Market Insights9 min read
The Comps Engine: How Comparable Sales Actually Shape a Property Price Prediction
PA
PropertyLens AI## Why Two Valuations of the Same Property Can Differ by $80,000
In the 12 months to June 2026, the median house price across inner Brisbane moved within a band of roughly $200,000 depending on the quarter you measured. That volatility isn't just market noise — it's a direct illustration of why comparable sales analysis is both powerful and genuinely difficult to do well.
Get the comps right, and a price prediction lands within 3–5% of the eventual sale price. Get them wrong — pull the wrong radius, use stale data, ignore a key feature — and you can be off by $80,000 to $150,000 on a $1.2M property. That's not a rounding error. That's a failed finance application or a buyer who overpaid by more than their annual salary.
So how does a comps engine actually work? What decisions get made, and why do two engines looking at the same property sometimes produce meaningfully different results?
## What a Comparable Sale Actually Is
A comparable sale — a "comp" — is a recently sold property that resembles the subject property closely enough that its sale price can inform what the subject property is worth. The word "resembles" is doing a lot of work in that sentence.
No two properties are identical. A 3-bedroom house on 450sqm in Paddington might sit next to another 3-bedroom house on 450sqm in Paddington, and one could be worth $300,000 more than the other because one was renovated in 2024 and one hasn't been touched since 1987. Comps analysis is the discipline of finding the closest available matches and then making principled adjustments for the differences.
At its core, every comps engine is making five decisions:
1. **Geographic radius** — how far from the subject property to search
2. **Recency window** — how far back in time to look
3. **Property type** — houses, townhouses, units treated separately or together
4. **Bedroom and bathroom count** — the most basic size filter
5. **Land size** — particularly important for houses in Brisbane's inner-ring suburbs
Each of these decisions involves trade-offs, and different engines make different calls.
## The Radius Problem
Start with geography. Brisbane's inner suburbs are not homogeneous grids. Paddington's character changes street by street — the ridge-top streets commanding valley views are a different market from the lower-lying streets closer to Given Terrace. A 500m radius in Paddington might cross three distinct micro-markets. A 1km radius in Kenmore might be entirely consistent.
A naive comps engine draws a circle and takes everything inside it. A more sophisticated one weights by proximity — sales 200m away count more than sales 900m away — and applies suburb boundary logic to avoid pulling comps from a different market.
The practical consequence: if a property in Bulimba sits near the boundary with Hawthorne, a comps engine that includes Hawthorne sales might produce a different result than one that stays strictly within Bulimba. Neither is automatically wrong. But if Hawthorne has transacted more recently at higher prices, the Hawthorne-inclusive engine will skew higher.
For tightly held streets — think parts of Ascot or Hamilton where turnover is low — engines sometimes have to stretch the radius to find enough comps. That stretch introduces noise. A good engine flags when it's had to reach further than ideal.
## The Recency Window: Fresh Data vs. Enough Data
Brisbane's inner-ring market moved materially in 2023 and 2024. Using a 24-month lookback in mid-2026 means including sales from mid-2024, when conditions were different. Using only a 6-month window gives you fresh data but might leave you with only 3 comparable sales — not enough to be statistically meaningful.
The tension is real. Stale comps understate current value in a rising market and overstate it in a falling one. Too-narrow windows leave you data-starved.
Most professional valuation engines use a rolling 12-month window as the default, with time-adjustment factors applied to older sales. A sale from 14 months ago gets discounted slightly to account for market movement since then. The discount rate itself is derived from the suburb's observed price trajectory — which means the engine is already making a market-trend assumption before it's even started the comps analysis.
This is one reason two engines diverge: they may use different lookback periods, and they may apply different time-adjustment factors based on different underlying trend data.
## Property Type: Why You Can't Mix Houses and Units
This sounds obvious, but it matters more than most buyers realise. In suburbs like West End or Fortitude Valley, the unit market and the house market operate on almost entirely different demand drivers, buyer profiles, and price trajectories. A unit in West End at $650,000 tells you almost nothing about what a house in West End is worth.
The subtler problem is townhouses. A townhouse in Woolloongabba might be priced closer to a small house than to a unit, or vice versa, depending on its land component, body corporate structure, and configuration. Engines that bucket all attached dwellings together as "units" will produce worse predictions for townhouses than engines that treat them as a distinct category.
For Brisbane's inner suburbs, where the mix of property types is dense and varied, this categorisation decision has a measurable impact on prediction accuracy.
## Bedroom Count and the Adjustment Problem
A 4-bedroom house and a 3-bedroom house on the same street in Annerley are not the same product. But if there have only been two 4-bedroom sales in the suburb in the past 12 months, the engine faces a choice: use those two thin comps, or pull in 3-bedroom comps and apply an adjustment.
The adjustment approach requires the engine to know what an extra bedroom is worth in that specific suburb. In Annerley, that might be $60,000–$80,000. In Ascot, it might be $150,000+. Getting this wrong by suburb is a common source of prediction error.
Bathroom count adds another layer. A 4-bedroom, 2-bathroom house and a 4-bedroom, 1-bathroom house in the same street in Tarragindi can differ by $50,000–$80,000 at current prices. Engines that only filter on bedrooms will blend these together and produce a muddier estimate.
## Land Size: Brisbane's Most Underweighted Variable
In Sydney, apartments dominate the inner ring and land size matters less. In Brisbane, the inner suburbs are still substantially characterised by houses on individual lots, and land size is often the single most important variable after location.
Consider two houses in Coorparoo: one on 405sqm, one on 607sqm. Both 3-bedroom, both renovated, same street. The larger block might be worth $120,000–$180,000 more — not because the house is better, but because the land carries development optionality or simply more usable outdoor space.
A comps engine that doesn't weight for land size will systematically misprice properties at the extremes. Small-block properties get overvalued. Large-block properties get undervalued. For a buyer trying to assess whether a 607sqm block in Coorparoo is priced fairly relative to recent sales, this matters enormously.
The best engines apply a land-size adjustment that's calibrated by suburb — because the value of an extra 100sqm in Paddington is different from the value of an extra 100sqm in Keperra.
## Why Two Engines Produce Different Numbers
By now the answer should be becoming clear. Two comps engines can look at the same property and produce results $80,000–$120,000 apart because they've made different legitimate choices across five dimensions:
- Engine A uses a 12-month window; Engine B uses 18 months
- Engine A applies a 600m radius; Engine B uses 1km and crosses a suburb boundary
- Engine A treats townhouses as a separate category; Engine B groups them with units
- Engine A applies a land-size adjustment; Engine B doesn't
- Engine A weights recent sales more heavily; Engine B treats all comps equally
None of these differences necessarily makes one engine right and the other wrong. They reflect different methodological assumptions. What matters is whether those assumptions are appropriate for the specific property type, suburb, and market conditions.
This is also why human valuers — and good AI systems — don't just run the algorithm and accept the output. They look at which comps were selected, whether they're genuinely comparable, and whether any of them are outliers that should be excluded or down-weighted.
## The Three-Layer Approach to Prediction
A single comps engine, however well-calibrated, is a single point of view. The more robust approach layers multiple methods:
**Layer 1: Comparable sales analysis** — the comps engine as described above, identifying the most similar recent transactions and adjusting for differences.
**Layer 2: Feature-based valuation** — a statistical model (typically a hedonic regression or machine learning variant) trained on thousands of transactions, which prices individual features: bedrooms, bathrooms, land size, renovation status, proximity to transport, school catchment. This layer doesn't depend on finding close comps — it builds a price from the property's attributes.
**Layer 3: Contextual analysis** — incorporating suburb-level trend data, days-on-market movements, clearance rates, and any known factors specific to the property (flood overlay, heritage listing, corner block premium, busy road discount).
When all three layers converge on a similar number, confidence is high. When they diverge, that divergence is itself informative — it suggests the property has unusual characteristics that the comps engine is struggling to price, or that the market is in a transitional phase where historical data is less predictive.
## What This Means for Buyers and Investors
If you're using any price prediction tool — whether it's a bank valuation, an automated estimate, or an AI-powered report — the single most useful question you can ask is: *which comps did it use?*
A prediction based on 12 genuinely comparable recent sales is worth far more than one based on 3 sales from 18 months ago that are only loosely similar. Transparency about comp selection is a proxy for the quality of the whole analysis.
For buyers in Brisbane's inner ring, a few practical implications:
- **In tightly held streets** (parts of New Farm, Ascot, Hamilton), thin transaction volumes mean comps engines have to reach further. Treat estimates in these areas with wider error bands.
- **For properties with unusual land sizes** — particularly large blocks in Paddington, Bardon, or Tarragindi — check whether the estimate accounts for land size properly. Many automated tools don't.
- **For townhouses** in suburbs like West End, Newstead, or Woolloongabba, ask whether the tool is comparing your property to other townhouses or mixing in unit data.
- **In fast-moving markets**, a 12-month lookback may include comps from a different rate environment. A good tool will flag this.
## The Accuracy Question
No prediction is a guarantee. A well-constructed comps analysis on a typical 3-bedroom house in a liquid Brisbane suburb — somewhere like Morningside, Coorparoo, or Mitchelton — should land within 4–6% of the eventual sale price in normal market conditions. That's roughly $50,000–$75,000 on a $1.2M property.
On unusual properties — large blocks, heritage-listed homes, properties with significant renovation upside — the error band widens. Not because the tools are bad, but because the market itself is less certain about how to price them. Fewer buyers compete for unusual properties, which means the eventual sale price depends more on who shows up on auction day than on any algorithmic analysis.
Tracking prediction accuracy publicly — showing how past predictions compared to actual sale prices — is the only honest way to demonstrate that a prediction engine is calibrated rather than just confident.
## Putting It Together
Comparable sales analysis is not a black box. It's a series of explicit methodological choices about radius, recency, property type, bedroom count, and land size — choices that any good tool should be able to explain and that any informed buyer should understand.
When two estimates diverge significantly, the answer isn't to average them and move on. It's to look at the underlying comps each engine selected and ask which set is more genuinely comparable to the property you're evaluating.
PropertyLens's price prediction reports show the comparable sales used in each analysis, along with the adjustments applied — so you can see the reasoning rather than just the number. If you're researching a property in Brisbane's inner suburbs, the free estimate tool at app.propertylens.au/estimate gives you a starting point, and the detailed report goes deeper into the comps layer for properties where the methodology matters most.
In the 12 months to June 2026, the median house price across inner Brisbane moved within a band of roughly $200,000 depending on the quarter you measured. That volatility isn't just market noise — it's a direct illustration of why comparable sales analysis is both powerful and genuinely difficult to do well.
Get the comps right, and a price prediction lands within 3–5% of the eventual sale price. Get them wrong — pull the wrong radius, use stale data, ignore a key feature — and you can be off by $80,000 to $150,000 on a $1.2M property. That's not a rounding error. That's a failed finance application or a buyer who overpaid by more than their annual salary.
So how does a comps engine actually work? What decisions get made, and why do two engines looking at the same property sometimes produce meaningfully different results?
## What a Comparable Sale Actually Is
A comparable sale — a "comp" — is a recently sold property that resembles the subject property closely enough that its sale price can inform what the subject property is worth. The word "resembles" is doing a lot of work in that sentence.
No two properties are identical. A 3-bedroom house on 450sqm in Paddington might sit next to another 3-bedroom house on 450sqm in Paddington, and one could be worth $300,000 more than the other because one was renovated in 2024 and one hasn't been touched since 1987. Comps analysis is the discipline of finding the closest available matches and then making principled adjustments for the differences.
At its core, every comps engine is making five decisions:
1. **Geographic radius** — how far from the subject property to search
2. **Recency window** — how far back in time to look
3. **Property type** — houses, townhouses, units treated separately or together
4. **Bedroom and bathroom count** — the most basic size filter
5. **Land size** — particularly important for houses in Brisbane's inner-ring suburbs
Each of these decisions involves trade-offs, and different engines make different calls.
## The Radius Problem
Start with geography. Brisbane's inner suburbs are not homogeneous grids. Paddington's character changes street by street — the ridge-top streets commanding valley views are a different market from the lower-lying streets closer to Given Terrace. A 500m radius in Paddington might cross three distinct micro-markets. A 1km radius in Kenmore might be entirely consistent.
A naive comps engine draws a circle and takes everything inside it. A more sophisticated one weights by proximity — sales 200m away count more than sales 900m away — and applies suburb boundary logic to avoid pulling comps from a different market.
The practical consequence: if a property in Bulimba sits near the boundary with Hawthorne, a comps engine that includes Hawthorne sales might produce a different result than one that stays strictly within Bulimba. Neither is automatically wrong. But if Hawthorne has transacted more recently at higher prices, the Hawthorne-inclusive engine will skew higher.
For tightly held streets — think parts of Ascot or Hamilton where turnover is low — engines sometimes have to stretch the radius to find enough comps. That stretch introduces noise. A good engine flags when it's had to reach further than ideal.
## The Recency Window: Fresh Data vs. Enough Data
Brisbane's inner-ring market moved materially in 2023 and 2024. Using a 24-month lookback in mid-2026 means including sales from mid-2024, when conditions were different. Using only a 6-month window gives you fresh data but might leave you with only 3 comparable sales — not enough to be statistically meaningful.
The tension is real. Stale comps understate current value in a rising market and overstate it in a falling one. Too-narrow windows leave you data-starved.
Most professional valuation engines use a rolling 12-month window as the default, with time-adjustment factors applied to older sales. A sale from 14 months ago gets discounted slightly to account for market movement since then. The discount rate itself is derived from the suburb's observed price trajectory — which means the engine is already making a market-trend assumption before it's even started the comps analysis.
This is one reason two engines diverge: they may use different lookback periods, and they may apply different time-adjustment factors based on different underlying trend data.
## Property Type: Why You Can't Mix Houses and Units
This sounds obvious, but it matters more than most buyers realise. In suburbs like West End or Fortitude Valley, the unit market and the house market operate on almost entirely different demand drivers, buyer profiles, and price trajectories. A unit in West End at $650,000 tells you almost nothing about what a house in West End is worth.
The subtler problem is townhouses. A townhouse in Woolloongabba might be priced closer to a small house than to a unit, or vice versa, depending on its land component, body corporate structure, and configuration. Engines that bucket all attached dwellings together as "units" will produce worse predictions for townhouses than engines that treat them as a distinct category.
For Brisbane's inner suburbs, where the mix of property types is dense and varied, this categorisation decision has a measurable impact on prediction accuracy.
## Bedroom Count and the Adjustment Problem
A 4-bedroom house and a 3-bedroom house on the same street in Annerley are not the same product. But if there have only been two 4-bedroom sales in the suburb in the past 12 months, the engine faces a choice: use those two thin comps, or pull in 3-bedroom comps and apply an adjustment.
The adjustment approach requires the engine to know what an extra bedroom is worth in that specific suburb. In Annerley, that might be $60,000–$80,000. In Ascot, it might be $150,000+. Getting this wrong by suburb is a common source of prediction error.
Bathroom count adds another layer. A 4-bedroom, 2-bathroom house and a 4-bedroom, 1-bathroom house in the same street in Tarragindi can differ by $50,000–$80,000 at current prices. Engines that only filter on bedrooms will blend these together and produce a muddier estimate.
## Land Size: Brisbane's Most Underweighted Variable
In Sydney, apartments dominate the inner ring and land size matters less. In Brisbane, the inner suburbs are still substantially characterised by houses on individual lots, and land size is often the single most important variable after location.
Consider two houses in Coorparoo: one on 405sqm, one on 607sqm. Both 3-bedroom, both renovated, same street. The larger block might be worth $120,000–$180,000 more — not because the house is better, but because the land carries development optionality or simply more usable outdoor space.
A comps engine that doesn't weight for land size will systematically misprice properties at the extremes. Small-block properties get overvalued. Large-block properties get undervalued. For a buyer trying to assess whether a 607sqm block in Coorparoo is priced fairly relative to recent sales, this matters enormously.
The best engines apply a land-size adjustment that's calibrated by suburb — because the value of an extra 100sqm in Paddington is different from the value of an extra 100sqm in Keperra.
## Why Two Engines Produce Different Numbers
By now the answer should be becoming clear. Two comps engines can look at the same property and produce results $80,000–$120,000 apart because they've made different legitimate choices across five dimensions:
- Engine A uses a 12-month window; Engine B uses 18 months
- Engine A applies a 600m radius; Engine B uses 1km and crosses a suburb boundary
- Engine A treats townhouses as a separate category; Engine B groups them with units
- Engine A applies a land-size adjustment; Engine B doesn't
- Engine A weights recent sales more heavily; Engine B treats all comps equally
None of these differences necessarily makes one engine right and the other wrong. They reflect different methodological assumptions. What matters is whether those assumptions are appropriate for the specific property type, suburb, and market conditions.
This is also why human valuers — and good AI systems — don't just run the algorithm and accept the output. They look at which comps were selected, whether they're genuinely comparable, and whether any of them are outliers that should be excluded or down-weighted.
## The Three-Layer Approach to Prediction
A single comps engine, however well-calibrated, is a single point of view. The more robust approach layers multiple methods:
**Layer 1: Comparable sales analysis** — the comps engine as described above, identifying the most similar recent transactions and adjusting for differences.
**Layer 2: Feature-based valuation** — a statistical model (typically a hedonic regression or machine learning variant) trained on thousands of transactions, which prices individual features: bedrooms, bathrooms, land size, renovation status, proximity to transport, school catchment. This layer doesn't depend on finding close comps — it builds a price from the property's attributes.
**Layer 3: Contextual analysis** — incorporating suburb-level trend data, days-on-market movements, clearance rates, and any known factors specific to the property (flood overlay, heritage listing, corner block premium, busy road discount).
When all three layers converge on a similar number, confidence is high. When they diverge, that divergence is itself informative — it suggests the property has unusual characteristics that the comps engine is struggling to price, or that the market is in a transitional phase where historical data is less predictive.
## What This Means for Buyers and Investors
If you're using any price prediction tool — whether it's a bank valuation, an automated estimate, or an AI-powered report — the single most useful question you can ask is: *which comps did it use?*
A prediction based on 12 genuinely comparable recent sales is worth far more than one based on 3 sales from 18 months ago that are only loosely similar. Transparency about comp selection is a proxy for the quality of the whole analysis.
For buyers in Brisbane's inner ring, a few practical implications:
- **In tightly held streets** (parts of New Farm, Ascot, Hamilton), thin transaction volumes mean comps engines have to reach further. Treat estimates in these areas with wider error bands.
- **For properties with unusual land sizes** — particularly large blocks in Paddington, Bardon, or Tarragindi — check whether the estimate accounts for land size properly. Many automated tools don't.
- **For townhouses** in suburbs like West End, Newstead, or Woolloongabba, ask whether the tool is comparing your property to other townhouses or mixing in unit data.
- **In fast-moving markets**, a 12-month lookback may include comps from a different rate environment. A good tool will flag this.
## The Accuracy Question
No prediction is a guarantee. A well-constructed comps analysis on a typical 3-bedroom house in a liquid Brisbane suburb — somewhere like Morningside, Coorparoo, or Mitchelton — should land within 4–6% of the eventual sale price in normal market conditions. That's roughly $50,000–$75,000 on a $1.2M property.
On unusual properties — large blocks, heritage-listed homes, properties with significant renovation upside — the error band widens. Not because the tools are bad, but because the market itself is less certain about how to price them. Fewer buyers compete for unusual properties, which means the eventual sale price depends more on who shows up on auction day than on any algorithmic analysis.
Tracking prediction accuracy publicly — showing how past predictions compared to actual sale prices — is the only honest way to demonstrate that a prediction engine is calibrated rather than just confident.
## Putting It Together
Comparable sales analysis is not a black box. It's a series of explicit methodological choices about radius, recency, property type, bedroom count, and land size — choices that any good tool should be able to explain and that any informed buyer should understand.
When two estimates diverge significantly, the answer isn't to average them and move on. It's to look at the underlying comps each engine selected and ask which set is more genuinely comparable to the property you're evaluating.
PropertyLens's price prediction reports show the comparable sales used in each analysis, along with the adjustments applied — so you can see the reasoning rather than just the number. If you're researching a property in Brisbane's inner suburbs, the free estimate tool at app.propertylens.au/estimate gives you a starting point, and the detailed report goes deeper into the comps layer for properties where the methodology matters most.