FI-TW: An Open Train-Weather Dataset for Railway Delay Analysis in Finland

Imagine you are trying to predict how long a road trip will take. You know the distance, the speed limit, and the car's condition. But what if you forgot to check the weather? In the world of science, this is the difference between looking at a map and looking at a map plus a live storm radar. This paper lives in the intersection of data science (using computers to find patterns in huge piles of numbers) and transportation engineering (keeping trains running smoothly). The core idea is simple: trains don't just run on tracks; they run through the air, and the air changes. Sometimes the air is calm, and sometimes it's a blizzard. For a long time, scientists studying train delays had two separate notebooks: one with train schedules and one with weather reports. They rarely looked at them together. This is a problem because if you want to know why a train is late, you need to know if it's stuck in a snowdrift or just waiting for a signal. The authors of this paper decided to glue these two notebooks together to see if the weather is the secret villain behind the delays.

The researchers, working in Finland, have built a massive new digital library called FI-TW (Finland Integrated Train–Weather). Think of it as a giant, seven-year-long diary that records exactly what every long-distance train was doing at the exact same moment the weather station nearby was measuring the wind, snow, and temperature. They didn't just guess; they took real data from the Finnish railway system and the Finnish Meteorological Institute and stitched them together like a patchwork quilt.

Here is the story of what they found and how they did it:

The Great Data Stitch
Finland is a long, thin country that stretches from the coast up into the Arctic Circle. It has about 5,915 kilometers of train tracks and over 90 million passenger trips a year. The weather there is no joke; it can swing from a mild summer day to a bone-chilling -40°C winter night where metal doors freeze shut. To understand how this weather affects trains, the team grabbed two huge datasets. First, they took the "train diary" from the Digitraffic service, which tracks every single train's schedule, its actual arrival time, and how many minutes it was late. Second, they grabbed the "weather diary" from 209 weather stations scattered across the country.

The tricky part was matching them up. A train moves, but a weather station sits still. To solve this, the team used a mathematical tool called the Haversine distance (which is just a fancy way of calculating the shortest distance between two points on a round ball like Earth) to find the closest weather station to every single stop a train made. If the closest station didn't have a specific measurement (like snow depth), they had a backup plan: they looked for the next closest station within a 50-kilometer radius. It's like if your local bakery is out of cinnamon rolls, you don't give up; you check the bakery down the street. This "radial fallback" strategy ensured they didn't lose data just because one sensor was missing a piece of the puzzle.

The Result: Winter is the Villain
Once they had this massive dataset—containing about 38.5 million observations and 138 different features—they started looking for patterns. The results were as clear as a winter sunrise. They found a strong seasonal rhythm: trains are much more likely to be late in the winter. In the cold months of December, January, and February, the delay rate jumps to over 25%. In the summer, it drops to below 20%.

They also noticed that delays aren't spread out evenly. They tend to cluster in the central and northern parts of Finland, where the weather is harshest. The team even created special "weather scenario" flags, like a digital red flag for "Blizzard" or "Black Ice," to help computers recognize these dangerous conditions instantly.

The Test: Can We Predict the Future?
To see if this new dataset actually helps, the authors ran a simple test. They used a computer model called XGBoost (a type of machine learning that is very good at finding patterns) to predict how late a train would be at a specific station in Oulu. They taught the model to look at the train's schedule, the time of day, and the weather.

The result? The model could predict the delay with an average error of just 2.73 minutes. That's pretty impressive! It suggests that when you give a computer both the train schedule and the weather report, it can figure out delays much better than if it only had the schedule. The model was especially good at predicting delays that happened at that specific station, rather than delays that had just piled up from earlier in the trip.

What This Means
This paper doesn't claim to have solved the problem of train delays forever. It's more like handing researchers a super-powered flashlight. Before this, studying how weather affects trains in the Arctic was like trying to solve a puzzle with half the pieces missing. Now, with the FI-TW dataset, scientists have all the pieces. They can use this data to build better prediction systems, help railway companies plan for snowstorms, and figure out which parts of the track are most vulnerable to the weather.

The authors are careful to say this is just the beginning. They built the dataset and showed it works, but they didn't build the ultimate "perfect" prediction system. That's up to the next generation of scientists. But by opening the door and showing that weather and train data belong in the same room, they've given everyone a much better chance to keep the trains running on time, even when the Arctic wind is howling.