Acumos Property Assistant
A python-powered machine learning model for determining the valuation of your property - powered by Acumos, Redfin, and sklearn.
Many real estate agents follow a notion of comps (or comparables) when pricing a home. Problem here is that this is often subject to an individual realtors' biases and often only looks at 2-3 properties in the area as a comparison. There are also many real estate websites (such as Zillow, Redfin, and Trulia) that do property valuations; however, the exact details that go into their pricing models are unclear. This program provides a simple open-source solution that can be easily extended for more niche property value modeling.
The Acumos Property Assistant is a machine learning model deployed on the Acumos store for pricing your home based on fundamental property characteristics.
More specifically, we survey the intrinsic properties of a home, such as number of bathrooms, number of bedrooms, square footage, and location, as sampled from the Redfin website (www.redfin.com), and allow users to price their home using a large dataset of collected properties.
By no means is this an exhaustive model, but should give users as idea of what their home (or someone else's home) is at worth in the current market.
What it does
Out of the box, the Acumos Property Assistant will take any property (that you find on Redfin or with the appropriate metrics) from the Boston area and appraise it based on all the properties that have sold in the last 3 months. This project contains two key components: A jupyter notebook containing Acumos model code / analysis done on the Redfin data, and the Acumos example website, which demonstrates live use of the deployed flask project with the Acumos model from a local container (see server.py in the linked github).
How I built it
This project uses a simple GLM (generalized linear model) across multiple property characteristics in order to appraise the value. In order to provide comparable property accuracy, the model should be trained on properties located within the same region as the property to be appraised; the training data was collected of a download of recent housing data available here:
This data could be exchanged for any other city in the model. As input, you'll want to provide a dataframe of properties to be tested and their appropriate features (square footage, bathrooms, bedrooms, etc - documented in the hosted model) as input.
Challenges I ran into
Documenting and converting the qualitative (non-numeric values) and re-encoding them in order to be used in the model.
Accomplishments that I'm proud of
What I learned
How to deploy and make a model available for public use.
What's next for the Acumos Property Assistant?
This model could be generalized and used on cities and regions all over the world. There is additional transparency from using this model to value your property, and the values/model parameters can be tuned to your liking. I may just use this to value my property in the future!
See the full code, deployment detail, and screenshots here: https://github.com/cbonoz/acumos-house-valuations