60% of all sales are occurring above list price. Last month I gave examples of 20 bidding wars where houses traded for 120% of their asking price. A good buyers’ agent provides value when asked what is a house worth, and what will it take to get it? The first value is in their advice. The second is unseen value when they communicate to the seller’s agent that this buyer is (or is not) serious, able and committed to seeing this transaction through. Most buyers do not see this second part of the process, so they do not understand that alongside the money comes trust that the buyer will follow through on the commitment. Trust comes from experience. Trust is an incredibly difficult thing to measure. This column is about measuring Trust, and an agent’s value as a Recommendation Engine, and why Zillow’s prediction engine fails.

Recommendation Engine. Many years ago, when Netflix was still mailing DVD’s, they discovered their key value proposition was in their ability to recommend your next movie. Netflix improved from a 2% success rate in 2000 to over 80% today. How? With a collaborative filtering algorithm. We are similar and we like similar things, so if I like it, you should like it. That algorithm missed out on suggesting older “long-tail” titles, so Netflix made the adjustment. Adding demographic data did not improve the algorithm’s predictive power. Netflix has spent billions over 25 years learning that what customers say they want and what they do is very different.

Any experienced Realtor could have told you that. Zillow introduced the Zestimate in 2006 with a 13.6% median error rate for off-market homes. By 2021, its Neural Zestimate, using a neural network to analyze data like square footage and location, reduced this to 7.49% nationally—but in Connecticut, the error rate is 8.02%, a $232,580 miss on the median sale price.

Online prediction markets like Polymarket outperform neural networks and polls. They gave Trump a 60% chance of winning the 2024 election when Nate Silver gave Harris 55%. Data scientist Alex McCullough reports Polymarket’s accuracy at 90.5% one month before an event, rising to 94.2% four hours before. Polymarket tends to slightly overestimate event probabilities due to biases such as herd mentality, low liquidity, acquiescence bias, and participants’ preference for high-risk bets. These factors can lead to market prices being overpriced, with “Yes” outcomes resolving less frequently than expected. Realtors in Spring share this bias.

Netflix and Zillow could have asked my agent Susan Engel, “What does John want?” Because when she asked me what I was looking for and I said, “I want a really good deal” she replied, “No you don’t. Be careful what you wish for. There’s a reason some houses sell cheaply. It may take us a long time to re-sell it.” She was right, a “deal” was not the most important thing.  I didn’t want a white elephant that only I could appreciate. I wanted a great location that everybody wanted, and I didn’t want to get stuck with it if I lost my job. I reflected and told her I wanted “a blue-chip stock, a house that I can always sell, in any market, for at least what I paid.”  We found that “blue-chip” location and it turned out to be a good investment. Sometimes our agents know what we want, even when we don’t and cannot articulate it. I’ll bet the Financial Planners in the room can tell similar stories. “You say you want risk, but you don’t.”

I take copious notes when a buyer calls. Sometimes the conversation starts with something specific, “I want a midcentury modern.”  “I want an open plan.”  As often as not I am told what they don’t want. “I do not want to live on a double yellow line.”  “I don’t want to live too close to the New York border.”  “I don’t want too many steps.”  A question we should ask is “How long do you expect to live there?” because our needs change, and picking a “forever house”, one that meets your needs now and in 20 years, is sometimes too much to ask of the market, or the buyer.

The average one of us looks at 40 choices before we pick one. I’m describing movies on Netflix, of course. Nobody looks at 40 houses before making an offer, or do they? Realtor.com says buyers on average will visit 10 homes over 10 weeks before making an offer. Expectations are changing. NAR says 12 houses over 12 weeks a decade ago has dropped to 8 houses over 8 weeks in 2021 and that number is still dropping.  In a supply-constrained market, with most houses selling quickly with multiple bids, it is more typical for a buyer to make an offer on the first good house they see (after vetting dozens online) than risk losing it.

Trust. Where does trust live in the homebuying process? When there is no contact between buyer and seller before the closing what mechanisms what do we have in place for establishing trust when we just met at an open house? Ebay’s feedback rating is a walled-garden, one of the first commercial trust metrics, and it seems to work for them. We learn by the ratings seller history and frequency, and are they rated highly on real transactions. We rely on Amazon reviews, but according to a 2020 study about 42% of Amazon reviews are fake. 

Where then? NAR studies reveal that 43% of Realtors are hired based on the recommendations of friends and family. So analog! For the largest transaction most people ever make we have no objective criteria. According to NAR 90% of us would recommend our Realtor, but those recommendations are typically very subjective: “She was my Realtor, and she was so helpful” is a typical online review. Does it inspire trust? No. Google is no better. Searches for “Best Realtor in New Canaan” is garbage, nothing objective or believable. I searched for the top 15 Realtors in New Canaan and the lists include NONE with an office here, NONE who have sold or listed a house here in the last 90 days. Useless. Zillow’s ratings system falls short as an objective metric: success on Zillow comes to those agents who do a better job asking for reviews, and who pay Zillow. Gaming the system has become an art form and the New York attorney general doesn’t like it. She set up a fake yoghurt shop to catch the perpetrators writing fake reviews. And she’s prosecuting the agents who hire fake reviewers. She secured a $47 million judgement from an apartment finder website for defrauding renters. RateMyProfessors.com was founded in 25 years ago and includes 19 million ratings on 1.7 million professors at 8000 schools. That’s the same number of Realtors we have in the U.S.  No. “Easy” courses result in higher ratings. “Helpfulness” does too. There is very little correlation between online ratings and formal in-class evaluations of professors, leading many to conclude that here too, there is no accountability. 

Conclusion: The Limits of Prediction, the Value of Trust

While prediction engines like Polymarket (94.2% accuracy) and Netflix (80% success) excel in their domains, Zillow’s Zestimate, with an 8.02% error rate in Connecticut, shows AI’s limits in real estate. Agents remain the ultimate recommendation engine, building trust and understanding needs in ways algorithms can’t, ensuring buyers find not just a house, but a home.

John Engel is a broker with the Engel Team at Douglas Elliman in New Canaan and on ebay he has a reputation as “Nice transaction. Positive buyer” but on Google he really shines: “They did a great job finding us a house…and when it was time to sell we called on them and they did not disappoint.” High praise, indeed!

Check out John Engel’s Podcast, Boroughs & Burbs, the National Real Estate Conversation here.

Read this article on the New Canaan Sentinel website here.