Machine Learning in Real Estate: How It Is Actually Used
By The Realtify · Published · Updated
Machine learning in real estate is older than the current excitement by about two decades, and it does something quite specific: it estimates a number or a category from data you already hold. What a property is worth. Which enquiry is serious. Which tenant will not renew. It does not write your listing copy and it does not answer your phone — those are different technologies that got filed under the same three letters, and the confusion is expensive because it makes people buy the wrong thing.
This page covers what these models genuinely predict, how much of your own data each one needs before it can work at all, the two validation mistakes that make most real estate models look brilliant in testing and useless in production, and the fair housing constraint that turned modelling choices into legal ones. It is more technical than a vendor blog and more practical than a paper, because the question underneath is a business decision: is one of these worth having, and how would you know if it were any good?
Machine learning in real estate is used for five jobs: valuing property (automated valuation models), scoring leads and predicting seller intent, forecasting tenant renewal and arrears risk, setting rent and forecasting demand, and classifying listing photographs with computer vision. All five are prediction, not generation — they return a number or a category from historical data rather than producing text or images. Each needs a specific amount of your own labelled history to work, must be validated on time-ordered data rather than random splits, and in housing must be checked for proxy discrimination, because the Fair Housing Act applies to algorithmic decisions as much as human ones.
Machine Learning Is the Prediction Layer, Not the Chatbot
One distinction resolves most of the confusion in this topic, and it is worth being precise about because the two technologies fail in opposite ways and need completely different things from you.
| Compared on | Machine learning (prediction) | Generative AI |
|---|---|---|
| What it returns | An estimate: ₹1.42 crore, or 78% likely to renew | A draft: a description, an email, a summary, an image |
| What it needs from you | Years of your own labelled history — the outcomes, not just the records | Almost no setup, but access to current facts if it is to be accurate |
| How you know it works | Backtest it against outcomes you already know. It is genuinely measurable | Someone reads it. There is no accuracy score for a paragraph |
| How it fails | Quietly. The market shifts, the model drifts, and the numbers slowly stop meaning anything | Loudly. It invents a possession date and puts it on a portal |
| Who it suits | Organisations with volume and history: portals, lenders, portfolios, large brokerages | Everyone, from day one, including a solo agent |
The practical translation: if your question is *what will this be worth* or *which of these should I call first*, that is machine learning and it needs data you may not have. If your question is *this takes four hours a week and nobody enjoys it*, that is generation and you can start on Monday. Small teams overwhelmingly have the second problem and are sold the first. The generation side is covered in the guide to generative AI in real estate; the whole three-layer picture is in the guide to AI in real estate.
One more piece of vocabulary, then we will stop. Supervised learning needs examples with known answers attached: a thousand sales with their prices, ten thousand leads with whether they converted. Unsupervised learning finds structure without answers, which in this industry mostly means segmentation. Almost everything commercially useful in property is supervised, which is why the availability of labelled outcomes — not the algorithm — is the binding constraint.
How Is Machine Learning Used in Real Estate? The Five Real Jobs
Strip out the marketing and there are five jobs where prediction models are genuinely in production across the industry in 2026. Each has a different maturity and a different data prerequisite.
| The job | What the model returns | What it needs to exist first |
|---|---|---|
| Valuation | An estimated value with a confidence range | Thousands of comparable transactions in the same market, with attributes recorded |
| Lead scoring and seller intent | A probability that an enquiry converts, or that an owner sells within a window | Several thousand past leads with the outcome recorded, including the ones that went nowhere |
| Renewal, churn and arrears risk | A ranked watchlist of tenants unlikely to renew or likely to fall behind | Three or more years of renewal and payment outcomes, with reasons |
| Rent pricing and demand forecasting | A recommended rent, or expected absorption over a period | Dense unit-level pricing and occupancy history — and now a legal review |
| Computer vision on listing photos | Room type, features, condition and quality scores from an image | Nothing from you — this is bought, not built, because the training set is millions of photos |
Notice that four of the five require your own labelled history, and that the requirement is measured in years and thousands, not in weeks and hundreds. That is not a limitation of the technology. It is what supervised learning is.
Valuation: The Oldest Model and Its Honest Error Rate
Automated valuation models are the most mature application in the industry and the only one where a major operator publishes its own accuracy. Zillow states a national median error rate for the Zestimate of roughly 2% on homes currently listed and around 7% on off-market homes, and publishes the figures itself. That transparency is genuinely unusual and worth reading carefully, because three subtleties in it apply to every valuation model you will ever be sold.
- 1Median means half are worse. A 7% median error is not a 7% ceiling. Half of all estimates are further off than that, and the tail on unusual properties is much longer than the headline suggests.
- 2The on-market figure is flattered by the listing. Once a property is listed, the asking price becomes available information and estimates converge toward it. Comparing an on-market accuracy figure to what a model can do before a price exists is not a like-for-like comparison.
- 3National accuracy tells you nothing about your submarket. Model error is driven by transaction density and attribute quality. In a liquid suburb of standardised housing stock the model is excellent; in a thin market of heterogeneous properties, which describes much of India, it degrades sharply and the confidence interval is the only honest output.
The technique itself is not exotic. Most production AVMs are gradient-boosted decision trees over tabular features — location, size, age, floor, condition, recent nearby sales, time — and on this kind of structured data those models still outperform deep learning comfortably. The competitive advantage was never the algorithm. It was always the data: transaction coverage, attribute accuracy, and how quickly recent sales enter the training set.
Lead Scoring and Seller Intent
This is the most heavily marketed application to brokerages and the one most likely to disappoint, for a reason that is arithmetic rather than technological.
Suppose 3% of your enquiries become transactions. A model that predicts *no* for every single lead is 97% accurate. It is also completely worthless. Accuracy is the wrong metric for any rare outcome, which is why a vendor quoting accuracy on a lead-scoring model has either misunderstood their own product or is relying on you to.
- Ask for lift, not accuracy. If you call the top-scored decile of leads, how many more conversions do you get than calling a random decile? A model that produces two or three times lift in the top decile is genuinely valuable. One that produces 1.1 times is a rounding error dressed as intelligence.
- Ask about precision and recall together. A model can flag almost nothing and be right every time, or flag everything and catch every deal. Neither on its own is useful, and vendors quote whichever flatters them.
- Check whether the score changes behaviour. A score nobody acts on is a report. If the top decile does not get called first, the model has changed nothing regardless of its statistics.
- Seller intent is harder than buyer intent and sold more confidently. Predicting which homeowner will list within twelve months from public records and behavioural signals is a legitimate research problem with modest achievable accuracy. Treat any vendor promising confident seller predictions as making a claim they cannot support on your market.
There is also a sequencing point that matters more than the model. Scoring leads is only useful if the leads are being answered. If enquiries are sitting overnight, a scoring model tells you precisely which of the deals you lost was the most valuable. Fix the response first — that is the argument in the guide to AI chatbots for real estate, and the reason lead qualification automation usually outranks scoring in the queue.
Renewal, Arrears and Rent Pricing
Property management is where prediction has the cleanest economics, because the outcomes are frequent, recorded and financially explicit. A tenant either renewed or did not. A payment either arrived or did not. That is a labelled dataset generating itself every month, which is exactly what supervised learning needs and exactly what a brokerage's sales pipeline does not have.
| Prediction | How well it works | The catch |
|---|---|---|
| Which tenants will not renew | Useful at the margin — a ranked watchlist that beats instinct by enough to justify earlier conversations | Needs three or more years of renewal outcomes with reasons recorded, which most managers do not capture |
| Which accounts will fall into arrears | The strongest of the group. Payment behaviour is dense, regular and highly predictive | Acting on it requires judgement. A prediction is not a reason to treat someone differently |
| Maintenance and equipment failure | Genuinely good where sensors exist; weak where the only data is work orders | The data comes from hardware you may not have installed |
| Rent pricing | Technically effective and now legally constrained | Inputs matter more than output. See below |
| Demand and absorption forecasting | Reasonable directionally, poor at turning points — like most economic forecasting | Every model is trained on the past and turning points are where the past stops applying |
Rent pricing deserves its own warning, because in 2026 it stopped being purely a modelling question. The US Department of Justice and a group of states sued RealPage over its revenue-management software, alleging that competing landlords were pricing apartments using each other's non-public data. The case settled in November 2025 without an admission of liability, under a consent decree that bars the software from using competitors' non-public information at runtime, bars training on active lease data, restricts geographic modelling and installs a court-appointed monitor.
The Two Validation Errors That Ruin Real Estate Models
Here is the part missing from almost every article on this subject, including the technical ones. Two specific mistakes explain most real estate models that test beautifully and fail in production. Both are avoidable and neither is subtle once you have seen it.
Error one: random train/test splits. The standard tutorial approach is to shuffle your data and hold back 20% for testing. In a property market this is invalid, because prices, rents and demand are time-ordered. A randomly shuffled test set contains transactions from *before* some of the training data, so the model is effectively being told about the future and then congratulated for predicting the past. The fix is a time-based split: train on everything before a date, test on everything after it, exactly as the model will be used.
Error two: data leakage. A feature that only becomes known after the outcome guarantees a perfect backtest and a useless model. In real estate the classic offenders are almost invitations: the number of viewings on a property when predicting whether it sells; the final agreed price when predicting the valuation; the number of follow-up calls when predicting whether a lead converts. Each of these is a consequence of the outcome, not a cause of it.
- 1Split by time, always. And test across at least one market turn if your history is long enough to include one.
- 2For every feature, ask when it becomes known. If the answer is 'after the thing we are predicting', remove it.
- 3Backtest on a period nobody tuned on. A model tuned and then evaluated on the same window measures memorisation.
- 4Re-measure quarterly and watch for drift. Markets change regime; a model trained on a rising market applies rising-market logic until someone retrains it. This is not hypothetical — the most expensive lesson in residential real estate came from a pricing model taking balance-sheet risk into a turning market.
- 5Keep a champion-challenger setup. Run the new model alongside the old one in shadow mode before it makes decisions. It is cheap and it catches the failures that matter.
How Much Data Do You Actually Need?
Vendors answer this with "it depends", which is true and unhelpful. Here are working thresholds. They are rules of thumb rather than laws, and they are approximately where projects stop being wishful.
| Model | Rough minimum | Why that number |
|---|---|---|
| Local valuation model | Several thousand transactions in the specific market, with attributes | Value is driven by location and property type; the model needs enough examples within each combination, not in total |
| Lead conversion scoring | A few thousand leads including at least a few hundred conversions | The rare class sets the limit. Two hundred conversions across five thousand leads is workable; twelve is not |
| Renewal prediction | Three or more years of renewal decisions across a few hundred units | Renewals happen annually, so a year of data is one observation per tenant |
| Arrears prediction | Twelve to twenty-four months of payment records | Monthly events mean the data accumulates twelve times faster than renewals |
| Anything with computer vision | Buy it | Training an image model needs millions of labelled photographs. This is a category to license, never to build |
Two qualifications. First, features multiply the requirement: a model with forty inputs needs materially more examples than one with eight, and small-data projects almost always do better with fewer, better-chosen features. Second, and more often decisive, the data must be *labelled and clean*. Ten thousand leads in a CRM where the outcome field was filled in optionally is not a training set. It is ten thousand rows with an unusable target, and cleaning that is the project.
Fair Housing Is Now a Modelling Constraint
A model does not need a protected characteristic as an input to discriminate on it. It needs a correlate, and in housing the correlates are everywhere: postcode, school district, name, language of enquiry, source of income, even the device someone browses on. A model optimising purely for conversion will happily learn a pattern that a regulator reads as disparate impact, and it will do so without anyone intending it.
In May 2024 the US Department of Housing and Urban Development issued guidance stating that the Fair Housing Act applies to tenant screening and to housing-related advertising that relies on algorithms and artificial intelligence. That is the whole point in one sentence: using a third-party model does not transfer the obligation. If a screening or targeting system produces discriminatory outcomes, the housing provider is exposed regardless of who built it.
- Screening and advertising are the high-risk zones. Deciding who gets housed and who gets shown a home carries obligations that a lead-scoring model for sales enquiries does not.
- Audit outcomes by group, not inputs by name. Removing race from a model proves nothing. Testing whether approval or targeting rates differ by protected class is the actual check.
- Insist on explainability where decisions affect people. If the system cannot say why it declined an applicant, you cannot defend the decision, and "the model said so" is not a defence.
- Keep a human decision-maker on anything adverse. A person who can be asked to justify a rejection, with a documented reason.
- Ask the vendor for their bias testing. A serious provider has done it and will show you the method. A vendor who has never been asked is telling you something.
This applies well beyond the United States in substance if not in statute. Anywhere with anti-discrimination law in housing, the reasoning transfers, and the disclosure and transparency obligations that arrived with the EU AI Act in August 2026 point the same way. Governance is not a brake on this technology; it is the difference between a model you can deploy and one you have to withdraw.
Buy, Build or Skip
Most content on this topic implicitly assumes you should have a model. You often should not, and knowing that early saves a year. Here is the rule we use.
Pros
- Buy when the value is in a dataset you cannot assemble — valuation coverage, ownership records, computer vision on photographs
- Buy when the problem is common across firms, so somebody has already trained a better model than you will
- Build when the predictive signal lives in your own history and nobody else has it: your leads, your renewals, your arrears, your pipeline
- Build when the decision it informs is frequent enough to justify maintaining a model — and maintenance, not construction, is the real cost
- Build small first. A well-chosen logistic regression on clean data beats an elaborate model on messy data, and it is explainable, which matters legally
Cons
- Skip if the outcome field in your CRM is inconsistently filled in — there is nothing to learn from
- Skip if the decision it informs happens twice a year; a spreadsheet and judgement will do
- Skip if nobody will change behaviour based on the output. A score nobody acts on is a cost centre
- Skip if you have fewer than a few hundred positive outcomes recorded, whatever a vendor tells you
- Skip if your real problem is response time or data entry. Those are orchestration problems and prediction will not touch them
For teams under about twenty people the honest answer is usually skip, then buy, and rarely build — and the year spent recording outcomes properly is not wasted, because it is the prerequisite for every option above. Who sells what in each of these categories is mapped in the guide to AI real estate companies, and the portfolio-side view is in AI tools for commercial real estate.
What to Ask Before You Sign
Six questions. They are readable off this page by someone with no statistical background, and they are difficult to answer with a slide.
- 1What is the error rate in my market, not nationally? And what is the sample size behind that figure? A vendor who cannot break performance down by geography has not tested it there.
- 2How was the model validated — random split or time-ordered? If the answer is a random split, or if nobody knows, the reported performance is unreliable and everything after this question is moot.
- 3What lift do I get in the top decile? For any scoring product. Refuse accuracy as an answer.
- 4What features does it use, and when does each become known? This is the leakage check. Any feature that only exists after the outcome invalidates the backtest.
- 5How often is it retrained, and what triggers a retrain? "Annually" is a warning in a market that moves quarterly. "When performance degrades past a threshold" is a good answer, if they can name the threshold.
- 6Have you tested for disparate impact, and can I see the method? Mandatory for anything touching screening, targeting or who gets shown which property.
A vendor who answers all six plainly, including the parts that make their product look ordinary, is a different proposition from one who returns to the deck. In our experience the second question filters hardest, because it is the one nobody expects a real estate buyer to ask.
Frequently Asked Questions
How is machine learning used in real estate?
For five jobs, all of them prediction rather than content generation. Valuation, through automated valuation models that estimate a property's worth from comparable transactions. Lead scoring and seller intent, estimating which enquiries convert or which owners will list. Renewal, churn and arrears risk in property management. Rent pricing and demand forecasting. And computer vision, classifying room types, features and condition from listing photographs. Each returns a number or a category from historical data, and four of the five require years of your own recorded outcomes to work at all.
What is the difference between machine learning and generative AI in real estate?
Machine learning predicts: it takes your historical data and returns an estimate, such as a valuation or a probability that a tenant renews. You can measure whether it is right by backtesting against outcomes you already know. Generative AI produces: listing descriptions, emails, reports, images. It needs almost no setup but there is no accuracy score for a paragraph, so a person has to read it. Small teams usually have generation-shaped problems and are sold prediction tools, which is why so many AI purchases in real estate disappoint.
How accurate are machine learning property valuations?
Zillow publishes a national median error rate of roughly 2% for homes currently on the market and around 7% for off-market homes, which is about as good as automated valuation gets at national scale. Three caveats matter more than the headline. Median means half of all estimates are worse than that figure. On-market accuracy is flattered because the asking price is already public information. And national accuracy says nothing about a thin or heterogeneous submarket, where error rises sharply and the confidence range is the only honest output.
How much data do you need to train a real estate machine learning model?
As rough thresholds: several thousand local transactions with attributes for a valuation model; a few thousand leads including at least a few hundred recorded conversions for lead scoring; three or more years of renewal decisions across a few hundred units for renewal prediction; twelve to twenty-four months of payment records for arrears. Computer vision should always be licensed rather than built, since it needs millions of labelled images. Crucially the data must be labelled and consistent — a CRM where the outcome field was optional is not a training set.
Can machine learning models break fair housing rules?
Yes, and without anyone intending it. A model does not need a protected characteristic as an input to discriminate; a correlate such as postcode, name or language is enough, and a model optimising for conversion will learn those patterns. In May 2024 the US Department of Housing and Urban Development issued guidance confirming that the Fair Housing Act applies to tenant screening and housing advertising that relies on algorithms and AI, so using a third-party model does not transfer the obligation. Audit outcomes by group, require explainability for adverse decisions, and keep a human decision-maker.
Should a small brokerage build its own machine learning model?
Usually not, and often not yet. Build only where the predictive signal lives in data nobody else has — your own leads, renewals or arrears — and where the decision recurs often enough to justify maintaining the model, since maintenance rather than construction is the real cost. Buy where the value is in a dataset you cannot assemble, such as valuation coverage or image recognition. Skip entirely if outcome data is inconsistently recorded, if nobody will act on the score, or if the actual problem is response time, which prediction does not address.
Find Out Whether You Have the Data Yet
Before anyone sells you a scoring or valuation model, it is worth knowing whether your own data could support one. In a free 45-minute review we look at what your CRM and records actually contain — how consistently outcomes were captured, how many usable examples exist, and which single prediction would be worth making — and tell you plainly whether the answer is build, buy or not yet. You leave with the assessment either way.
Book a Free 45-Minute Audit ↗Conclusion
Machine learning has been quietly doing real work in real estate for twenty years, and the work is narrower than the marketing and more valuable than the sceptics allow. It values property, ranks enquiries, flags tenants at risk, prices rent and reads photographs. It does not decide anything, it does not write anything, and it cannot be evaluated by how impressive the demo was.
What separates the firms getting value from the ones with an abandoned pilot is unglamorous and consistent: they had labelled history, they validated on time-ordered data, they checked whether anyone acted on the output, and they retrained when the market moved. If you can only take one thing away, take the second validation question — random split or time-ordered — because a model tested on a shuffled property dataset has been marked on the wrong exam, and no amount of accuracy after that means anything at all.