Late Delivery Risk Prediction in Supply Chain: Moving from Reactive to Proactive Delay Management

Imagine the world of logistics as a giant, chaotic game of "Catch Me If You Can," where packages are the players and delivery trucks are the chasers. Sometimes, the packages get lost, the trucks hit traffic, or the weather turns grumpy, and the game ends with a late arrival. For a long time, companies played this game reactively: they waited until the package was late, then panicked and tried to fix it. But what if you had a crystal ball? That's the world of Machine Learning, a branch of computer science where we teach computers to spot patterns in history so they can guess what happens next. Think of it like a detective who has read a million mystery novels; they know that if a suspect wears a red hat and runs at night, they are likely to be the culprit. In this story, the "suspects" are shipping orders, and the "crime" is being late. The paper we are looking at asks a simple but powerful question: Can we use these digital detectives to predict a late delivery before it even leaves the warehouse, so we can stop the crime before it happens?

---

The Late Delivery Crisis: A Story of Broken Promises

APL Logistics, a company that moves hundreds of global orders every day, was facing a massive headache. Imagine running a pizza delivery service where half your pizzas arrive cold and an hour late. That's roughly what was happening here: 54.8% of their shipments were arriving late. This wasn't just annoying for customers; it was costing the company money in penalties and ruining their reputation. The old way of doing things was "reactive," meaning they only tried to fix problems after the damage was done. The team wanted to switch to being "proactive," using a new tool to see the future.

The Crystal Ball: How They Built the Predictor

To build their "crystal ball," the researchers had to clean up a messy pile of data first. Imagine trying to predict the weather using a notebook that has pages torn out, ink blots, and notes about what happened after the storm. That's what raw data often looks like.

First, they had to stop the computer from using information that wouldn't be available in real life. In the data, there were columns like "Actual Delivery Status" and "Real Days for Shipping." If the computer saw these, it would just look at the answer key to guess the result, which is useless in real life. They threw those columns away, ensuring the model only used information available before the package shipped.

Next, they had to teach the computer to understand human language. Computers only speak numbers, so they turned words like "Standard Class" or "First Class" into code (0s and 1s). They also filled in missing pieces of the puzzle, like zip codes, so the computer didn't get confused.

One tricky part was that the data was unbalanced. It was like trying to learn to recognize lions by looking at a zoo where 99% of the animals are zebras. The computer would just guess "zebra" every time and be right, but it would never learn to spot a lion. Since late deliveries were actually quite common (54.8%), the team used a clever trick called SMOTE. Think of SMOTE as a photocopier that creates new, fake examples of the "late" deliveries to make the training data fair, so the computer learns to spot both on-time and late packages equally well.

The Secret Ingredients: Feature Engineering

The researchers didn't just feed the computer raw numbers; they cooked up some special ingredients called "features" to help the model think smarter.

1. Shipping Pressure Index: They calculated how much work was piled onto a single shipment. If you have a huge order but only a tiny window to ship it, that's high pressure.
2. Mode Risk Flag: They gave a "danger tag" to certain shipping speeds. If a package was marked "First Class" or "Express," the model knew these were high-risk because they promised a very tight delivery window that was hard to keep.
3. Region Congestion Index: They looked at history to see which places were always traffic jams. Some regions were like a clogged drain, always causing delays.
4. Order Complexity Score: They combined the price of the item and how hard the location was to reach. Expensive items going to tough places got a high "complexity" score, warning the system to be extra careful.

What the Data Revealed: The Shocking Truths

When the team let the computer crunch the numbers, it found some surprising secrets that the managers had missed.

• The "Second Class" Trap: The data showed that "Second Class" shipping was a disaster. It was promised to take 2 days, but in reality, it took nearly 4 days. It was almost always 2 days late. Similarly, "First Class" was promised for 1 day but took 2. The company was making promises they couldn't keep, and customers were getting angry.
• The Danger Zones: The map of the world lit up with red in specific areas. Central Africa, South Asia, and East Africa were the riskiest zones, with a late delivery probability averaging near 60%. It wasn't just bad luck; these areas had systemic issues with customs and local partners.
• The Money Leaks: The model found specific products and places where the company was bleeding cash. Shipping to the state of Minnesota was losing -$464 per order. One specific product, the "Total Gym 1400," was losing -$441 every time it was ordered. The company was essentially paying people to take their products away.

The Magic 72%: How Good is the Crystal Ball?

The team trained their computer model on 80% of the data and tested it on the remaining 20%. The result? The Random Forest model (a type of computer algorithm that uses many decision trees to make a choice) achieved 72% accuracy.

Now, 72% might sound like a B- grade, but in the messy, unpredictable world of global shipping—where traffic jams, storms, and customs officers can change everything overnight—this is a huge win. The model doesn't just say "Yes" or "No"; it gives a probability score, sorting orders into Low, Medium, and High Risk. This allows the operations team to focus their energy on the "High Risk" orders that need immediate attention.

From Code to Action: The Live Dashboard

The researchers didn't just leave the model in a computer lab. They built a Streamlit Dashboard, a live web tool that looks like a video game control panel.

• The Overview: It shows a big picture of how many orders are at risk.
• The Order Check: A worker can type in a new order, and the dashboard instantly says, "This one is High Risk!"
• The Action Plan: It tells the team exactly what to do. For the "Second Class" shipping, they need to talk to their partners and fix the delivery promises. For the "Total Gym 1400" and Minnesota orders, they might need to stop selling them or find a cheaper way to ship them.

The Future of Shipping

This project is a shift from reacting to disasters to predicting them. The author suggests that while the current model is a great "safety net," the future could be even brighter. Imagine adding real-time weather data or live port traffic updates to the mix. The goal is to move from just predicting delays to actually prescribing the best route to avoid them.

In the end, this paper shows that by cleaning up messy data, teaching computers to spot patterns, and building tools that humans can actually use, a logistics company can stop the bleeding, keep customers happy, and turn a chaotic game of catch into a well-oiled machine. The crystal ball isn't perfect, but it's the best tool they've ever had to see the future.