Suneet Singal on how to read a property-technology claim

Suneet Singal’s published background is real estate and corporate finance, which is a useful way to read a property-technology claim: separate the product sentence from the table that would make the sentence true. A platform may say it knows prices, predicts rents, finds deals, or replaces an inspection. Those are different claims. Each one needs a dataset, an error, a time period, a geography, and a statement of what the model was not trained on. This entry is a list of those questions. It does not evaluate a named startup, and it does not say that Suneet Singal built a model. It points at public series a claim often borrows credibility from, including house-price indexes and Census housing statistics, and it asks the reader to check whether the claim is actually about those series. The result is a reading method, not an investment recommendation and not a product review.

A claim is a sentence that is missing a table

Property technology, in the loose sense used in pitches, is software attached to a real estate job: pricing, underwriting, marketing, property management, construction tracking, or search. The pitch sentence is usually short. The table that would support it is usually absent from the slide. Suneet Singal’s method is to write the missing table’s column headers before agreeing with the sentence. The headers are the data source, the geography, the time period, the thing being predicted or ranked, the error on a named sample, and the cases that were excluded. If the pitch cannot fill the headers, the sentence remains a sentence. It does not become a measurement because the interface looks precise.

The header set is the same habit as the housing-data entry, moved from a government release to a vendor. A government release at least prints its universe. A vendor slide sometimes prints a result and implies a universe. Asking “which sales, in which city, in which years?” is not hostility. It is the minimum required to tell a back-test from a demonstration on five friendly buildings. Five buildings can be a legitimate illustration. They are not a census, and they are not the Federal Housing Finance Agency’s house-price index. The moment the slide says “the market,” it has claimed a universe it may not have measured.

Keep the claim’s verb accurate. “Lists homes” is a database claim. “Predicts the sale price within a stated error” is a statistical claim. “Finds you a better loan” is a claim about a recommendation and, often, about a regulated activity. They do not share a standard of proof. A reader who accepts the first because the second sounds technical has mixed the verbs. Write the verb at the top of the note, and do not let a later sentence quietly upgrade “lists” into “predicts” or “predicts” into “guarantees.”

What data the product used

Ask where the rows came from. Recorded deeds, multiple-listing-service feeds, assessment rolls, rental listings, photographs, sensor feeds, and user-entered rents are different sources with different lags and different incentives. A deed is a recorded price and is silent on condition. A listing price is an asking price and may never become a sale. An assessment is a tax value and, in California, follows the base-year rules in the property-tax entry rather than the current market. A product that trains on assessments and then speaks as if it had trained on sales has changed the target. The question “what data?” includes “what was the target variable, in one sentence?”

Ask about coverage. A model trained on large coastal metros can be applied, in the software, to a thin rural county with a button that looks identical. The button is not a training set. Ask whether the geography in front of you was in the data, and in what volume. Ask how recent the last sale, listing, or rent in the file is. A model can be retrained annually and marketed as live. The refresh date belongs in the note next to the accuracy claim. The data-reading entry’s vintage rule is the same rule: a number without a period is not ready.

Ask what the product did with missing data. Houses without a recorded square footage, rents without a lease term, and sales that were not arm’s length are ordinary problems. Dropping them, filling them with averages, or refusing to score them are three different choices, and each choice changes the error. If the vendor says the data are proprietary and will not describe even the categories, you can still describe the claim as uncheckable. Uncheckable is a conclusion. It is a better conclusion than repeating the slide.

What error the product admits

Error needs a definition. A median absolute percentage error, a share of estimates within 10 percent of a later sale, and an average error that lets large misses cancel are three different sentences. Ask which one was computed, on which sample, over which period, and whether the sample was held out from training. A performance number computed on the same sales the model was fit to is not a forecast test. It can still be disclosed. It has to be labeled as fit, not as prediction.

Ask about the misses, not only the middle. A model that is close on typical houses and far off on unusual ones can publish a comforting median. The houses that hurt people are often the unusual ones. Ask for the error at a high percentile, or for a description of the worst misses, and ask whether those misses were removed before the chart was drawn. A chart with no sample size is decoration. A chart with a sample size and no statement of exclusions is half a chart.

Compare the claim with a dumb baseline. Predicting that a house sells for the recent median in its tract is a baseline. Predicting last year’s rent plus a flat percentage is a baseline. If the product does not beat a baseline the vendor is willing to name, the complexity is not yet an argument. This page does not set the baseline for every product. It says the pitch should contain one, and that the baseline should be dumber than the product, so the comparison means something. Beating a baseline the vendor chose after seeing the results is not the same achievement. Ask whether the baseline was fixed first.

What the model was not trained on

Every model has a world it did not see. The useful question is which parts of that world the product will still score without warning. New construction that has no sale history, renovations that are not in the records, condition that a photograph misses, disaster damage, rent control, unusual lot sizes, mixed-use buildings, and markets with very few annual sales are common gaps. A vendor who says “it handles everything” has not answered. A vendor who says “we do not score condominiums” or “errors are higher after major renovations” has given you a boundary you can use.

The boundary belongs next to the accuracy number. An error measured on arm’s-length single-family resales in dense tracts does not travel, automatically, to the house the boundary excludes. If the software produces a number anyway, ask whether that number is flagged as outside the training support. A flag is a feature. The absence of a flag means the user has to remember the boundary the slide admitted in a footnote. Users will not remember the footnote. The reading method writes the footnote into the note so the number cannot outrun it.

Suneet Singal did not, on the evidence of this site, build the model you are reading about. There is no claim of a proprietary score, a patent, or a back-test. The questions are the contribution. They can be asked of a portal’s value estimate, a landlord’s revenue-management tool, a lender’s automated valuation, or a slideshow at a conference. The automated-valuation entry asks a tighter version for the case where the output is a value used in lending or in a listing. This entry stays at the level of any property-technology claim, including claims that are not about value at all.

A demonstration set is not a market

Demos choose their buildings. A live demo on an address the salesperson picked is a story. A test on a random sample of sales, frozen before the model was tuned, is an evaluation. Ask which one you are watching. Ask whether addresses that produced embarrassing outputs were removed from the demo script. This is ordinary skepticism about demonstrations in any industry. It matters more when the demo is used to stand in for a housing statistic a reader could have looked up.

Public series are the check against a universal claim. If a product says it tracks “the market” in a metropolitan area, open the FHFA house-price index for the geography the agency actually names, and open the Census housing topics pages for the survey statistic that matches the verb. If the verb is rent, a price index is the wrong check. If the verb is sale prices, a median gross rent is the wrong check. The product can be more local than the public series and still be honest. It becomes dishonest when it borrows the public series’ authority without matching the public series’ definition. The Sacramento entry shows how many different objects share one place name. A proptech claim that says Sacramento has the same obligation to pick one.

The National Association of Realtors research page and the CFPB’s owning-a-home material are checks of a different kind. One is a market-research portal. The other describes the mortgage process a piece of software sometimes claims to streamline. A product can sit beside either process. It does not become the process because it exports a PDF. If a human still has a legal duty, the software has not erased the duty by being fast. Ask who signs, who is licensed, and who is responsible when the output is wrong. Those are product questions and, often, legal questions. This page does not answer the legal version. It tells you not to skip the question because the dashboard loaded quickly.

Regulated steps the software does not erase

Brokerage, lending, appraisal, and property management are activities with licenses and rules in the jurisdictions where they are practiced. A website can gather documents, suggest a price, or route a message without becoming the broker or the lender. The pitch should say which regulated role, if any, a licensed person still performs. If the pitch says the software “replaces” an appraiser or an inspector, read the inspection-and-appraisal entry and ask to see the engagement that was actually replaced. Often what was replaced was a lookup, and the engagement remains. Sometimes a lender really does use a model in place of a traditional appraisal, under a program with conditions. The conditions are the claim. “Replaces the appraisal” without the program name is an advertisement.

Fairness questions belong in the same pause. Ask whether the vendor has measured whether error differs across neighborhoods, and what they do when it does. This page does not assert that a named model is biased. It asserts that a claim of neutral accuracy is incomplete without a look at where the misses fall. A reader who wants a legal conclusion needs the law and the facts, not this essay. A reader who wants to avoid repeating a marketing sentence can stop at the unanswered question.

Security claims on a property platform are a separate worksheet. A sentence about encryption, passkeys, or “quantum-safe storage” should be read with the passkeys entry and the post-quantum entry, not accepted because the rest of the demo was about buildings. The cryptographic-inventory entry says what a real inventory contains. A badge on a login page is not an inventory. Property technology does not get a discount on those questions because the subject matter is houses.

A question list to keep

Use a short list and do not add a score the vendor did not earn. What is the verb of the claim? What data, from what period, in what geography? What is the target the model learned? What error, on what held-out sample, against what baseline? What was excluded, and does the product still return a number there? Who remains licensed and responsible? Which public series, if any, is the claim being compared with, and do the definitions match? What is the date of the methodology note?

If most of the list is blank, the honest summary is that the product was not checkable from the materials you were given. You can still try the product for a narrow job, such as keeping a list of addresses, without believing the widest sentence on the homepage. Suneet Singal’s reference recommends that separation. It does not recommend a vendor, a valuation, or a loan. The next page to open is the methodology note, or the public series the homepage invoked. The homepage itself has usually finished its work once the question list is written down.

What this entry does not claim

Suneet Singal’s published professional background is in real estate and corporate finance. Company filings linked from this site record a chief executive officer departure from FC Global Realty Incorporated effective 2 January 2018 and an acting chief executive officer appointment at First Capital Investment Corporation effective 12 September 2017. The Suneet Singal Scholarship, Grant and Giving websites describe him as the founder of those programs. This entry explains a general subject. It is not a private track record, a current job title, an investment result, or a quote.

Continue with questions to ask about automated valuation, why an article’s date is part of its meaning, and post-quantum planning in plain language.