What are 'likely to sell' prediction models in real estate, and how accurate are they in 2026?

A 'likely to sell' prediction model is a machine learning system that estimates the probability that a specific property will be listed for sale within a defined time window — typically 6, 12, or 24 months. These models sit at the center of what the industry calls seller-lead prediction or listing prediction, and by August 2026 they have become one of the most commercially contested applications of artificial intelligence in residential real estate. The core output is deceptively simple: a score between 0 and 100 (or a probability between 0 and 1) attached to each home in a market, telling an agent, investor, or platform which owners are statistically most likely to put their house on the market soon.

What a Likely-to-Sell Model Actually Predicts

Also worth reading: How accurate are AI property valuation models in 2026, and should buyers and sellers trust them? · How do AI-driven real estate matching and property discovery platforms achieve EU AI Act compliance in 2026? · What are the definitive real estate voice agent ROI metrics for 2026?

The first thing to understand is that these models do not predict whether a home will sell — they predict whether it will be listed. Once a property hits the MLS with competent pricing and representation, the sale rate in most US markets runs between 90 and 96 percent within 90 days. The hard problem is upstream: identifying which of the roughly 140 million housing units in the United States will enter that funnel. A typical county might see only 4 to 7 percent of its owner-occupied homes list in any given year, so a model working against a base rate of 5 percent needs to be genuinely discriminative to be useful.

Most production systems frame this as binary classification: did the owner list within 12 months of the snapshot date, yes or no? Better systems use survival analysis, which produces a time-to-event curve rather than a single yes/no flag, letting you ask not just "will they sell" but "when." Uplift modelling — a technique borrowed from customer relationship management, where it powers up-sell, cross-sell, and retention campaigns — has also migrated into real estate. Instead of asking who is likely to sell, uplift models ask who is likely to sell because they were contacted, which is a materially different population and one that matters enormously when you are spending marketing dollars.

How These Models Are Built: Features and Data

The predictive power of a likely-to-sell model comes almost entirely from feature engineering rather than algorithm choice. Gradient boosted trees (XGBoost, LightGBM) still outperform deep learning on tabular data of this kind in most published benchmarks, and the features fall into four broad categories.

Life-event signals are the strongest predictors. Divorce filings, probate records, job changes inferred from commute-pattern shifts, marriage, birth records, and retirement-age thresholds all correlate strongly with listing behavior. Property signals come next: length of ownership (owners at the 7-year mark sell at roughly twice the baseline rate), equity position (owners with more than 40 percent equity have both the ability and incentive to move), mortgage rate lock-in (a household holding a 2.8 percent rate faces a payment shock of 60 to 80 percent if they rebuy at prevailing rates, which suppressed move rates measurably through 2023–2025), deferred maintenance visible in imagery, and expired listings or withdrawn listings from prior attempts. Behavioral signals include online activity — repeated searches for school districts in another area, valuation-tool lookups on the subject property, browsing of moving services — though this data raises privacy questions discussed below. Market signals round out the set: local price appreciation, inventory levels, days-on-market trends, and new-construction pressure nearby.

Training data is usually assembled from historical MLS records joined to tax assessor data, credit-header files, and consumer datasets. A typical enterprise model trains on millions of property-months of history. Class imbalance is severe — remember the ~5 percent annual base rate — so practitioners use techniques like focal loss, SMOTE oversampling, or simply calibrating probabilities carefully rather than chasing raw accuracy numbers.

Accuracy: What the Numbers Actually Say

Here is where marketing and reality diverge sharply. Vendor claims of "80 percent accuracy" almost always refer to accuracy on a balanced test set, which is meaningless when the base rate is 5 percent — a model that predicts nobody sells achieves 95 percent accuracy while being useless. The metrics that matter are precision-at-k (of your top 100 scored households, how many actually list?), recall, and calibration.

Realistic performance figures from industry practice and academic work: top-decile precision typically lands between 15 and 30 percent, meaning that if you mail or call the 10 percent of households with the highest scores, roughly one in four to one in six will list within the target window — a 3x to 6x lift over random targeting. Top-percentile precision can reach 35 to 50 percent but covers so few households that volume becomes impractical. Recall is the weak point: even excellent models capture only 20 to 40 percent of eventual sellers in their top decile, because a large share of listings are triggered by genuinely unpredictable events — a sudden divorce, a death, a job relocation announced with three weeks' notice. No model predicts those ex ante. Anyone selling you a system claiming to identify most future sellers is misrepresenting the state of the art.

Model decay is also underappreciated. Seller propensity shifts with mortgage-rate regimes, seasonal patterns, and macro conditions; a model trained on 2021 data degraded visibly during the 2022–2023 rate shock. Production systems need retraining at least quarterly, with drift monitoring on input distributions and calibration curves.

Comparison: Build vs. Buy vs. Platform Access

Building in-house makes sense only above a certain scale. If your team transacts fewer than 150 sides per year, the economics rarely justify nine-figure data pipelines and a dedicated ML hire. Buying scores from vendors like those offering pre-mover lists gets you speed but commoditizes your advantage — every competitor can buy the same list. Platforms that embed prediction inside a matching workflow occupy a middle ground: the agent does not see raw scores but receives ranked property-owner matches, which trades transparency for usability.

Practical Steps If You Want to Use One

Start by defining the decision the score feeds. A score without an action is trivia. If the action is direct mail, you care about top-decile precision and cost-per-listed-lead; if it is door-knocking in a farm area, geographic concentration matters as much as rank order. Second, demand validation on your own market before paying anything: ask the vendor for backtested precision-at-100 and precision-at-1000 figures for your metro, computed on a holdout period after their training cutoff. Third, run a controlled pilot — split your target geography, send to scored households in half and randomly selected households in the other half, and measure listing conversion over two quarters. Fourth, budget for follow-up cadence: predicted sellers convert over months, not days, and most teams abandon the program before the signal matures. Fifth, track cost per closed transaction, not cost per lead; a $4 postcard campaign that yields one listing per 800 sends at a $9,000 gross commission is wildly profitable even at a 0.125 percent hit rate.

Common Mistakes and Failure Modes

The most frequent error is treating the score as deterministic. A 70 percent likelihood-to-sell score means exactly what it says — three in ten such households will not list — and agents who burn relationships by presumptuously calling owners "ready to sell" damage their reputation in exactly the neighborhoods they farm. The second mistake is ignoring the privacy and compliance layer. Predicting life events from consumer data touches FCRA boundaries, state privacy laws (California's CCPA/CPRA, and similar statutes now active in a dozen states), and TCPA rules around automated calls and texts. Purchasing behavioral data scraped without consent carries legal exposure that several 2024–2026 enforcement actions have made concrete. Third, teams cherry-pick vendor case studies: a testimonial from a Phoenix brokerage tells you little about performance in a low-turnover Midwest market where annual sale rates run near 3 percent. Fourth, organizations conflate correlation with causation and skip uplift analysis entirely, spending money contacting people who would have listed anyway — wasted spend that can represent 40 to 60 percent of a naive campaign. Finally, many buyers never re-validate: a model validated in Q1 2025 may be miscalibrated by mid-2026 if rate cuts reshaped move-up behavior.

When It Makes Sense to Act — and When to Wait

Timing matters on both the adoption and deployment side. For brokerages, the argument for adopting some form of seller prediction in 2026 is competitive parity rather than edge: the technology has diffused enough that being absent from predicted-seller outreach means ceding ground to teams that use it. For individual agents, the calculus depends on farming intensity — if you already work a defined geographic farm of 500 to 2,000 homes, layering propensity scores onto existing mailings costs little and reliably improves hit rates by a factor of two to four. Waiting makes sense if your market is supply-starved and turnover is below 3 percent annually, if you lack the operational capacity to sustain 6-plus months of follow-up, or if your business is concentrated in rentals or commercial where these residential models transfer poorly.

On the deployment calendar, spring remains structurally favorable: listing volumes peak between March and June, so scoring refreshes completed by January position outreach ahead of the wave. Models trained on rolling 24-month windows with quarterly recalibration held up best through the volatility of 2023–2025, and that cadence remains the sensible default heading into late 2026.

Where AI-Driven Matching Fits In

Prediction is only half the value chain; matching is the other half. Knowing a household is likely to sell is less useful than knowing what they would buy next and connecting them to it before the listing goes public. This is the thesis behind AI-driven property discovery platforms: combine seller-propensity scoring with buyer-preference inference, then match off-market or soon-to-be-listed inventory to likely buyers on both ends. For consumers, the practical benefit is earlier visibility into homes that fit their actual criteria rather than filtered portal searches; for agents, it converts a cold propensity score into a warm two-sided introduction. The platforms executing this well treat the prediction as a ranking input to a discovery experience rather than a lead-scoring product sold by the thousand names, which tends to produce better long-term engagement and fewer compliance headaches. Expect the next 18 months to push further toward agentic systems that monitor a matched pipeline continuously and alert both parties when conditions align — financing readiness, target inventory availability, and seller timing converging in the same window.