🚍 Predicting Chicago Bus Delays with Weather Data

Inspiration

The idea for this project came from a real-world frustration: standing in freezing Chicago weather with a broken arm, waiting for a bus that never seemed to come. We wanted to build something that could help riders anticipate delays based on real conditions — especially when being stranded is not just an inconvenience, but a real hardship.

What We Built

We developed a machine learning model to predict bus delays in Chicago based on weather conditions like temperature, precipitation, and snow.
Our model achieved 93% recall for actual delays and a ROC AUC of 0.83, meaning it is highly sensitive to detecting when buses are likely to be late — empowering both commuters and transit operators to prepare ahead.

How We Built It

We combined historical weather data with bus service reliability data.
We engineered features like temperature, rain, snow, hour of day, and weekday to capture patterns.
We trained a classification model (Random Forest) to predict the probability of a delay occurring.
We built a Streamlit app to make predictions based on live or user-entered weather conditions. (Not complete yet)

Challenges We Faced

Data merging issues: Pandas was not robust enough to easily merge bus reliability data and weather data that didn’t have exact matching timestamps. We had to build a custom fuzzy matching approach to align records sensibly.
Finding enough data: The bus reliability API was down, so we had to creatively find and pre-process historical data ourselves, and we only worked with 2 weeks of data in April, despite access to plenty more weather data.
Resource limitations: Building and training models on a local machine made optimization and experimentation slower and more constrained than we'd hoped.

What We Learned

Real-world data is messy — merging datasets often requires flexible, creative solutions beyond simple joins.
Sensitivity (recall) is critical for public-facing predictions: it’s better to flag potential delays and be cautious than to miss them.
Streamlit made it surprisingly easy to turn a machine learning model into an interactive tool that others can use.