Detecting Geographic Leakage in Property Price Prediction

1. Introduction

Property price prediction is central to appraisal, taxation, lending, investment, and urban planning. Recent predictive work has moved well beyond linear hedonic models toward machine learning and deep learning systems that ingest heterogeneous variables, including structural attributes, environmental context, and locational information. That shift is reasonable because housing value does not depend on floor area, age, and quality alone. It also depends on neighborhood access, regional market conditions, and the spatial organization of demand. Recent house-price studies explicitly include spatial effects, location variables, and amenity-related predictors, while some models incorporate broader regional or temporal structure such as intercity spillovers and suburb-level price–volume dynamics. These developments improve predictive realism, but they also make evaluation more fragile when train–test separation does not respect geographic dependence. [1] [2] [3] [4]

This article addresses a specific methodological question: how can geographic leakage be detected in property price prediction before reported performance is interpreted as evidence of generalization? In this context, leakage does not refer only to the direct use of future target values or global normalization errors. It refers to a broader problem in which training data contain geographic information that is so proximate, structurally duplicated, or spillover-linked to test data that the evaluation no longer represents deployment to new areas or genuinely independent spatial units. A model can then appear accurate because nearby or connected observations in the training set partially encode the price regime of the test locations. [1] [4]

The proposed contribution is bounded. This manuscript does not report completed experiments and does not claim a new predictive model for house prices. Instead, it develops a conceptual detection framework for leakage risk in geographic price modeling, together with an explicit audit protocol that can be applied to existing forecasting pipelines. The focus is on settings where the target is property sale price, average community price, or suburb-level price index, and where predictors include coordinates, local amenities, mobility summaries, or lagged market variables.

2. Evidence scope and targeted narrative synthesis

The evidentiary scope is a targeted narrative synthesis of sources directly relevant to spatial dependence and locational predictors in housing-price forecasting. The synthesis is not presented as a systematic review and does not claim exhaustive coverage. One source is especially important as a conceptual basis because it models directed intercity housing-price spillovers with transfer-entropy-based network methods and describes a ripple pattern in which core cities transmit price signals to other regions. That evidence is directly relevant to leakage because it implies that market information can cross administrative boundaries in asymmetric and nonlocal ways. A spatial split based only on municipal borders may therefore fail to produce independent train and test partitions when spillovers remain active across the boundary. [4]

A second key source models suburb-level real estate prices and transaction volumes using six lagged values of each variable and emphasizes that price–volume relationships vary across forecast horizons, market regimes, and local suburbs. This matters for leakage detection because it shows that local housing dynamics are heterogeneous rather than exchangeable across space. If neighborhoods or suburbs with closely aligned temporal patterns are split arbitrarily, evaluation can benefit from shared local signatures that are not available when predicting truly distinct areas. The source also treats elasticity-based sensitivities as model-based rather than causal effects, which is a useful caution for any leakage diagnostic based on feature influence. [3]

Additional application evidence comes from community-level and city-level price prediction studies that rely on rich locational covariates. One dataset for Nanning uses residential communities as the unit of analysis, includes longitude and latitude, computes nearest distances and counts to urban amenities within a 1-km radius, and derives land-use proportions within 1-km buffers. This demonstrates how modern housing models encode place through both direct coordinates and engineered neighborhood context. Another study predicts Budapest real estate prices from mobility-derived call detail record summaries aggregated to cells, including entropy, gyration, and home–work distance. These variables are not coordinates in a narrow sense, yet they still embed geographically structured information about urban zones. House-price studies summarized in recent methodological work likewise describe datasets with structural, locational, and amenity attributes, and one such study used a conventional 70:20:10 split while applying normalization using only the training subset. [2] [5] [1]

Taken together, these sources support three premises for the present proposal: first, housing prices are spatially dependent across cities and neighborhoods; second, modern predictors contain dense geographic information even when geography is not represented solely by raw coordinates; third, standard random splits or weakly geographic partitions can therefore exaggerate predictive performance by allowing spatially linked information to cross the training–testing boundary. [4] [3] [2] [5] [1]

3. Conceptualizing geographic leakage

Geographic leakage can be defined as contamination of model evaluation caused by residual spatial dependence or duplicated locational information between training and test sets. The definition has two components. The first is dependence leakage: prices in the training set remain informative about test prices because the two sets are linked by distance decay, commuting structure, amenity catchments, market regime synchrony, or directed spillovers. The second is feature leakage: locational variables or engineered context features identify the test area so precisely that the model effectively interpolates within an already observed geographic manifold rather than extrapolating to a new area.

This concept differs from standard data leakage in two ways. Standard leakage usually concerns information from the target or future period entering model fitting directly, for example by normalizing with statistics from the full dataset. Geographic leakage may occur even when temporal ordering is respected and preprocessing is confined to training data. The issue is not that the test target was seen explicitly, but that the test location was represented implicitly through neighboring, nested, or spillover-connected observations. A model can therefore pass common leakage checks while still failing the more relevant question for many deployment settings: can it predict prices in areas whose geographic context was not already represented in the training sample? [1]

Several forms of geographic leakage are especially plausible in property prediction. First, coordinate interpolation occurs when longitude and latitude or fine-grained geocodes place nearby homes from the same micro-market in different folds. Second, buffer overlap occurs when amenity counts, land-use shares, or nearest distances are computed from overlapping spatial windows, as in community-level analyses using 1-km buffers. Third, network leakage occurs when mobility or spillover graphs connect train and test units through strong flows or transmission pathways. Fourth, temporal-spatial echo occurs when lagged price series from connected suburbs or cities preserve common market states across folds. These forms may coexist and should be audited jointly rather than one feature type at a time. [2] [5] [3] [4]

4. Lessons from prior predictive methods

Prior work illustrates both the value of locational modeling and the reason leakage detection is necessary. The Nanning dataset operationalizes location at fine resolution through longitude, latitude, proximity to amenities, facility counts within 1 km, and land-use shares within 1-km buffers, with residential communities as the basic unit. Such design is appropriate for learning neighborhood price gradients, but it also means that nearby communities may share substantial feature overlap. If those communities are split randomly, the model may be evaluated on locations whose contextual signature is nearly duplicated in training data. [2]

The Budapest mobility study shows a different route to geographic encoding. It aggregates call-detail-record mobility metrics to cells and links normalized real estate price to home-based or work-based movement patterns. Because cell size varies spatially and predictors summarize the behavior of residents or workers anchored to specific cells, these variables embed neighborhood structure even without explicit amenity buffers. A model trained on one set of cells may therefore profit from train–test partitions that preserve adjacent or socially linked cells across the split. [5]

The suburb-level price–volume forecasting study offers a further complication. Each observation uses six lagged price values and six lagged transaction-volume values to forecast future market conditions for a suburb. The study reports that volume-to-price relationships differ across horizons, regimes, and suburbs. For leakage analysis, this implies that random partitions over panel observations could mix highly similar local temporal histories across train and test data. Spatially grouped and temporally blocked evaluation may be required if the intended use is forecasting unseen suburbs or newly entered markets. [3]

Recent general house-price prediction work also highlights a narrower form of leakage control: one study emphasizes leakage-free normalization applied using only the training data in a 70:20:10 split. That is an important preprocessing safeguard, but it does not address geographic leakage. A split can be preprocessing-safe and still geographically permissive. The distinction is crucial because house-price datasets often contain structural, locational, and amenity-related attributes simultaneously, so train–test contamination can persist after standard machine-learning hygiene has been satisfied. [1]

Finally, the intercity spillover literature indicates that geographic leakage is not reducible to Euclidean closeness. Directed transfer of housing-price information can create ripple effects from dominant hubs to downstream markets, and central network positions may matter more than economic scale alone. A city excluded from the training set may still be strongly represented if its upstream spillover sources remain in training data. This observation motivates dependence diagnostics that extend beyond simple distance bands. [4]

5. Proposed methodology for detecting geographic leakage

The proposed methodology is a leakage-audit protocol that can be run alongside any property price prediction model. It is model-agnostic: the predictive learner may be linear regression, random forest, gradient boosting, neural networks, or a time-series forecaster. The audit uses five classes of variables. First, outcome variables: transaction price, log price, price per square meter, community average price, or suburb-level price index. Second, structural predictors: floor area, age, room count, property fees, building type, and related characteristics. Third, explicit locational predictors: longitude, latitude, administrative unit, distance to center, distance to transit, and nearest-facility distances. Fourth, contextual spatial predictors: counts within a spatial buffer, land-use shares, environmental measures, mobility summaries, or market activity summaries. Fifth, dependence variables used only for auditing: spatial distance matrix, adjacency graph, mobility-flow graph, or directed spillover matrix inferred from lagged price interactions. [2] [5] [3] [4]

The core operational idea is to compare an intended evaluation split with a geography-respecting comparator and to quantify residual dependence between training and test partitions. Let U denote the set of geographic units, such as properties, communities, cells, suburbs, or cities. Let A and B denote the training and test subsets for a split, with A ∩ B = empty set. Define w(i,j) as a nonnegative dependence weight from unit i in A to unit j in B. The weight may be based on inverse distance, graph adjacency, mobility flow, or directed spillover strength. The proposed Geographic Leakage Index for a given split is: GLI(A,B) = (1 / |B|) × sum over j in B of max over i in A w(i,j). A high GLI means that many test units retain a strong dependence link to at least one training unit.

An exact decision rule can then be stated. For a chosen dependence definition and a hypothetical threshold tau, flag the split as leakage-prone if GLI(A,B) > tau. Because tau depends on domain and weight scale, it should be prespecified before model evaluation. As an illustrative example only, if w(i,j) is standardized to lie in [0,1], a researcher might set tau hypothetically at 0.30 for strict audits or at a higher value for exploratory work. The threshold is not claimed as empirically validated here; its purpose is to force explicit tolerance choices rather than implicit assumptions of independence.

The audit consists of seven steps. Step 1: define the deployment target. The researcher must state whether the model is intended for interpolation within already sampled neighborhoods, transfer to unseen neighborhoods in the same city, or transfer to unseen cities or regions. Leakage is judged relative to this target. Step 2: document feature provenance. Each predictor is labeled as structural, explicit-location, contextual-spatial, temporal, or derived-from-graph. Step 3: construct dependence weights. At minimum, build a distance-based matrix; where possible, also build a graph from commuting, mobility, transaction interaction, or spillover estimates. Step 4: generate candidate splits, including the original split and at least one geography-respecting comparator. Step 5: compute GLI and subsidiary diagnostics for each split. Step 6: fit the predictive model under identical preprocessing and hyperparameter rules across splits. Step 7: compare performance degradation and error geography between splits. Large performance drops from random to geographic splits, especially when accompanied by high GLI, are interpreted as evidence consistent with geographic leakage.

6. Validation design and comparators

The proposed audit requires a validation design that matches the intended use case. Three comparator families are recommended. The first is random splitting, included not as a preferred design but as a baseline to expose sensitivity to geography. The second is grouped geographic splitting, where all observations from the same spatial unit or contiguous cluster are assigned to one fold. The third is blocked spatiotemporal splitting, where time ordering is respected and test units are also geographically separated from training units. For suburb or city panels with lagged features, the third design is often the most credible because it avoids both temporal and spatial echo. [3] [4]

Outcome measures should include at least one scale-dependent metric and one scale-free metric. Suitable choices are mean absolute error, root mean squared error, mean absolute percentage error, and median absolute error. However, aggregate error is insufficient for leakage detection. The audit should also report: (i) GLI for each split; (ii) nearest-train distance for every test unit; (iii) the proportion of test units whose spatial buffers overlap with any training unit; (iv) error stratified by distance to the training set; and (v) if a graph is available, error stratified by spillover centrality or inbound dependence from training units. If performance is high only for test units near the training set or strongly linked to it, the evaluation may still be geographically contaminated. [1] [2] [4]

For interpretability, feature-influence analysis can supplement leakage detection but should not replace it. The suburb-level forecasting study uses elasticity-based sensitivities as local model diagnostics, and the Nanning dataset discusses SHAP-based interpretation. In a leakage audit, such tools are useful for asking whether coordinate-like or contextual variables dominate predictions under random splits but become less influential under geographic splits. Yet influence values are not proof of leakage on their own. They reveal how the model uses features, not whether the evaluation design is valid. [3] [2]

7. Boundary conditions, conflicting signals, and failure cases

Not every use of locational information constitutes leakage. If the intended deployment setting is valuation within neighborhoods already represented in the training data, then local interpolation may be appropriate, and random or fine-grained cross-validation may answer a legitimate question. In such a setting, coordinates and amenity buffers are part of the target signal, not contamination. Leakage arises when evaluation claims are broader than the split justifies, such as presenting within-neighborhood interpolation accuracy as if it established transfer to unseen suburbs or cities.

A second boundary condition concerns scale. Intercity spillover evidence shows that markets can be connected over long distances through directional transmission, but this does not imply that every city pair is strongly dependent. Conversely, neighborhood-level studies show intense local overlap, but local dependence need not dominate at broader regional aggregation. The appropriate audit therefore depends on the unit of analysis. Property-level splits require attention to coordinate duplication and buffer overlap; community-level splits require neighborhood-cluster blocking; suburb- and city-level panels require graph- or spillover-aware separation. [2] [3] [4]

Conflicting signals may also appear. A random split can yield modest accuracy improvements over a geographic split even when GLI is low, perhaps because the geographic split is simply harder due to covariate shift rather than leakage. The converse is also possible: GLI may be moderate while performance remains stable, suggesting either robust generalization or an audit weight that poorly approximates dependence. For this reason, the protocol uses triangulation rather than a single indicator. Leakage is inferred from the joint pattern of residual dependence, split sensitivity, and spatial concentration of errors.

Several failure cases deserve explicit mention. First, if geographic coordinates are noisy or privacy-jittered, distance-based GLI can understate dependence. Second, if mobility or spillover graphs are unavailable, the audit may miss nonlocal connections. Third, if sample size is small, aggressive geographic blocking can produce unstable estimates and make the model appear worse for reasons unrelated to leakage. Fourth, in markets with rapid regime change, even geographically valid splits may perform poorly because the underlying process shifts over time. This concern is consistent with evidence that price–volume relationships vary across horizons and regimes. A failed geographic split is therefore not itself proof of leakage in the original split; it is a warning that claims of transferability require more care. [3]

8. Prospective evaluation protocol

A prospective study implementing the proposed framework should begin by selecting one or more datasets that differ in geographic resolution, such as property listings with coordinates, residential communities with amenity buffers, and suburb-level or city-level panels with temporal dynamics. For each dataset, the protocol would define the intended deployment target and then create three split designs: random, grouped geographic, and blocked spatiotemporal when applicable. The same predictive model or set of models would be trained under identical preprocessing rules, with leakage-free scaling or encoding based only on training data. The only planned difference across runs would be the split structure and, if necessary, the unit used for grouping. [1] [2] [3]

The primary hypothesis would be falsifiable: if evaluation performance drops materially when moving from random to geography-respecting splits and if the original random split has high GLI or extensive buffer/graph overlap, then the original estimate was likely inflated by geographic leakage. A secondary hypothesis would examine feature dependence: models that rely heavily on explicit coordinates or local context variables should be more sensitive to geographic splitting than models using only structural attributes. A third hypothesis would test whether directed dependence diagnostics based on spillover or mobility graphs provide information beyond Euclidean distance alone. [5] [4]

Planned analyses should include descriptive maps of splits, distributions of nearest-train distance, GLI by fold, aggregate errors, and error gradients over space. Where time series are available, results should be reported separately by forecast horizon because prior work indicates that local interactions can weaken at longer horizons. Where market-state labels are available, one may also inspect whether leakage risk differs across low-, normal-, and high-price regimes, echoing regime-aware forecasting analysis. Any thresholds used for GLI or overlap rates should be prespecified and treated as hypotheses subject to revision after empirical testing rather than as settled constants. [3]

9. Conclusion

This article has developed a conceptual framework for detecting geographic leakage in property price prediction. The argument rests on targeted evidence that housing markets exhibit spatial dependence across cities, heterogeneous local dynamics across suburbs, and strong predictive contributions from coordinates, accessibility, amenity context, and mobility structure. These conditions make property prediction especially vulnerable to train–test contamination when evaluation splits ignore geographic dependence. The manuscript therefore proposed a model-agnostic leakage-audit workflow centered on deployment-target definition, feature provenance review, dependence-weight construction, geography-respecting comparators, and a formal Geographic Leakage Index. [4] [3] [2] [5]

What has been established here is conceptual and methodological rather than empirical. The manuscript specifies operational definitions, a decision sequence, comparators, outcomes, assumptions, and major failure modes. It also distinguishes geographic leakage from ordinary preprocessing leakage and clarifies that locational features are not inherently problematic; the problem arises when evaluation claims exceed the independence supported by the split. What remains untested is the empirical calibration of dependence weights and thresholds, the relative usefulness of distance versus graph-based diagnostics, and the magnitude of leakage inflation across different housing datasets and forecast tasks. Those questions require prospective implementation of the proposed protocol under clearly stated deployment scenarios.

References

1. 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.
2. Lin, Wenbo; Wu, Kaichuang; Zhang, Xiaolu; Gu, Xiaoling; Jiang, Ming; Wang, Jun; Han, Feng (2026). A comprehensive dataset for uncovering housing price drivers in emerging cities: Nanning as a case study. Data in Brief. 66, 112884. doi:10.1016/j.dib.2026.112884.
3. Le, Dat; Rajasegarar, Sutharshan; Luo, Wei; Nguyen, Thanh Thi; Angelova, Maia (2026). From prediction to explanation for price–volume dynamics in real estate market. Interface Focus. 16(3), 20250069. doi:10.1098/rsfs.2025.0069.
4. Qiu, Lu; Jiao, Yanzhe; Dong, Gege; Cui, Guangcan (2026). Finite-Length Spatiotemporal Modelling for Housing Price Network Spillovers. Entropy. 28(5), 537. doi:10.3390/e28050537.
5. Pinter, Gergo; Mosavi, Amir; Felde, Imre (2020). Artificial Intelligence for Modeling Real Estate Price Using Call Detail Records and Hybrid Machine Learning Approach. Entropy. 22(12), 1421. doi:10.3390/e22121421.