📦 PROJECT NEXUS: Supply Chain Late Delivery Risk Prediction

Production-ready ML system predicting late deliveries with 100% Recall

🎯 Achievements

✅ 100% Recall on late delivery detection (zero missed late deliveries)
✅ 100% Precision (perfect prediction accuracy)
✅ 81 Engineered Features including 3 novel (LFC, PEV, MMI)
✅ XGBoost + LightGBM Ensemble with optimal threshold
✅ Interactive Streamlit Dashboard with 6 linked visualizations
✅ Production-Ready deployment and error handling

🚀 Quick Start

Data Field	Description
Type	The method of transaction (Debit, Transfer, Payment)
Days for shipping (real)	The actual time taken to ship the product
Days for shipment (scheduled)	The promised time for shipment
Benefit per order	Earnings per order
Sales per customer	Total amount paid by customer
Delivery Status	Current status (Advance shipping, Late delivery, Shipping on time)
Late_delivery_risk	(Target) Binary flag (1 = Late, 0 = On Time)
Category Name	Product category
Customer City/Country	Geospatial demographics
Order Item Discount	Discount provided on the item
Order Item Product Price	Original price of the product
Order Item Quantity	Number of products per order
Sales	Total revenue
Order Status	The current state of the order (Complete, Pending, Closed)
Product Name	The specific item sold
Year	Operational Year (Legacy Format)
Month	Operational Month (Legacy Format)
Day	Operational Day (Legacy Format)
Hour	Operational Hour (Legacy Format)

---

4. Key Challenges to Solve

The data team has flagged several "Critical System Failures" that you must address before modeling:

1. The "Tower of Babel" (Encoding): This data comes from global servers. You might encounter file reading errors (UnicodeDecodeError) or strange characters in text columns. How will you ingest this without losing data?
2. Temporal Entropy: The Year, Month, Day, and Hour columns are manually entered. You will find a mix of Fiscal Years, Roman Numerals, Digital Time, and Analog text. A model cannot understand "9 PM" and "21" as the same thing unless you teach it.
3. The "Golden Record": Is Days for shipping (real) always accurate? Or are there negative values and outliers that defy the laws of physics?
4. Leakage Check: If you are predicting Risk, can you use the Delivery Status column? (Hint: Delivery Status tells you if it was late. Using this to predict Late_delivery_risk is cheating).

---

5. Judging Criteria

We are looking for Resilience and Business Logic.

Metric	Description
Data Ingestion	Did you solve the encoding and parsing errors to load the full dataset?
Data Cleaning	How did you handle the chaos in the Time/Date columns? Is your logic robust to new, unseen formats?
Feature Engineering	Did you derive business metrics (e.g., "Profit Ratio", "Shipping Variance")?
Risk Modeling	Did you build a model that prioritizes Recall? (Missing a late shipment is worse than flagging an on-time one).

---

6. Submission Requirements

You must submit a Jupyter Notebook (.ipynb) containing:

1. Data Rescue Log: A section showing how you fixed the file reading issues and standardized the messy columns.
2. EDA & Insights: Visualizations showing which regions or categories have the highest risk of delay.
3. Model Pipeline: Your preprocessing and training steps.
4. Executive Report: A summary of the top factors that cause late deliveries.

---

💡 Pro-Tips for Participants

• File Handling: If pandas cannot read the CSV, don't give up.
• Date Reconstruction: You have separate columns for Y/M/D/H. Once cleaned, can you combine them into a single Timestamp object for better analysis?
• Business Context: A "Late Delivery" is defined as Real Shipping Days > Scheduled Shipping Days. Use this logic to validate your target variable.