Can Artificial Intelligence/AI reduce your property taxes?

AI based real estate property comparables — finally, a proper application for the nearest neighbor algorithm.

Photo by Kelly Sikkema on Unsplash

In this article, I will talk about how to potentially reduce the amount of property taxes. As Benjamin Franklin said, “in this world nothing can be said to be certain, except death and taxes.” But can we “solve for” lower taxes using AI?

This might seem presumptuous. But as I see it, you can either ignore this article and continue complaining about the level of property taxes or allow me 5 minutes to show how I reduced my house assessed value by ~$10K and property tax by ~$2K. The following is from the Cook County Assessor Office website for my PIN:

If my example will not convince you, then I will provide you with 2 other groups that can benefit from this AI based approach.

First, let’s start with a bit of context. If you live in the US and you own a house, you pay property taxes. I am not going to get into the historical details of property taxes. My goal here is to attempt to show you, at a high level, how deep learning (tensorflow) could potentially minimize your real estate tax by generating better comparables to leverage during the appeal process available to all tax payers in the county where I reside (Cook County Assessor Office/CCAO).

Let me also point out that this is not tax advice and I am not a lawyer.

What is the current state?

As a real estate owner your goal is to minimize the property taxes that you pay annually. Local governments provide an appeal process to allow you to challenge the assessment of your property and, therefore, how much taxes you pay. During this process you supply other comparable properties as arguments for reducing your taxes. Local governments offer a way to search and find comparables. However, this is not very effective as there are only a few factors you can change and you are only allowed to use a range. Here is an example:

On average, 17% of the owners submit an appeal and ~5% of them are successful. These 2022 numbers vary from year to year and are only based on CCAO appeals (not Board of Review appeals), but this should give you an idea of the appeal success rate.

If the above summary is too high level, there is an excellent Cook County Property Taxes 101 here.

What is the proposed change?

I developed my own approach to come up with comparables using deep learning/tensorflow. There are 2 steps:

1. A custom developed autoencoder model to create expressive embeddings. My model leverages data from the model used by the Cook County Assessor Office to determine the assessed values. You can find more details on the CCAO model and approach here
2. A nearest neighbor algorithm to find the most similar neighbors and, therefore, comparables. I use Google’s SCANN. Google Vertex AI Matching Engine simplifies deployment considerably

I package all this minimum viable product/MVP into a container (docker) that runs on Google Cloud Run. There are some other minor integrations (Google Maps). The MVP allows me to search for a PIN and then get the details in a table comparing the target PIN with the most similar PINs. Here is the user interface/ui:

What is different and maybe better?

1. It actually utilizes local government assessment data that is used to determine property taxes. It includes more data (100 factors), not just a few factors.
2. It leverages one of the most sophisticated and scalable algorithms to find nearest neighbors. Normalizing 100 factors (structured and unstructured) will give you vectors that are very large (40K elements in one vector describing one house). I am using Google’s TPUs to train the autoencoder.
3. It shows if any of the neighbors have been successful with their previous appeals.
4. It shows the agents/lawyers that have been successful in my area with the other appeals.

Who else can benefit?

Still not convinced about the applicability and utility of this approach for a property owner? I think there are other 2 groups that might benefit from this:

1. Assessor office — can use a similar approach to provide better comparables. The current approach is sub-optimal because the CCAO valuation model uses 100+ factors; however, we are only allowed to use a handful of factors to search for comparables.
2. Law firms — can increase their appeal success rate by: a) generating qualifying leads based on the PINs that have been successful and are very similar; b) understanding their competition; c) generating custom letters instead of generic ones when they market their service via direct mail.

The purpose of this article is to solicit feedback on the utility of this MVP. If I get a lot of interest/thumbs up, I might attempt to find a way to make this MVP available. I appreciate your comments.