI computation, exploratory analysis, modeling (XGBoost, clustering), and forecasting have been implemented in Analysis.ipynb

• Customer Segmentation Output The final customer segmentation results (5 customer personas derived using K-Means clustering) are available in customer_clusters.csv

• Project Synopsis A comprehensive project report summarizing objectives, methodology, insights, predictive models, and strategic recommendations is provided in SupplyChainPerformanceandAnalysisProject.pdf

---

Overview

• Analysed a 3-year online retail dataset (1.8L+ order-item transactions) from DataCo, performing rigorous data cleaning, deduplication, missing-value imputation, ZIP-code reconciliation, and full date standardization to convert item-level entries into reliable order- and customer-level aggregates.

• Computed 30+ operational, customer, and financial KPIs covering delivery performance, churn behavior, discount dependency, regional contribution, SKU profitability, fraud hotspots, and working-capital risks which revealed the structural inefficiencies and revenue leakage across the supply chain.

• Developed an interactive, multi-tab Streamlit dashboard using Python, Plotly, and NumPy to visualize real-time insights on revenue, profit, shipping delays, churn, category margins, fraud regions, and geographic order heatmaps.

• Built a 2-layer XGBoost model — a classifier to predict late deliveries (SLA breaches) and a regressor to predict shipment delays — enabling proactive logistics routing and risk-aware order handling.

• Performed customer segmentation using K-Means clustering and identified 5 high-impact customer personas, uncovering a 25–30% margin uplift opportunity and mismatches in the premium shipping experience.

• Implemented exponential forecasting using log-linear regression to project full-year 2018 revenue & profit, accounting for partial Jan-2018 data and month-wise trend momentum lost during 2017.

• Diagnosed key bottlenecks such as 57% late deliveries, 51% churn risk, 98% discount dependency, 22% payment failures, and 19% loss-making orders, highlighting critical operational leakage points.

• Delivered a 4-phase strategic roadmap (0–3, 3–6, 6–12, 12+ months) recommending premium-shipping redesign, discount optimization, predictive logistics, market-specific expansion, SKU rationalization, and partner renegotiation for sustainable margin improvement.

---

Executive Summary

Objective

The goal of this project was to analyze a 3-year online retail supply chain dataset, diagnose operational bottlenecks and revenue leakage, build predictive models to optimize logistics performance, segment customers for targeted strategies, and generate a strategic roadmap to improve margins, delivery reliability, and customer retention.

Tools, Language & Analytical Stack

Programming Language: Python Data & Analytics: Pandas, NumPy, Scikit-learn (K-Means, Linear Regression), XGBoost Visualization & Dashboarding: Streamlit (multi-tab UI), Plotly, Matplotlib, Seaborn Modeling: XGBoost Classifier & Regressor, Exponential/log-linear forecasting Environment: Jupyter Notebook and Streamlit App

---

Dataset Overview

Data Source

Dataset: DataCo Smart Supply Chain for Big Data Analysis Repository: Mendeley Data (Version 3) DOI: 10.17632/8gx2fvg2k6.3 Authors: Fabian Constante, Fernando Silva, António Pereira (2019)

Time Coverage

January 2015 → January 2018 (Complete 3-year dataset + partial month for 2018)

Granularity & Structure

The dataset is provided at an order-item level, where each row represents a unique product within an order. This required transforming and aggregating data into:

• Order-level metrics (order value, delivery performance, AOV, shipment delays)
• Customer-level metrics (RFM features, churn risk, behavior patterns)

Data Volume & Summary Metrics

• Total Records: 1,80,519 order-item transactions
• Unique Customers: 16,206
• Completed Orders: 36,866
• Number of Features: 53 columns

Data Scope

• Geographic Coverage: Southeast Asia, South Asia, Oceania, Eastern Asia, West Asia, West of USA , US Center , West Africa, Central Africa, North Africa, Western Europe, Northern Europe, Central America, Caribbean, South America, East Africa, Southern Europe, East of USA, Canada, Southern Africa, Central Asia, Eastern Europe, South of USA (23 regions)
• Total SKUs: 118 (Smart Watch, Perfect Fitness Perfect Rip Deck, Under Armour Girls' Toddler Spine Surge Runni, Nike Men's Dri-FIT Victory Golf Polo, etc)
• Product Categories: 51
• Customer Segments: Consumer, Home Office, Corporate
• Order Status: COMPLETE, PENDING, CLOSED, PENDING PAYMENT, CANCELED, PROCESSING, SUSPECTED FRAUD, ON HOLD, PAYMENT REVIEW

---

Understanding the KPIs

The key performance indicators I calculated are-

• Total Unique Customers and total orders
• Repeat Purchase Rate
• Contribution in Revenue of Top 10% and 30% customers
• Total Revenue & Profit
• Average revenue per customer
• Revenue across different regions
• Average Order Value, Average Order Quantity
• Late Delivery Rate, Cancellation Rate, Fraud Rate
• Top 10 products sold
• Churn Rate
• Average Customer Lifespan
• Customer LTV
• Average Time Between Orders

Out of all these KPIs the Customer Churn Rate, Late Delivery Rate across all Delivery Classes, Profit Margin, LTV, and Discount Penetration were directly tied to customer retention and operational savings and I find them most impactful.

NOTE

• Churn Rate: percentage of customers who stop using the service during a specific period (Customers who did not order for > 240 days)

• Average Customer Lifespan: average length of time a customer remains a customer (here: average duration from a customer's first order until their last order)

• Customer LTV: total predicted revenue a customer will generate for the business over the entire period of their relationship (total amount of money an average customer is expected to spend from the store before they stop shopping there)

• Customers inactive for > 240 days were considered churned (50%)

• Average lifespan ≈ 280 days (~9.3 months).

• LTV = Avg Revenue per Customer × Avg Sales per Customer × (Avg Lifespan / 365)

The main causes of revenue leakage

• Pending orders: 40% of all orders — huge working capital block.
• Pending Payment Status: 22% of transactions stuck, impacting cash flow.
• 19% orders unprofitable after costs
• 2018 profitability 35% below 2017
• Churn rate: ~51% — half of customers lapse after 240 days of inactivity
• Root causes: payment delays, pricing gaps, and supply chain inefficiency

---

XGBoost

XGBoost builds an ensemble of decision trees, where each new tree corrects the errors made by the previous ones. It optimizes a custom loss function using gradient descent and adds regularization to reduce overfitting.

What XGBoost Can Do:

Task Type	Example Use Cases
Classification	Predict late deliveries, churn, fraud
Regression	Forecast shipment delay, revenue
Ranking	Product recommendation, search ranking

How It Works:

1. Starts with an initial prediction (e.g., average value).
2. Computes the residuals (errors).
3. Builds a new tree to predict those residuals.
4. Adds that tree to the model using a learning rate.
5. Repeats steps 2–4 for N rounds.

But why XGBoost?

• I had to Select an algorithm for classification of late vs on-time.
• I chose XGBoost because it handles mixed data types well as I had numerical features like shipping days and categorical ones like shipping modes, order city, etc.
• It's also robust against overfitting because I can tune the Tree depth and learning rate.
• Moreover, I compared the results of this model with Linear Regression & Random Forest as well. XGBoost gave me a better accuracy score.

---

The 5 Customer Clusters

Clustering is an unsupervised learning task where the goal is to group a set of data points into clusters such that the points within the same cluster are more similar to each other than to points in other clusters. This is done to discover inherent groupings or structure in unlabeled data.

Cluster 0 – “Mass-Market High-Variety Buyers”

Type: General, diverse, multi-category shoppers Cluster Size: 11,687 customers (largest) Revenue Share: 91.95% (dominant)

Characteristics

• High order count per customer (≈5 orders)
• High variety: many categories & many countries (broad demand pattern)
• Moderate recency → periodic repeat buyers
• Medium delay rate (54%)
• Mix of Standard (61%) and Second-Class (18%) shipping

Cluster 1 – “Delay-Prone Low-Frequency Buyers”**

Type: Low-engagement, operationally risky Cluster Size: 1,812 Revenue Share: 1.98%

Characteristics

• Low frequency (~1 order)
• High delay rate (78%)
• Dominated by Second-Class shipping (98%)
• Higher-than-average profit margin
• Medium recency (inactive recently)

Cluster 2 – “Standard-Class Low-Touch Value Customers”

Type: Simple, low-maintenance buyers Cluster Size: 679 Revenue Share: 0.69%

Characteristics

• Low frequency (~1 order)
• Lower delay rate (46%)
• Strong preference for Standard Class (73%)
• Slightly negative profit margin (due to discounts)
• Small basket sizes

Cluster 3 – “Price-Sensitive Occasional Buyers (High On-Time Rate)”

Type: Low recency, stable, low-risk Cluster Size: 5,048 Revenue Share: 3.82%

Characteristics

• Very low delay rate (37%) – best logistics behavior
• High Standard-Class usage (91%)
• Low frequency (≈1)
• Moderate variety, small baskets
• Negative average delay (delivered before schedule)

Cluster 4 – “Premium Service Buyers with Extreme Delay Issues”

Type: Premium shipping, extremely late Cluster Size: 1,426 Revenue Share: 1.56%

Characteristics

• 98% First-Class usage, but 94.8% late rate
• Low frequency (~1)
• Higher than average profit margins
• High delay & high dissatisfaction risk
• One of the worst segments operationally

---

Insights & Recommendations

Category	Challenge Identified	Strategic Recommendations (Stronger & More Actionable)
Delivery & SLA	• 57% late deliveries • Premium shipping models consistently fail • Intercontinental routes highly unreliable	• Segment logistics partners by region (EU, Americas, APAC) and route type (air/sea) • Introduce Predictive Delay Engine using XGBoost outputs for proactive re-routing • Redesign premium shipping with guaranteed SLA, penalty clauses & priority lanes
Customer Retention	• 51% churn rate • Avg. lifespan 9 months • Avg. 60-day reorder gap	• Launch tiered Loyalty Program (Silver/Gold/Elite) based on RFM score • Introduce replenishment subscriptions for recurring-use SKUs • Personalized reactivation journeys using past order behavior
Fraud Detection	• 9 products & 3 regions account for 60% of fraud • High-risk ZIP clusters	• Build Product-Level Fraud Score based on ZIP code, discount usage & order velocity • Vendor audit for high-fraud SKUs • Deploy address verification, device fingerprinting & anomaly detection
Discount Strategy	• 98% of orders discounted • Avg discount = $21 • Margin dilution on bestsellers	• Run A/B price elasticity tests per category • Set Discount Guardrails—minimum required margin per product • Migrate to behavior-based discounting (repeat buyers get benefits, not all buyers)
Profitability	• 19% orders sold at a loss • Low-margin SKUs eroding performance	• Dynamic pricing tied to competitor & demand signals • Bundle low-margin SKUs with top sellers • Rationalize unprofitable SKUs and renegotiate supplier terms
Working Capital	• 23% orders stuck in pending payment • $8M revenue + $900K profit locked	• One-click payment UX revamp + auto-retry logic • Introduce alternate payments (UPI, wallets, BNPL) • Automate reconciliation to reduce payment friction
Regional Growth	• Africa = +32% higher margin but only 2% penetration • Overdependence on Europe & Central America	• Launch Africa-focused product catalog • Partner with regional logistics firms • Targeted marketing via localized campaigns
Product Growth	• Fitness category = 45% margin but low sales • High revenue concentration in apparel/footwear	• Create premium bundles & influencer partnerships for Fitness vertical • Expand SKU depth in high-margin categories • Cross-sell fitness items during sports/seasonal campaigns

Organizational Strategy — Implementation Roadmap

Phase 1 — Immediate (0–3 Months)

Objective: Stabilize delivery operations & stop financial leakage

• Establish a Premium Delivery Taskforce to fix SLA failures
• Resolve payment flow friction (auto-retry, simplified checkout, alternate methods)
• Deploy SKU-level profitability dashboards and kill-loss SKUs
• Introduce high-margin bundles (Fitness, Accessories, Lifestyle)
• Quick-win route optimization for delayed intercontinental segments

Phase 2 — Short Term (3–6 Months)

Objective: Build predictive capabilities & unlock customer value

• Deploy Predictive Delay Model into operations workflow (XGBoost output → routing actions)
• Launch VIP customer program for top 5% customers (85% revenue retention)
• Implement regional route optimization using shipping performance data
• Roll out discount A/B tests to cut unnecessary margin loss
• Introduce fraud risk scoring at SKU, ZIP and region level

Phase 3 — Medium Term (6–12 Months)

Objective: Expand capacity and optimize the supply chain end-to-end

• Pilot operations in Africa, the highest-margin region
• Establish Regional Hubs (EU, Americas, APAC) for better SLA performance
• Integrate logistics partner scorecarding with penalty/reward structure
• Expand dynamic pricing to category-level elasticity models
• Implement automated replenishment orders for repeatable SKUs

Phase 4 — Long Term (12+ Months)

Objective: Scale globally, digitize operations, and ensure sustainable growth

• Roll out global expansion blueprint for new markets
• Integrate advanced AI/ML forecasting, including demand sensing and optimal inventory placement
• Transition to an API-first logistics architecture with real-time carrier updates
• Build sustainability program: route carbon optimization, reducing packaging waste, green logistics
• Implement end-to-end supply chain digital twin for scenario simulation