DocsCompliance & assurance · AVM
Performance & Validation
Valtaic evaluates its AVM internally through development and release testing. The assessment combines historical accuracy analysis with checks of valuation behaviour and deployed API responses.
Evaluation Approach
| Assessment | Purpose |
|---|---|
| Historical accuracy | Compare estimates with recorded transactions using separate training, validation and test periods. |
| Segmented analysis | Examine results across property types, locations and selected property populations. |
| Scenario checks | Examine how estimates respond to changes in supplied property characteristics. |
| Property diagnostics | Investigate individual results and their supporting evidence. |
| Serving checks | Confirm that deployed responses preserve the expected valuation behaviour and API contract. |
Each assessment has a defined scope. Historical model results describe the model and population tested. An assessment of the complete API also needs to cover the active release, endpoint, input handling and outcome policy. Chronological separation should be assessed alongside the information available at each historical valuation date.
Performance Measures
Read accuracy alongside the proportion of properties receiving a usable estimate.
| Measure | Interpretation |
|---|---|
| Median absolute percentage error | The middle absolute percentage error across evaluated properties. |
| Within ±10% and ±20% | The proportion of estimates within each tolerance of the observed sale price. |
| Signed error | The direction of error; positive values indicate overvaluation relative to the sale. |
| Large-error rates | The frequency and size of errors beyond a stated threshold, especially overvaluations. |
| Acceptance and referral rates | The proportion of the submitted population receiving each outcome. |
| Range coverage and width | How often observed outcomes fall within returned ranges, considered alongside how wide those ranges are. |
For a positive observed sale price, percentage error is 100 × (estimate − sale price) / sale price. Absolute percentage error uses its absolute value.
Performance evidence should identify the release, evaluation dates, sample sizes, exclusions and property segments. Sale timing, transaction eligibility and related properties affect interpretation.
Coverage and Referrals
Report submitted, eligible, valued, referred, rejected and failed cases with explicit denominators. Accuracy for accepted cases should be assessed together with the acceptance rate and the characteristics of referred properties.
Review results across geography, property type, tenure, price band and development status where relevant. Segment sizes and uncertainty help distinguish a reliable pattern from a small-sample result.
Prediction Ranges
Nominal coverage is the intended coverage associated with a prediction range. Observed coverage measures how often eligible outcomes fall inside their returned ranges on an evaluation population.
Read coverage with interval width and the proportion of results for which a range is available. Confidence is a separate evidence-strength signal: interpret it according to the response definition, rather than as a percentage accuracy figure.
Assessing a Release
For adoption, request evidence matched to the release_id, endpoint and intended use. Establish acceptance criteria for your property population, review representative cases and define when a result needs professional review.
Setting an earlier valuation_date uses the live service's supported historical-date behaviour. A point-in-time accuracy assessment additionally needs a defined historical information set and evaluation design. See Versioning and Releases.
API examples illustrate response structure. Use a release-matched evaluation for performance decisions and discuss available evidence through hello@valtaic.io.