As part of my 2025 internship with Data Detectives, I chose to work on a simple housing AVM as my data project. I started simple — very, very simple — with a dataset of just 15 homes manually entered from Redfin's website. Using these 15 house sales, I was able to implement simple tests and pipelines. I used SDV (Synthetic Data Vault) to generate synthetic data for the project and built a simple model. But I knew the project was far from complete.
I've recently taken this project back up. Using a scraper and data from Realtor.com, I've updated and expanded my dataset to include more than 3,000 homes. My results can be found in the code on my Github, linked above.
During my internship, I worked with a very limited set of features, selecting only the variables that were most important during my initial data analysis. These features only explain about 86% of the variance in house value in the full dataset — adding bathrooms, lot size, and neighborhood-level features will close most of the remaining gap.