Getting Real about Real Estate: A Machine Learning Categorization Model

Abstract

For many, especially those who may be lower-income and/or minorities, real estate is their first step towards the American dream. Given the large capital and risk required, they may only get one chance at it. As with many things in life, there are good and bad picks. A fair amount of financial fluency is required to pinpoint what makes a sound investment choice in the long run. While luck sometimes is enough, many times it is not and thus a gap is created between those able to afford the additional cost of an advisor. In the past, researchers such as Chiarazzo [7] and Ravikumar [19] have applied machine learning methods such as neural networks and random forests to create a price forecast for individual listings. In this study we focus on creating a model capable of indicating what can be considered a convincing residential real estate investment. More concretely, we use a random forest categorization model to predict if a certain listing will be below, at, or above average by price relative to its respective zip code. We also run a natural language processing model to retrieve non-quantitative information about location. Our training data is a combined dataset of Zillow listings (a reputable online marketplace) and corresponding Airbnb listings and reviews. Through our random forest categorization model, we find that square footage and location are some key drivers of price category. We also find that, unsurprisingly, demand has an inverse relationship with price. Our model was able to achieve a 62% accuracy rate.