Fair warning: this issue is about AI and commercial real estate. If neither is your thing, delete it with my blessing; it won’t hurt my feelings. If you’re even a little curious, stay with me. It’s a bit technical but I’ll translate along the way.
Here’s a number every commercial broker lives with: in any given year, only about 5% of commercial properties change hands. Make 100 cold calls and, statistically, maybe 5 are sitting on a deal that trades in the next twelve months. The other 95 are practice. Prospecting is the whole art of finding that 5% before another broker does.
Cut to – About four months ago I started building a tool to fix my own problem of trying to track the 5% over time. It began as a CRM for an ADHD brain. Brokers today juggle a stack of software subscriptions that are single point solutions but none of them talk to each other and they all have rules and challenges about pulling information into your CRM (SOOO annoying). The administrative data entry is the ADHD Achilles heel. I wanted one brain instead of a Franken-stack.
Four months later and that exploration in vibe coding has turned into a multi-tool platform, been accepted by the NVIDIA Inception program, and crashed my computer…four times. UGH, why am I still doing this?!?!
Here’s why, and it has me genuinely rattled: it started making predictions that were coming true.
Sounds crazy? – yeah me too. Case and point. I got a ‘signal’ that University Town Center was going to sell. I dismissed it and told the system it was wrong. The next day it re-affirmed the prediction based on Rainier being announced as the developer for the north side development of UTC in Norman, the need for a liquidity event and a chain of entities and capital moving in a really peculiar way. Okay…interesting. But would it sell?
The dead gum thing closed in May of this year, 2 months after the prediction surfaced – WHAT!?!
Meet the CUBE. Picture an actual cube of data. One edge is a subset of commercial properties in Oklahoma & Cleveland County; 23,744 of them, to be precise. The second edge is time: every month from 2000 to 2025, or 312 months. The third edge is facts about each property in each month, 97 of them: assessed value, years since it last sold, that month’s mortgage rate, the income of its neighborhood, how its price compares to its true peers (just to name a few). Multiply it out and you get a little over 7.4 million calculations. Each one is a snapshot and a claim of what you could have known from public data about a property in a given month. All of this is assembled by a harnessing program called (wait for it) the Cube Factory. (florals for spring… groundbreaking)
How do you know a thing like this works? Two report-card numbers but I’m going to focus on the main one.
If you learn anything about predictive analytics you need to know ROC AUC. It stands for Receiver Operating Characteristic, Area Under the Curve, which is an MIT mouthful that means almost the accuracy of a prediction between two variables. So let me put it like this: hand the model one property that sold and one that didn’t, blind, and see if it can tell them apart. A coin flip scores 0.50. A perfect model scores 1.00. Banks and day traders use this metric and aim for a rating of 0.70 to 0.80….so better than a coin flip.
My first version of the cube scored 0.44 to 0.48. okay ouch; guessing would have been a better strategy. Painful. But By version 3 it was hitting 0.66 to 0.84. Now we’re getting somewhere. As of version 6, it scores 0.899 to 0.920. HOLY CRAP!
Why does that matter? Remember the 5%. Working the top of the CUBE’s list of “likely to sell” instead of guessing, real sellers show up 5 to 7 times more often; and in its best months, roughly 70 of its top 100 picks sold within 12 months. It doesn’t replace making the call and winning the listing – but it’s more of a triage tool to tell you which 5% to focus on.
Now, I know what you’re thinking. Everything I just showed you is historical; it’s the model proving itself against sales that already happened. Neat, sure, but nobody ever earned a commission on a deal that closed three years ago.
Exactly right. So in June, I stopped testing the past and pointed it at the future. I locked in a set of live predictions before the month started, then sat on my hands and let June play out. The month has closed, it’s been scored, and while technically I have to wait at least 6-12 months to declare it’s ‘official’ accuracy…it did predict 22 properties that did close in June…and the rest…oooo….this is just the beginning of the shenanigans.
More on the horizon… stay tuned for the progress.