Zillow's zestimate has been under a lot of scrutiny lately due to an increase in inaccurate estimation leading to losses for the company.
This analysis takes a different approach, where we try to predict the log error in the zestimate prediction. Thus, instead of revisiting the actual model for Zestimate calculation (which is a black box to us), we try to analyze what parameters actually cause the Zestimate to be off.
The data is provisioned by Zillow through a Kaggle competition.
Upon analyzing the feature importance for the best chosen model, the following features seem to be the most impactful in contributing to the error in Zestimates:
- Property/County Land Use Code
- Presence in Zip Code 964XXX (Properties in this area are generally incorrectly priced)
- Bedroom count (Properties with lower bedroom counts tend to have a higher error in the Zestimate)
- Garage Count
- Buildings that are rated at a Quality level of 5