Table 1 Evidence map for part II:...

Table 1

Evidence map for part II: central findings, supporting sources, evidence strength and jurisdictional depth

Finding	Supporting sources	Evidence strength	Jurisdictional depth
Out-of-sample validation and backtesting are prerequisites for defensible AVM deployment	Schulz et al. (2014), Steurer et al. (2021), Krause et al. (2020)	Established	Broad
Prediction intervals, calibration and distributional or conformal methods should accompany every AVM output; in line with supervisory expectations	Rigby and Stasinopoulos (2005), Wheeler et al. (2014), Angelopoulos and Bates (2023), Granna et al. (2025), European Union (2024), Federal Register (2024)	Established (methods); emerging (regulatory codification)	EU, US, UK
Data quality, density and granularity set the upper bound of AVM performance more fundamentally than model complexity	Openshaw (1984), Granna et al. (2022), Krause et al. (2020), Steurer et al. (2021)	Established	Broad
Concept drift and temporal instability materially affect AVM reliability and require systematic monitoring, not ad hoc maintenance	Gama et al. (2014), Gudigantala and Mehrotra (2024)	Established	Primarily US illustration; drift theory jurisdiction-agnostic
Alternative and multimodal data can enrich AVM inputs, but added value is context-dependent and often proprietary	Koch et al. (2019), Despotovic and Brunauer (2024)	Emerging	Broad
ESG signals influence property values (green premiums, brown discounts, stranding risk), but ESG data infrastructure and standardisation lag behind regulatory demand	Eichholtz et al. (2010), Cajias et al. (2019), Hirsch et al. (2019), RICS (2024), European Union (2020, 2022)	Emerging	EU-focused
Synthetic data can support prototyping and stress testing but are not a substitute for real transaction data; risks include model collapse	Goodfellow et al. (2014), Kingma and Welling (2014), Shumailov et al. (2024), Alemohammad et al. (2023), Bidanset (2025)	Emerging	Broad
AVMs can reproduce and amplify historical bias; disparities across racial and socioeconomic groups are empirically documented in US markets	Howell and Korver-Glenn (2021), Freddie Mac (2021), Zhu et al. (2024), Neal et al. (2020), CFPB (2023)	Established (US); Under-researched (elsewhere)	US-focused
Formal fairness criteria cannot be jointly satisfied; mitigation is a governance choice, not a purely technical one	Kleinberg et al. (2017), Barocas et al. (2019)	Established (theoretical)	Broad
Hybrid human AVM architectures are the most resilient operating model under current regulatory expectations; practitioner adoption depends on institutional trust and clarity of responsibility	Kasim et al. (2026), Renigier-Biłozor et al. (2022), Glumac and Des Rosiers (2021)	Emerging (practitioner evidence); Conceptual (framework)	Broad (with EU emphasis)

Finding	Supporting sources	Evidence strength	Jurisdictional depth
Out-of-sample validation and backtesting are prerequisites for defensible AVM deployment	Schulz et al. (2014), Steurer et al. (2021), Krause et al. (2020)	Established	Broad
Prediction intervals, calibration and distributional or conformal methods should accompany every AVM output; in line with supervisory expectations	Rigby and Stasinopoulos (2005), Wheeler et al. (2014), Angelopoulos and Bates (2023), Granna et al. (2025), European Union (2024), Federal Register (2024)	Established (methods); emerging (regulatory codification)	EU, US, UK
Data quality, density and granularity set the upper bound of AVM performance more fundamentally than model complexity	Openshaw (1984), Granna et al. (2022), Krause et al. (2020), Steurer et al. (2021)	Established	Broad
Concept drift and temporal instability materially affect AVM reliability and require systematic monitoring, not ad hoc maintenance	Gama et al. (2014), Gudigantala and Mehrotra (2024)	Established	Primarily US illustration; drift theory jurisdiction-agnostic
Alternative and multimodal data can enrich AVM inputs, but added value is context-dependent and often proprietary	Koch et al. (2019), Despotovic and Brunauer (2024)	Emerging	Broad
ESG signals influence property values (green premiums, brown discounts, stranding risk), but ESG data infrastructure and standardisation lag behind regulatory demand	Eichholtz et al. (2010), Cajias et al. (2019), Hirsch et al. (2019), RICS (2024), European Union (2020, 2022)	Emerging	EU-focused
Synthetic data can support prototyping and stress testing but are not a substitute for real transaction data; risks include model collapse	Goodfellow et al. (2014), Kingma and Welling (2014), Shumailov et al. (2024), Alemohammad et al. (2023), Bidanset (2025)	Emerging	Broad
AVMs can reproduce and amplify historical bias; disparities across racial and socioeconomic groups are empirically documented in US markets	Howell and Korver-Glenn (2021), Freddie Mac (2021), Zhu et al. (2024), Neal et al. (2020), CFPB (2023)	Established (US); Under-researched (elsewhere)	US-focused
Formal fairness criteria cannot be jointly satisfied; mitigation is a governance choice, not a purely technical one	Kleinberg et al. (2017), Barocas et al. (2019)	Established (theoretical)	Broad
Hybrid human AVM architectures are the most resilient operating model under current regulatory expectations; practitioner adoption depends on institutional trust and clarity of responsibility	Kasim et al. (2026), Renigier-Biłozor et al. (2022), Glumac and Des Rosiers (2021)	Emerging (practitioner evidence); Conceptual (framework)	Broad (with EU emphasis)

Note(s): Evidence strength for commercial systems is based on public disclosures and secondary literature; internal validation data are not accessible. “Emerging” reflects findings supported by fewer than three independent studies or restricted to single jurisdictions. “Established” reflects findings replicated across multiple studies and jurisdictions

Source(s): Compiled by the authors on the basis of the literature discussed in Sections 3–7

[ViewLarge]

Sharing Unavailable