How can predictive "likely to sell" scores be used without treating them as proof an owner will sell?

A likely to sell score orders records inside one file, which makes it a queue position rather than a finding about any owner. The score decides which records in an already defined segment get worked first, while any statement about one owner's plans comes from a recorded document or from the owner directly.

Contact data is metered, so treating a top band as a list of sellers has a price attached. PropertyRadar's $119 Solo plan includes 250 phone and 250 email contacts a month and bills 8 cents for each one past that, while Team includes 500 at 6 cents and Business includes 2,500 at 4 cents.[1] Pulling 3,000 top-band records on Solo runs $220 in overage before anyone has read a single county filing.

The test that settles whether a band earned that budget runs on your own records, over the six-month window the score itself states.[2] No 0 to 100 number identifies who can sign a deed, which is why a top-band label changes the order of the work rather than the checks that come before a call.

A propensity model outputs a raw number that a threshold converts into a category, which is the mechanic underneath every band label a property platform prints. Google's machine learning course describes a classification threshold as a threshold probability used to convert a model's raw numerical output into one of two categories, with examples above the line assigned to the positive class.[5]

Likely Transactions is PropertyRadar's version of that mechanic, with four scores per property covering sale, purchase, refinance and HELOC activity.[2] Each one runs 0 to 100 with plain-language bands from Very Low to Very High over a stated window of the next six months, and the monthly rescore means a record's band can move without any filing behind it.[2]

A band is a bucket of rank positions, so a top-band label means those owners scored above the rest of the file. Two records reach the same band from different inputs, since a model weighing equity and length of ownership can place a long-held free-and-clear house beside a recently remortgaged one.[2] Neither path is visible from the number, which is why the band works as a queue position rather than a description of the person.

The monthly rescore moves records across those boundaries without publishing a reason.[2] A record that dropped two bands since last month is a question for the county index, because the model may have responded to a data refresh rather than to anything the owner did.

Base rates cap how many sellers any band can hold, because a model can only redistribute a small number of positives across a large file. Google's course defines prediction bias as the difference between the mean of a model's predictions and the mean of the ground-truth labels, and gives the example of a model trained on data where 5% of emails are spam, which should predict on average that 5% are spam.[6]

The Census Bureau put the national mover rate at 8.4% in 2021, a new historical low over more than seven decades.[8] That figure counts everyone who changed residence, renters included, and a six-month scoring window covers half of any annual rate, so the share of a county's owners who record a transfer inside one stays small.

An invented file shows the shape without borrowing any product's performance. Take 10,000 owners in a county where 3% record a transfer in six months, which is 300 sellers, and a model that concentrates a third of them into a top band of 1,000 records. That band holds 100 sellers and 900 owners who did nothing, better than three times the base rate and still a band where nine of every ten conversations reach somebody who was never going to sell. Every figure there is illustrative arithmetic rather than a result from any market or product.

Calibration asks whether the number carries probability meaning or only rank order. The prediction-bias definition supplies the test, because a score of 0.10 applied to 1,000 records should be followed by roughly 100 positives whenever the number is a probability.[6] A 0 to 100 display score carries probability meaning only when the vendor states that mapping, and PropertyRadar's own wording puts its scores outside that category.[3]

You can ask about calibration without asking a vendor to open its model. Which population the score was fitted on matters, because a model trained statewide behaves differently inside one zip code, and which event counted as a positive matters just as much.

SmartZip Analytics sells a comparable propensity product to listing agents.[9] Its marketing homepage leads with the header "We Predict Listings" alongside the claim that 70% of homeowners choose the first agent they meet, which is marketing about how sellers pick an agent rather than information about how often a score is right.[9]

The size of the top band is a budget decision before it is a modeling decision, because the plan's included contacts set how many owners you can reach this month. A Solo account carrying 250 included phone contacts works a top band of 250 records with no overage, while a band of 2,000 adds $140 at 8 cents a contact.[1] Cutting the band to fit the allowance settles the month's spend before the dialing starts, which is harder to do once a 40,000-row sort is already open.

PropertyRadar describes customers filtering by score and sorting lists high to low, with the scores layered on top of criteria the platform already carries.[2] Its glossary counts 250+ search criteria available for building a list, and the score arrives as one more column beside them.[4]

Running the criteria first and the score second keeps every band inside a population somebody can defend to a partner. A record with a top-band score and no working phone produces no conversation this month at any price, while a middle-band record with an active mobile and a recorded notice of default produces one as soon as somebody calls it.

Precision and recall name the two readings of a holdout result, and they move against each other as the band widens. Google's course defines precision as the proportion of a model's positive classifications that are actually positive, and recall as the proportion of all actual positives the model classified correctly.[7] On a scored owner file, precision becomes the share of the worked band that recorded a transfer, while recall becomes the share of the segment's transfers the band captured.

Narrowing a band raises precision and gives up recall. The thresholding page states the general form, that a higher threshold makes a model predict fewer positives overall, both true and false.[5] Running the illustrative county file through that trade, a 1,000-record band holding 100 of the 300 sellers gives precision near 10% against recall near 33%, and doubling the band to 2,000 records would raise recall while precision fell.

A 250-record holdout against a 3% base rate expects around 8 transfers, a count small enough that a two-record difference between groups can come from chance. Testing on a larger slice, or repeating the design across two consecutive windows, buys confidence a single thin sample cannot. Publish the raw counts beside any lift figure so a reader can weigh transfer totals against the base rate the segment produced.

A recorded event carries a date and an office behind it, which a monthly model output does not. A notice of default or a listing that expired last month gives a caller something to reference and a timestamp to work against, while the band supplies the order those records get called in.

A monthly rescore and a monitored list answer separate questions.[2] PropertyRadar rescores every property on its Likely Transactions signals monthly, while a monitored list reports that a new document appeared under one record.[2]

Verification stays a human step at every band, since no score establishes who can sign a deed or whether the mailing address still reaches that person. Ownership of record, entity status, and occupancy get checked before a contact is purchased, and a top-band label changes nothing about that sequence.

PropertyRadar publishes the endpoint labels and the 0 to 100 range without publishing the cut points between bands or an accuracy figure for the scores.[2][3] A band's performance on your file is therefore a number you produce yourself, over the score's own six-month window.

Holding back a matched slice costs conversations you might otherwise have had, and 250 untouched records is a full month of a Solo plan's included phone contacts.[1] That cost buys the only comparison that separates the ranking from the outreach, so one holdout window per scoring change usually settles whether the band deserves its allowance.

The feature page limits the use as well as the reading, since it calls likely scores estimates for marketing use only and not a measure of creditworthiness.[3] A score has no place in a decision about credit, insurance, or tenancy, because treating it as a screening input puts it outside the vendor's own stated scope.

It is not a probability

A 0 to 100 display score reports rank order unless the vendor states a mapping to probability, and PropertyRadar calls its likely scores directional, model-based estimates for marketing use only.[3] A 90 therefore places the record near the top of the ordering, without any claim that nine such owners in ten will sell.

It is not intent

No model observes an owner's plans. PropertyRadar's chief executive Mark Hockridge is quoted in the product announcement saying it "is not a crystal ball," and the same announcement frames a high score as a reason to reach out that produces more conversations.[2]

It is not a dated event

A recorded notice of default or an expired listing carries a date and an office behind it, while a score carries a monthly rescore.[2] Sorting by score alone pulls records with no filing attached, which is why the criteria segment gets built before the sort.[4]