Repeat-sales methodology estimates property-price change from multiple transactions on the same assets, reducing sales-mix bias while introducing distinct sample risks.
Repeat-sales methodology estimates property-price change by comparing two or more eligible transactions involving the same property and combining many such pairs in a statistical model. Matching a property with itself reduces distortion from changes in the mix of homes sold, but it does not automatically adjust for renovations, deterioration, or the fact that frequently traded properties may differ from the wider housing stock.
The method is best known for residential House Price Index (HPI) construction. It can also be applied to other assets with stable identifiers and repeated observations, although the data requirements and biases may differ.
The repeat-sales methodology is the procedure used to select pairs, estimate period effects, weight observations, aggregate results, and revise the series. A repeat-sales index is the numerical output of that procedure.
That distinction matters because two providers can both use repeat sales and still publish different results. Their methodologies may differ in:
Calling an index “repeat sales” therefore describes its broad design, not every rule needed to reproduce it.
A compiler generally needs:
| Data element | Why it matters |
|---|---|
| Stable property identifier | Matches the same parcel or dwelling across records |
| Transaction or valuation date | Assigns each observation to a model period |
| Sale price or eligible valuation | Provides the price relative for a matched pair |
| Transaction type | Helps distinguish market sales from transfers, refinances, or other events |
| Property type and geography | Supports eligibility filters and local aggregation |
| Record-quality fields | Help identify duplicates, incomplete addresses, implausible values, or coding errors |
| Property-change information, when available | Can flag major renovations, demolition, subdivision, or conversion |
The exact fields depend on the index. A public-record index may use deed and assessor data. A mortgage-based index may use purchase prices and refinance appraisals tied to loan records. Those sources observe different parts of the market.
The compiler defines the target market before matching properties. Rules may specify:
Transfers without a meaningful market price, duplicate records, incomplete identifiers, and implausible observations are normally excluded. Provider rules can also remove or separately classify foreclosures, short sales, rapid resales, and properties whose type changed.
Eligible records are grouped by a stable property identifier. A home sold in two periods produces one sales pair. If it sells three times, the methodology may form multiple pairs depending on the provider’s rules.
Accurate matching is not trivial. Parcel identifiers can change after subdivision or consolidation, addresses can be recorded inconsistently, and condominium unit numbers can be missing. A false match combines different properties; a missed match discards usable information.
For property (i), observed in earlier period (s) and later period (t), a simplified repeat-sales return is:
where:
The logarithm turns a price ratio across several periods into an additive change. It does not reveal how much of that change occurred in each intervening month or quarter; the regression estimates those period effects using all overlapping pairs.
A simplified model is:
where:
The model uses many pairs with different start and end dates. Together, they form a network that lets the regression infer market-level movement across periods.
If the base index is 100, a simplified index level can be recovered as:
Production methodologies can use different parameterizations, weighting stages, aggregation rules, and robustness procedures. These equations explain the core logic; they are not a substitute for a specific provider’s methodology document.
An unweighted model gives each eligible pair the same regression weight. A weighted repeat-sales model recognizes that error variance may differ across observations.
Longer holding periods often create more opportunity for unobserved renovations, deterioration, or other property-specific changes. A common three-stage approach is:
Pairs expected to be noisier receive less influence. Some production indexes add robust weights to limit extreme observations or use value weights when aggregating market segments. The phrase weighted repeat sales does not identify one universal weighting formula.
Local estimates may be combined into regional or national indexes. Providers can weight component areas by housing stock, transaction value, market value, or another measure. The aggregation choice determines how much influence each market has on the published total.
The final index is normalized to a base such as 100. Changing the base rescales index levels without changing period-to-period growth rates.
Consider three eligible sales pairs:
| Property | Earlier observation | Later observation | Raw pair change |
|---|---|---|---|
| A | Q1: $300,000 | Q2: $315,000 | 5.0% |
| B | Q1: $400,000 | Q3: $436,000 | 9.0% |
| C | Q2: $250,000 | Q3: $260,000 | 4.0% |
Property B spans both intervals. The regression uses all three pairs to estimate a Q1-to-Q2 market effect and a Q2-to-Q3 market effect that best fit the observed log price relatives. It does not simply average 5%, 9%, and 4%, because those changes cover different periods.
For illustration, if the estimated effects were 5% in the first interval and 4% in the second, an index beginning at 100 would chain as follows:
1Q1 index = 100.0
2Q2 index = 100.0 x 1.05 = 105.0
3Q3 index = 105.0 x 1.04 = 109.2
That chain is illustrative rather than an estimate from the small table. A production index uses far more observations, statistical weights, quality filters, and explicit aggregation rules.
Repeat matching alone does not create a defensible index. Common controls include:
Filters improve consistency but can also change the represented population. Removing distressed sales, for example, may answer a useful question while producing an index that no longer describes the full transaction market. Users need the exact rule set, not only the index label.
The method naturally uses information that arrives over time. A property’s first transaction does not become a repeat pair until another eligible sale or valuation occurs. When the later observation arrives, the new pair adds information about price movement since the earlier date.
Revisions can also result from:
For historical research, the current series may differ from the data available to an analyst on the original decision date. Retaining release vintages is therefore important when testing forecasts or policies in real time.
| Question | Repeat sales | Hedonic regression | Stratification or mix adjustment |
|---|---|---|---|
| How is quality controlled? | Match each property with itself | Model observed property characteristics | Compare prices within defined groups |
| Must a property sell more than once? | Yes | No | No |
| Characteristic data required | Relatively limited, beyond matching and filters | Often extensive | Enough to define useful groups |
| Main coverage risk | Repeat-sale properties may not represent all housing | Missing or mismeasured attributes | Quality can still vary within groups |
| Renovation treatment | Usually imperfect unless changes are observed or filtered | Can include observed improvements | Usually indirect |
| Revision tendency | New pairs can revise earlier estimates | Re-estimation and new data can revise results | Weight and source revisions can change history |
| Small-area performance | Often constrained by too few repeat pairs | Depends on transaction count and attribute quality | Depends on sufficient observations in each group |
No method is automatically best for every market. Repeat sales is attractive when reliable property matching and long transaction histories exist. Hedonic methods may use more observations but require richer attribute data and model choices. Stratification is transparent but may leave material differences within each group.
The method treats a matched property as the same asset, but the structure can change. Remodeling, additions, damage, deferred maintenance, and normal aging can make the price relative differ from pure market appreciation.
Some providers remove suspected changes or reduce their influence. Complete renovation data are rarely available, so constant quality remains an approximation.
Only properties with multiple eligible observations enter the model. Frequently traded homes may differ from long-held homes in age, location, price, condition, investor ownership, or borrower characteristics.
The estimated trend therefore describes the repeat-observation sample most directly. It may not represent newly built homes, never-resold properties, or segments with low turnover.
Small areas and quiet markets may produce too few pairs for stable monthly or quarterly estimates. Transaction volume can also fall sharply during periods of financial stress, increasing uncertainty when users most want timely local evidence.
A deed-record index, mortgage index, and appraisal-based index observe different events. Cash purchases, jumbo mortgages, government-backed loans, refinances, and non-market transfers may enter one dataset but not another.
Weights, filters, aggregation, and seasonal procedures are modeling choices. New records can change previous estimates, and results can differ across providers even when both use valid repeat-sales methods.
$250,000 price.Repeat-sales indexes can support:
The output should be combined with property-level evidence, sales volume, inventory, income, mortgage rates, delinquency, and other relevant data. A broad repeat-sales index is not an appraisal, an affordability measure, or a forecast.
Before relying on a repeat-sales index, verify:
This article is educational and does not provide an appraisal, lending decision, investment recommendation, or individualized financial advice.