Counterfactual Explanations for Mortgage Valuation Support: A Research Protocol for Explanation-Aware, Uncertainty-Sensitive Automated Appraisal

1. Introduction

Mortgage lending requires a valuation process that is accurate enough for collateral assessment and transparent enough for audit, dispute resolution, and expert review. Automated valuation models are already used to estimate property prices from transaction and property data, and recent work indicates that machine-learning approaches can improve valuation accuracy relative to many traditional statistical approaches, especially when nonlinear relations and heterogeneous features matter. At the same time, mortgage decisions differ from generic price prediction tasks because they must support an action: whether the estimated collateral value is sufficiently reliable for underwriting, whether manual review is needed, and what property or data factors drive a lower estimate. [1] [2]

A point estimate alone is insufficient for this task. Mortgage practitioners need at least four additional capabilities: local explanation of an estimate, quantified uncertainty around the estimate, a workflow for human review, and a disciplined way to answer counterfactual questions such as: under what plausible and legally acceptable changes would the estimated value move above a required collateral threshold? The present manuscript proposes a research protocol for such a system. It focuses narrowly on mortgage valuation support for residential collateral rather than broader property analytics or urban digital-twin applications.

The central claim is limited and conceptual. Existing evidence supports the plausibility of combining interpretable machine-learning valuation, uncertainty-aware regression, and hybrid human–machine review. What remains underdeveloped for mortgage support is a counterfactual layer that translates model behavior into bounded, auditable recourse statements while respecting valuation practice and feature constraints. This manuscript therefore proposes a method and prospective evaluation design rather than reporting unperformed experiments. [1] [3] [4]

2. Research question and evidence scope

The research question is: how can counterfactual explanations be designed for mortgage valuation support so that they are technically valid, interpretable to appraisers and underwriters, uncertainty-sensitive, and constrained by the institutional realities of collateral assessment? A secondary question is how climate-risk information should enter such explanations without overstating what current valuation evidence can support. [5]

The evidence base used here is a targeted narrative synthesis. It includes direct valuation studies on residential and commercial real-estate appraisal with interpretable or hybrid machine learning, a human–machine collaboration study in real-estate valuation, a study on feature-rich image fusion with SHAP in housing valuation, a hybrid deep-learning house-price study as evidence on high-capacity predictive modeling, an uncertainty-aware ensemble study from construction cost estimation as a transferable methodological template, and a systematic map of climate-risk valuation evidence. This is not presented as a systematic review, and no search counts or exhaustive inclusion claims are made. [1] [3] [6] [4] [2] [5]

The scope of inference is deliberately narrow. Direct evidence from real-estate valuation supports the relevance of nonlinear models, interpretable post hoc analysis, hybrid workflows, and richer feature spaces. Transferable evidence from adjacent regression problems supports the value of calibrated prediction intervals and explanation-aware ensembles. Climate-risk evidence is included because mortgage collateral can be exposed to physical and transition risks, but the literature also reports mixed capitalization and strong dependence on information availability, disclosure, and hazard measurement. Therefore, climate variables are treated as potentially relevant valuation inputs, not as universally priced determinants. [5] [4]

3. What prior work establishes and where the gap remains

Several source studies support the move beyond linear hedonic valuation. In residential valuation, machine-learning models are reported to outperform traditional methods in handling large datasets, nonlinear relations, and feature selection. In the Hong Kong study, tree-based models were emphasized as strong performers, feature configuration materially changed results, and image-derived features contributed substantially to valuation. SHAP was used there to analyze both global and local feature importance and to reveal nonlinear associations between image features and prices. This matters for mortgage support because the valuation explanation must often account for interactions among structural, locational, and environmental variables that are not well represented by simple linear effects. [1]

Commercial appraisal evidence reaches a related conclusion from a different angle. The study of U.S. commercial real estate appraisals argues that structured deviations between appraised values and subsequent transaction prices can be captured by boosting trees, with the practical implication that supervised machine learning may guide valuers toward more dependable estimates. That article is important here not because the asset class is identical to owner-occupied mortgage collateral, but because it frames machine learning as a support mechanism for appraisal practice rather than a replacement for judgment. [2]

Human–machine collaboration evidence also points to a specific design requirement. In the Vienna experiment, a hybrid approach in which experts saw model estimates achieved the highest accuracy relative to the other tested conditions, yet the same study also reports that automated guidance can conflict with expert intuition and degrade performance for some individuals or tasks. The implication is not that any model explanation will help; rather, explanation must be structured to support scrutiny, adjustment, and trust repair after apparent errors. [3]

Evidence on predictive architecture is not uniform. The Hong Kong study found Extra Trees superior among the tested models in its setting, whereas the hybrid deep-learning study reports strong results for an optimized GRU–MLP on a small publicly available dataset and cites other studies where XGBoost performs strongly. This is a boundary condition, not a contradiction to be erased: model performance depends on data scale, feature engineering, tuning, and market context. A mortgage-support protocol should therefore avoid hard-coding one learner as universally best and should instead specify model comparison and calibration procedures. [1] [6]

Finally, work on search-guided regression ensembles in construction cost estimation is methodologically relevant because it combines three ingredients often absent from valuation support: dynamic learner selection, calibrated prediction intervals, and SHAP-based interpretability. The domain differs from property valuation, so no performance transfer is assumed. Still, the design principle is directly transferable: a support system should produce both a point estimate and an uncertainty statement, and the explanation layer should be compatible with ensemble prediction rather than restricted to single models. [4]

4. Conceptual framework for counterfactual mortgage valuation support

The proposed system has four linked outputs for each property: (1) a point estimate of market value, (2) an uncertainty interval, (3) an attribution explanation showing which observed features moved the estimate up or down relative to a reference expectation, and (4) one or more constrained counterfactual scenarios describing the smallest plausible changes in mutable inputs that would move the estimated collateral value across a decision threshold. The system is intended as decision support for mortgage valuation review, not as an autonomous lending decision engine.

Operationally, the target variable is estimated market value at the valuation date, expressed either in currency units or log price depending on model form. For mortgage use, a derived underwriting quantity is the collateral coverage ratio CCR = V_hat / L, where V_hat is the model-estimated market value and L is the proposed loan amount. A hypothetical underwriting trigger may then be defined as CCR >= tau, where tau is a lender-set threshold. Because institutional thresholds vary and are not supplied here, tau is treated as a protocol parameter rather than an asserted industry constant.

The explanatory distinction between immutable, conditionally mutable, and freely revisable variables is central. Immutable variables include coordinates, parcel characteristics, building age at the valuation date, and externally observed hazard-zone designations. Conditionally mutable variables include property condition, documented refurbishment status, energy label, or legal encumbrance status, but only if a specific evidentiary pathway exists to update them. Freely revisable variables are not property traits at all but data-quality fields, such as corrected floor area or rectified geocoding, where counterfactual changes represent data remediation rather than physical intervention. Counterfactual explanations for mortgage use should prioritize this third class, then conditionally mutable characteristics, and should never suggest changing protected or impossible attributes. [5]

In this framework, a counterfactual is not an invitation to manipulate credit decisions. It is an explanation of model sensitivity under constrained, documentable changes relevant to valuation review. For example, if the estimate is below a collateral threshold because the system inferred poor exterior condition from image features and sparse structured data, the support output may state that a documented refurbishment update or corrected condition coding would likely shift the valuation range, subject to re-estimation and uncertainty bounds. This differs from generic recourse literature because the aim is appraisal clarification and review prioritization, not borrower behavior modification. [1] [3]

5. Proposed methodology

The proposed protocol uses a staged model architecture. Stage 1 estimates market value using a candidate set of regressors that may include tree ensembles, gradient boosting, regularized linear regression, and, where data volume and governance allow, a higher-capacity hybrid network. The candidate set is justified by the prior evidence that tree-based models often perform strongly in valuation and that hybrid nonlinear models can be competitive under some data conditions. Stage 2 calibrates uncertainty intervals around the selected predictor or ensemble. Stage 3 computes SHAP values for local explanation. Stage 4 generates constrained counterfactuals relative to a mortgage decision threshold and the uncertainty band. [1] [6] [4]

Input variables should include: structural attributes such as living area, rooms, floor level, age, condition, quality grade, tenure, and amenity indicators; locational variables such as coordinates, distance-based accessibility, neighborhood composition, and local market indicators; image-derived features where governance permits; and climate-risk variables, separated into physical-risk exposure and transition-risk indicators. Physical-risk variables may include official flood-zone class, modeled inundation depth proxy, heat exposure proxy, or coastal proximity where directly relevant. Transition-risk variables may include energy label or policy-relevant compliance indicators where a transition framing is explicit. The climate-risk map cautions that valuation effects are mixed and mediated by salience and disclosure, so these variables should be encoded as contextual features rather than assumed monotonic penalties. [1] [5]

One exact operational rule is proposed for accepting a counterfactual explanation. Let x be the observed feature vector, x' a candidate counterfactual, f(x) the point prediction of market value, and [L_V(x), U_V(x)] a calibrated prediction interval for market value. Let the underwriting quantity be either value itself, Q(x) = f(x), or collateral coverage, Q(x) = f(x) / L, where L is the proposed loan amount and is fixed for the case under review. The decision threshold T must be defined on the same quantity Q used in the underwriting rule. Accordingly, the induced interval for collateral coverage is [L_CCR(x), U_CCR(x)] = [L_V(x) / L, U_V(x) / L]. A candidate counterfactual is admissible only if all three conditions hold: (i) feasibility: x' satisfies all feature-domain and immutability constraints; (ii) sparsity-cost: C(x, x') <= lambda, where C is a weighted cost over mutable features and lambda is a policy-set maximum review burden; and (iii) robust crossing on the decision scale: if underwriting is defined on value, then L_V(x') >= T, whereas if underwriting is defined on collateral coverage, then L_CCR(x') >= T. This rule avoids the threshold-mapping error that would arise from comparing a lower bound in value space with a threshold stated in coverage-ratio space. It is intentionally stricter than requiring only Q(x') >= T, because mortgage support should not treat a threshold as met when only the point estimate crosses while the calibrated lower bound remains below it.

The weighted counterfactual cost function is defined for mixed data types. Let M_c, M_o, and M_n denote respectively the sets of mutable continuous, ordinal, and nominal features. Then C(x, x') = sum over j in M_c of w_j times abs((x_j' - x_j) / s_j) + sum over j in M_o of w_j times abs(r_j(x_j') - r_j(x_j)) + sum over j in M_n of w_j times I(x_j' != x_j), where s_j is a pre-registered scaling constant for continuous feature j, r_j maps ordinal categories to consecutive integer ranks, and I(.) is the indicator function. This makes costs commensurable across variable types. Default scaling should use the interquartile range of each continuous feature estimated from the training split only; if the interquartile range is zero, use the median absolute deviation or set s_j = 1 as a fallback. Data-correction variables should usually receive lower weights than physical-improvement variables. Immutable variables receive infinite weight or are excluded from the mutable sets. For nominal and ordinal variables, admissible transitions must be specified in advance by a transition matrix A_j, where A_j(a,b) = 1 if category a may legally and evidentially change to b and 0 otherwise. Thus, unknown-to-verified condition status may be admissible, whereas arbitrary reassignment of tenure, hazard zone, or protected characteristics is not.

The counterfactual search proceeds in a reproducible sequence. First, fit the valuation model on the training data and construct calibrated intervals by split conformal prediction on a held-out calibration set: train the point predictor on the proper training subset, compute absolute residuals on the calibration subset, choose a target miscoverage level alpha in advance, and set the interval for a new case to [f(x) - q_(1-alpha), f(x) + q_(1-alpha)], where q_(1-alpha) is the empirical conformal quantile of the calibration residuals. If the final predictor is an ensemble, the same conformal calibration is applied to the ensemble point prediction rather than to each base learner separately. Second, compute SHAP values for local attribution and use them only as a search heuristic, not as proof of causal effect. Third, restrict the search space to revisable or mutable features with admissible transitions and nonzero remaining headroom under the cost budget. Fourth, solve for x' by minimizing C(x, x') subject to feasibility and robust crossing; a practical default is best-first search over categorical and ordinal transitions combined with bounded continuous line search on the remaining mutable coordinates. Fifth, present one minimal-cost counterfactual. Sensitivity analysis may additionally report whether the recommendation changes when alpha, lambda, or the feature weights are varied over a pre-registered range. This default rule is preferable to an unconstrained menu because it yields a single auditable acceptance criterion and a single calibration procedure. [4] [1]

This proposal differs from the specific methods in prior work. The Hong Kong image-fusion study uses SHAP for interpretation but does not formulate mortgage-specific counterfactual rules or uncertainty-based robust threshold crossing. The Vienna human–machine study compares experts, model, and hybrid support but does not specify a formal counterfactual engine. The search-guided ensemble paper includes SHAP and prediction intervals but addresses construction costs rather than collateral valuation and does not define mortgage-oriented recourse constraints. The present protocol integrates these ingredients into a valuation-support sequence tailored to underwriting review. [1] [3] [4]

6. Comparator, outcomes, and prospective validation design

Because this manuscript is a protocol rather than an empirical report, the evaluation design is prospective. The primary comparator should be a conventional AVM output without counterfactuals: point estimate plus any standard confidence flag already used by the institution. A second comparator should be AVM plus SHAP explanation without counterfactual support. A third comparator, where feasible, is expert-only review using customary appraisal materials. These comparators allow separate estimation of the added value of attribution explanations and constrained counterfactuals. [3] [1]

Primary quantitative outcomes should include valuation accuracy on held-out transactions, calibration of prediction intervals, and decision consistency for threshold-based mortgage review. Accuracy can be reported with metrics already common in valuation studies, such as MAE, RMSE, MAPE, and proportion of predictions within specified error bands. The latter is particularly relevant because one cited valuation study notes that financial institutions often evaluate AVMs by error-margin coverage rather than by mean squared error alone. Interval quality should be assessed by empirical coverage and interval width. Decision consistency should measure whether equivalent cases receive similar review recommendations under the counterfactual protocol. [6]

Explanation-specific outcomes in this protocol are proposed validation criteria, not findings established by the cited valuation studies. Local fidelity should test whether the sign and rank order of suggested counterfactual changes agree with the model’s behavior in a defined neighborhood around the case. Actionability should be scored by independent reviewers who assess whether each suggested change is documentable, lawful, and meaningful for valuation practice. For hybrid use, expert outcomes should include review time, adjustment frequency, and agreement on whether a suggested counterfactual warrants reinspection, document request, or no action. Because prior human–machine evidence indicates that collaboration effects can vary across experts, subgroup analyses by experience and familiarity with data-driven tools should be specified prospectively rather than inferred from the present conceptual article. [3]

A suitable validation structure is nested and temporally aware. Model selection and hyperparameter tuning should occur only within training folds. Final testing should use later transactions than those used for model development to reflect market drift. If the institution operates across regions, geographic holdouts should be included because valuation models may not transfer well across submarkets. The Hong Kong evidence that feature configuration changes model performance supports explicit sensitivity analysis rather than a single fixed pipeline. [1]

Prospective user evaluation should randomize case review across support conditions: no model support, AVM only, AVM plus SHAP, and AVM plus SHAP plus counterfactual module. Outcome collection should include not only error and time but also the type of review action triggered, frequency of override, and post hoc trust repair behavior when users encounter a clear model mistake. Such a design follows the logic of hybrid experimentation in real-estate valuation while adding the explanation conditions needed for this protocol. [3]

7. Climate-risk integration and boundary conditions

Climate risk matters to mortgage collateral because physical hazards and transition pressures can affect market value, insurance costs, liquidity, and long-run salability. Yet the evidence base summarized in the climate-risk map is fragmented. Many studies find negative capitalization of physical risks, especially hydrological hazards, but others find limited or no effect. The map attributes part of this variation to information availability, salience, and disclosure frictions. Therefore, a mortgage valuation support system should not encode a universal penalty for climate exposure; it should instead include measured exposure variables, record the provenance of those measurements, and expose uncertainty about whether the market currently prices them in the local context. [5]

This has direct consequences for counterfactual design. Hazard-zone classification is usually immutable at the property level and must not be presented as an actionable recourse item. By contrast, some transition-related variables, such as verified energy label or documented mitigation upgrades, may be conditionally mutable if the lender’s use case concerns prospective collateral improvement or post-renovation valuation review. Even then, the system should state that the expected valuation effect is context dependent and not guaranteed, because the climate-risk literature shows uneven evidence across geographies, asset classes, and operationalizations. [5]

A further boundary condition concerns data richness. Image features improved valuation in one direct real-estate study, but their use in mortgage support raises governance issues: image recency, privacy, representativeness, and the possibility that visual proxies entangle unobserved neighborhood characteristics. Explanations based on image-derived SHAP values may be informative for review, yet counterfactuals should avoid requesting cosmetic or ambiguous visual changes unless they correspond to recognized appraisal attributes and documentable property condition states. [1]

8. Failure cases, risks, and limits of generalization

Several failure modes follow from the evidence. First, local explanations can be unstable when input features are highly correlated or when predictions rely on sparse neighborhoods. SHAP values help interpretation, but they do not prove causal effects. Counterfactuals based on such attributions may therefore reflect model contingencies rather than true market mechanisms. Second, a counterfactual that crosses a threshold in model space may still be implausible in market practice if the required feature change is not verifiable through standard appraisal evidence. [1] [4]

Third, human users may over-rely on model outputs or react defensively against them. The hybrid valuation evidence indicates both performance gains and cases where automated guidance degrades outcomes. For this reason, the protocol should enforce explanation display standards that separate factual observation, model inference, and suggested review action. Users should also be able to reject a counterfactual explanation and record the reason, creating an audit trail for subsequent refinement. [3]

Fourth, generalization across markets is limited. Commercial appraisal evidence, residential AVM evidence, and adjacent cost-estimation methods are informative, but they do not imply universal transfer. Asset heterogeneity, transaction sparsity, regulatory rules, and data availability differ sharply across countries and mortgage segments. Even within housing markets, climate-risk capitalization, image usefulness, and best-performing algorithms vary by local conditions. The protocol is therefore a design template requiring local calibration, not a ready-made global model. [2] [1] [5]

9. Conclusion

This manuscript has established a bounded conceptual contribution: a research protocol for mortgage valuation support that integrates ensemble-based value estimation, calibrated uncertainty, SHAP-based explanation, and constrained counterfactual scenarios tailored to underwriting review. The proposal is anchored in direct evidence that machine-learning valuation can capture nonlinear housing-price structure, that SHAP can clarify local and global feature effects, that hybrid expert–machine workflows can improve decisions under some conditions, and that uncertainty-aware ensemble design is a useful transferable principle for decision support. [1] [3] [4] [2]

It has also specified operational elements that can be falsified in future work: the mutable-versus-immutable feature taxonomy, the weighted counterfactual cost function, the robust crossing rule based on prediction-interval lower bounds, comparator conditions, outcome measures, and a prospective validation design. Climate-risk information has been incorporated with explicit caution, reflecting mixed capitalization evidence and the importance of disclosure and salience. [5]

What remains untested is substantial. No empirical claim is made here about predictive superiority, interval calibration, user efficiency, fairness, or acceptance in actual mortgage operations. Those questions require prospective evaluation with local transaction data, institution-specific thresholds, and controlled human-review studies. The value of the present manuscript is therefore methodological: it turns dispersed evidence on interpretable valuation, uncertainty, hybrid workflows, and climate-relevant collateral factors into a coherent protocol for counterfactual mortgage valuation support.

References

1. Deng, Lin (2025). Real estate valuation with multi-source image fusion and enhanced machine learning pipeline. PLOS One. 20(5), e0321951. doi:10.1371/journal.pone.0321951.
2. Deppner, Juergen; von Ahlefeldt-Dehn, Benedict; Beracha, Eli; Schaefers, Wolfgang (2025). Boosting the Accuracy of Commercial Real Estate Appraisals: An Interpretable Machine Learning Approach. The Journal of Real Estate Finance and Economics. 71(2), 314-351. First published online 2023. doi:10.1007/s11146-023-09944-1.
3. Kmen, Christopher; Navratil, Gerhard; Kattenbeck, Markus; Giannopoulos, Ioannis (2026). Evaluating human–machine collaboration through a comparative analysis of experts, machine learning, and hybrid approaches in real estate valuation. Scientific Reports. 16(1), 4044. doi:10.1038/s41598-025-34099-9.
4. Chen, Lifei; Lim, Zhi Min; Lim, Wei Hong; Tiang, Sew Sun; Sharma, Abhishek; Syamsunur, Deprizon; Alharbi, Amal H.; Eid, Marwa E.; et al. (2026). Search-guided regression ensembles for accurate, interpretable, and uncertainty-aware construction cost estimation. Scientific Reports. 16(1), 21456. doi:10.1038/s41598-026-51706-5.
5. Piseddu, Tommaso; Vanhuyse, Fedra (2026). Existing evidence on the impact of climate risk on real estate valuations: a systematic map. Environmental Evidence. 15(1), 8. doi:10.1186/s13750-026-00389-6.
6. Alhammad, Sarah M.; Fouad, Yasser; Mahmoud, Amira A.; Osman, Ahmed M.; El-Bakry, Hazem M.; Elshewey, Ahmed M. (2026). House price prediction using a hybrid GRU–MLP based on binary whale optimization algorithm and ant colony optimization for hyperparameter tuning. Scientific Reports. 16(1), 23225. doi:10.1038/s41598-026-62854-z.