Predicting Zillow Estimation Error Using Linear Regression and Gradient Boosting

Owning property is one of the most important investments that a person can make in their lifetime. Therefore, being able to accurately know the real-time value of any property is crucial for making wise sales and purchases. Since the online real estate database company Zillow first developed a machine learning system to predict property sale prices in real time, it has continually worked to improve the accuracy of this prediction mechanism.In this paper, we describe our work to decrease the error of Zillow's price estimation by examining the effectiveness of several machine learning models and techniques at making property related forecasts. Specifically, we used property data to train linear regression and gradient boosting models with which we then made predictions about other properties. For the gradient boosting model, we used grid search to fine-tune the model's hyperparameters and observed the contribution of such tuning to the model's accuracy. Finally, we examined the effectiveness of several data preprocessing techniques, including the novel approach of treating time as a property feature.