The Number Behind Every Estimate

The median house price across Greater Brisbane hit approximately $1.02 million in mid-2026, up from $870,000 at the start of 2024. That single figure gets quoted constantly. What gets quoted far less often is how anyone arrives at a price estimate for a *specific* property — a three-bedroom Queenslander on a 607 sqm block in Annerley, or a four-bedroom brick house backing onto a creek in Carindale.

That's where price prediction tools come in. And the gap between a good one and a mediocre one can be $80,000 or more on a single estimate.

Understanding how these tools work isn't just academic. If you're buying in Paddington at $1.4M, or selling in Chermside and wondering whether to list at $850,000 or $920,000, the method behind any estimate you're relying on matters enormously.

There are three core approaches used across Australian property prediction tools: comparable sales analysis, hedonic regression modelling, and machine learning. Most serious platforms use some combination of all three. Here's what each one actually does.

Method One: Comparable Sales Analysis

This is the oldest method and still the most intuitive. The idea is simple: find recently sold properties that closely resemble the one you're trying to value, then adjust for differences.

A property valuer doing this manually might look at five or six sales within a two-kilometre radius over the past six months, then apply judgment-based adjustments. A bedroom more adds value. A busy road subtracts it. A renovated kitchen adds something. How much? That's where human judgment — and human inconsistency — enters.

Automated comparable sales systems do this at scale. They pull from databases of settled sales (typically sourced from state land title registries and aggregated through property data providers), filter by proximity, recency, and property type, then apply weighting algorithms to rank which comparables are most relevant.

What it does well: Comparable sales analysis is grounded in actual transaction data. In active markets — inner Brisbane suburbs like Woolloongabba, Morningside, or Ashgrove, where dozens of similar properties sell each year — it performs reliably.

Where it struggles: Thin markets. If you're trying to estimate a property in a suburb where only eight houses sold in the past twelve months, and your subject property has unusual features, the comparable pool shrinks fast. Accuracy degrades. The same problem hits unusual properties anywhere: a heritage-listed home, a house on a 1,200 sqm block in an area dominated by 400 sqm lots, or a property with a dual occupancy setup.

Recency is also a constraint. In a market that moved 6% in six months — as parts of Brisbane's middle ring did through 2024 — comparables from seven months ago are already stale. Good systems weight recent sales more heavily and apply time-adjustment factors. Simpler ones don't.

Method Two: Hedonic Regression Modelling

This is the statistical backbone of most serious automated valuation models (AVMs) used by banks, mortgage insurers, and property platforms across Australia.

Hedonic regression treats a property's price as the sum of its individual characteristics — each feature contributing a measurable dollar value. Bedrooms, bathrooms, land size, car spaces, property age, construction type, proximity to train stations, school catchment zones, flood overlay status, and dozens of other variables are fed into a regression equation. The model estimates the marginal price contribution of each feature based on historical sales data.

For example, a well-calibrated hedonic model for Brisbane's inner north might determine that, all else equal, a second bathroom adds approximately $45,000–$65,000 to a house price in that submarket. A 100 sqm increase in land size might add $55,000–$80,000. Being in the Kelvin Grove State College catchment versus a lower-ranked catchment might add a measurable premium on top of that.

What it does well: Hedonic models handle large datasets efficiently and produce consistent, explainable outputs. They're particularly good at capturing the structural value of a property — the things you can measure and count. Major banks and lenders rely heavily on hedonic AVMs for mortgage assessments because they're auditable and systematic.

Where it struggles: Hedonic models are only as good as the variables they include and the data quality behind them. Condition and presentation — arguably the biggest driver of price variation between two otherwise identical properties — are almost impossible to capture in a regression model. A renovated Queenslander in Paddington and an unrenovated one on the same street might differ by $200,000, but both have the same bedroom count, land size, and suburb code.

Models also struggle with structural market shifts. A hedonic model trained on 2023–2025 data may underweight the premium buyers are now paying for energy efficiency features, or overweight proximity to certain retail precincts that have declined. Recalibration frequency matters.

Method Three: Machine Learning and AI Analysis

This is where property prediction has evolved most significantly over the past five years. Machine learning models — particularly gradient boosting algorithms and neural networks — can identify non-linear relationships between variables that traditional regression misses entirely.

A hedonic regression assumes that the value of a third bedroom is roughly constant across a suburb. A machine learning model can detect that a third bedroom adds $90,000 in the $900,000–$1.1M price band in Coorparoo but adds almost nothing in the $600,000–$750,000 band, because buyers at different price points have different needs. It can find these patterns automatically, without a human analyst specifying them in advance.

More advanced implementations layer in unstructured data: satellite imagery to assess property condition and land coverage, street view analysis, planning documents, infrastructure announcements, and even sentiment signals from listing descriptions. AI tools can also synthesise qualitative market intelligence — the kind of contextual reasoning that asks "what does the recent rezoning of this corridor mean for land values here in three years?" — in ways that pure statistical models cannot.

What it does well: Handling complexity and non-linearity. Detecting patterns across large, messy datasets. Incorporating information types that don't fit neatly into a regression equation. In high-transaction suburbs with rich data, machine learning models consistently outperform simpler AVMs on accuracy metrics.

Where it struggles: Interpretability. A gradient boosting model might produce a highly accurate estimate but offer limited explanation for why. For buyers and sellers who want to understand the reasoning — not just the number — this is a real limitation. Machine learning models also require large amounts of training data, which means they can underperform in low-volume markets or for highly unusual properties.

They can also overfit to recent market conditions and produce overconfident estimates during rapid market transitions. The Brisbane market's sharp correction in late 2022 caught several automated tools badly wrong precisely because their models had been trained on a sustained growth environment.

What Accuracy Looks Like in Practice

No price prediction tool is perfectly accurate. The honest question is: how wrong, and how often?

The industry benchmark used by most serious AVM providers is the percentage of estimates falling within 10% of the eventual sale price. In well-data-rich Australian capital city markets, leading tools typically achieve 70–80% of estimates within 10% of sale price. In thinner markets or for unusual properties, that figure drops to 55–65%.

That means on a $1,000,000 property, a quality tool should land within $100,000 of the sale price roughly three-quarters of the time. On the remaining quarter, the error can be larger — sometimes significantly so.

The practical implication: price estimates are a starting point for analysis, not a substitute for it. A buyer using an estimate to decide whether to even inspect a property is using it correctly. A buyer using an estimate as their sole basis for a $1.2M offer is using it dangerously.

Median error (as opposed to the within-10% rate) is also worth understanding. A tool might achieve 75% within 10% but have a systematic bias — consistently underestimating renovated properties, or overestimating properties near flood-affected areas. Platforms that publish their accuracy metrics transparently, broken down by suburb and property type, are far more useful than those that report a single headline figure.

How PropertyLens Combines All Three

PropertyLens uses a three-layer prediction engine that runs comparable sales analysis, feature-based hedonic modelling, and AI analysis in parallel, then synthesises the outputs.

The comparable sales layer identifies the most relevant recent transactions within the target suburb and surrounding area, weighted by recency, proximity, and property similarity. The hedonic layer applies a feature-based valuation using property attributes drawn from council records, land title data, and listing history. The AI layer — powered by Claude with web search capability — adds contextual research: recent sales not yet in the database, planning changes, infrastructure developments, and qualitative market signals that the statistical layers can't capture.

The three outputs are reconciled into a price range rather than a single point estimate. This is deliberate. A range of $980,000–$1,080,000 is more honest than a single figure of $1,030,000, because it communicates the genuine uncertainty in any property estimate. The width of the range itself carries information: a narrow range on a standard three-bedroom house in Nundah reflects high data confidence; a wider range on an unusual property in a thin market reflects lower confidence.

The models are recalibrated weekly against new settled sales data. In a market moving as quickly as Brisbane's has over the past two years, weekly recalibration makes a meaningful difference to estimate freshness. A tool recalibrated quarterly can be significantly behind in a rising or falling market.

Accuracy tracking is published publicly at app.propertylens.au/predictions — not because every estimate is perfect, but because transparency about where the model performs well and where it doesn't is more useful to buyers and investors than a marketing claim.

What to Look for When Comparing Prediction Tools

If you're evaluating property price prediction tools — whether PropertyLens or any other platform — these are the questions worth asking:

Does it publish accuracy metrics? And are they broken down by suburb, property type, and price band — or just a single headline number?
How frequently is it recalibrated? Weekly recalibration against new sales data is meaningfully better than monthly or quarterly in an active market.
Does it give a range or a point estimate? A single number implies false precision. A range is more honest.
What data sources does it draw on? Land title registry data, council records, and listing history are the minimum. Tools that also incorporate planning data, flood overlays, and infrastructure announcements are capturing more of what actually drives prices.
Can it explain its reasoning? For a buyer trying to understand why a property is estimated at a particular level, explainability matters. A tool that just outputs a number is less useful than one that shows which comparables it used and which features it weighted most heavily.
Does it handle your specific property type and suburb? A tool with strong performance in inner-city Brisbane units may be much weaker on large-block properties in the outer south. Ask specifically about performance in your target area.

The Limits No Tool Can Overcome

Every price prediction tool — no matter how sophisticated — has limits that are structural, not technical.

Presentation and condition at the time of sale are largely invisible to data systems. A vendor who has spent $80,000 on a targeted renovation in the six months before listing will likely achieve a price that no model predicted, because the model couldn't see the renovation. Negotiation dynamics on the day — a motivated vendor, two competing buyers, a delayed settlement condition — can move a price 3–5% in either direction regardless of what any model says.

Market sentiment shifts are also hard to capture in real time. The period between when buyer sentiment changes and when that change shows up in settled sales data — typically six to twelve weeks — is a blind spot for all data-driven tools.

The right mental model is this: a good price prediction tool narrows the range of uncertainty. It doesn't eliminate it. Used alongside a building and pest inspection, a flood overlay check, a review of recent comparable sales, and ideally a conversation with a local agent who knows the street, it becomes genuinely powerful. Used in isolation, it's just a number.

For Brisbane buyers doing their own research, PropertyLens offers free suburb-level estimates at app.propertylens.au/estimate, with detailed AI-powered prediction reports available for any specific address — including the comparable sales and feature analysis that sits behind the estimate.