📈 economics

How Local Should an AVM Be? Spatial Resolution, Market Regionalization in Automated Valuation Models

This paper reframes the locality of Automated Valuation Models (AVMs) as a partial-pooling problem involving spatial resolution, market regionalization, and external anchors, demonstrating through a comprehensive analysis of Korean commercial property transactions that while an optimal spatial resolution exists and regionalization improves accuracy, these mechanisms are only partially interchangeable and external anchors can significantly alter model selection and performance.

Original authors: Dahoon Kim

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dahoon Kim

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to guess the price of a building just by looking at a map. You could say it is in a specific city, or perhaps in a specific neighborhood within that city, or maybe even down to the exact block. For decades, experts in automated property valuation have operated on a simple assumption: the more specific you are about a location, the better your guess will be. If a computer can see a street address, it should be better than if it only sees a zip code. This logic suggests that spatial detail is a good thing to have as much of as possible. But what if being too specific actually confuses the computer? What if the most accurate way to price a building is not by zooming in as far as the map allows, but by finding a "sweet spot" where the location is detailed enough to matter, but broad enough to have enough data to be reliable?

This is the central puzzle tackled by a new study of commercial real estate in South Korea. The researchers were not just asking how to make a better price estimate; they were asking how much "local" information a computer model should actually be allowed to use. They focused on three different ways to introduce locality into a valuation system. The first is the size of the map grid used to describe a location. The second is whether the model should treat the entire country as one big market or split it into separate models for different cities. The third is whether the model can use an outside hint, like a government-assessed land price from the previous year, to help it understand the local value without having to guess it from scratch. By testing these three factors against thousands of real property sales, the study reveals that there is a limit to how local a model should be, and that pushing past that limit makes the predictions worse, not better.

The researchers analyzed over 22,000 transactions of commercial and office buildings across seven major South Korean cities between 2018 and 2025. They built a computer model to predict prices and then tested it under hundreds of different conditions. They tried representing locations on grids ranging from no specific grid at all, down to 5 kilometers, 2 kilometers, 1 kilometer, and finally a very fine 500-meter grid. They also tested whether the model should be one single system for all seven cities or if it should be split into separate systems for each city, or even smaller groups of cities. Finally, they tested whether giving the model a "hint" in the form of last year's official land price helped it perform.

The results showed a clear and surprising pattern. When the model was allowed to use no sub-city location information at all, its predictions were off by a median of about 27 percent. When the researchers introduced a 5-kilometer grid, the error dropped significantly to about 22.6 percent. This was the best result. However, as they made the grid finer, moving to 2 kilometers, then 1 kilometer, and finally 500 meters, the accuracy got worse again. At the finest 500-meter scale, the error climbed back up to nearly 27.4 percent, which was actually worse than having no grid at all. The study proved that there is an optimal middle ground. The computer needs enough detail to distinguish between different areas, but if the grid is too fine, the specific squares become too empty of data. The model tries to guess the price for a tiny, empty square based on very few examples, and it makes mistakes.

The researchers also found that splitting the market into separate models for different cities helped, but only up to a point. Breaking the single national model into seven city-specific models improved accuracy, but breaking it down further into smaller groups did not help much more. The biggest gains came from simply moving away from a single "one-size-fits-all" model. However, this splitting did not fix the problem of the overly fine grid. Even with seven separate city models, the 500-meter grid still performed poorly. This means that having more specific location data cannot be fixed just by having more specific market definitions; the two issues are related but not interchangeable.

The most powerful tool the researchers tested was the "external anchor." This was a piece of information supplied from outside the model's own learning data: the official land price assessed by the government for the previous year. When the model was given this hint, the entire problem changed. The sharp drop in accuracy that happened when the grid got too fine disappeared. With the government price hint available, the model performed almost equally well whether the grid was 5 kilometers, 2 kilometers, or 1 kilometer. The external information was so strong that it removed the need for the model to guess the local price level from scratch. It also meant the model needed fewer separate city divisions to work well. Instead of needing five or seven different models, the best setup with the external hint only needed three.

The study concludes that the question for building these valuation systems is not simply "how local can we get?" but rather "how much locality can our data support?" The researchers found that for commercial buildings in these cities, a 5-kilometer grid is the most reliable scale for representing location when no external hints are used. Going finer than that introduces too much noise. If a reliable external price signal is available, the choice of grid size matters much less, and the model can be simpler. The findings suggest that in fields where data is sparse, like commercial real estate, trying to be too precise about location can actually hurt performance. The best approach is to find the level of detail that the available data can actually support, rather than assuming that more detail is always better.

Technical Summary: How Local Should an AVM Be? Spatial Resolution, Market Regionalization in Automated Valuation Models

1. Problem Statement

Automated Valuation Models (AVMs) traditionally treat location as a binary inclusion problem: whether to include spatial data and which fixed device (e.g., coordinates, postcodes, administrative boundaries) to use. The implicit design principle in the literature is that finer spatial detail is unambiguously superior. This paper challenges that assumption, arguing that locality is not a single design decision but a trade-off governed by three distinct channels:

Representation Locality ( $h$ ): The spatial resolution at which location is encoded within a single model (e.g., grid size).
Model Locality ( $K$ ): The number of independent regional models ( $K=1$ being complete pooling, $K=7$ being city-specific models) into which the market is partitioned.
External Locality ( $A$ ): Locational information supplied from outside the estimation sample (e.g., lagged assessed land prices).

The core research question is how much locality an AVM should introduce, through which channel, and at what degree of pooling. The paper posits that all three channels share a common partial-pooling logic: localization creates information gains but simultaneously fragments statistical support. The study investigates whether these channels are interchangeable substitutes or if they address distinct sources of error.

2. Methodology and Empirical Design

Data

The study utilizes a dataset of 22,063 arm's-length transactions of non-residential composite properties (commercial and office buildings sold with land) across seven metropolitan cities in South Korea from 2018 to 2025.

Temporal Split:
- Training: 2018–2020 ( $N=9,714$ ).
- Validation: 2021–2022 ( $N=6,770$ ).
- Future Test: 2023–2025 ( $N=5,579$ ).
Features: 150 baseline predictors including building characteristics, land attributes, and time indices.
External Anchor: The lagged officially assessed parcel-level land price ( $A_{i,t-1}$ ), used in anchored specifications.

Model Specification

Learner: XGBoost gradient-boosted regression trees, held fixed across all experimental conditions (same hyperparameters, seed, and preprocessing) to isolate spatial effects.
Experimental Surface: The author estimates the full risk surface across:
- Resolutions ( $h$ ): None, 5 km, 2 km, 1 km, and 500 m grids.
- Partitions ( $K$ ): All 877 possible set partitions of the seven cities.
- Anchoring: Unanchored vs. Anchored (with lagged assessed price).
Selection Protocol: Strictly nested. Configurations are fitted on 2018–2020, selected based only on the 2021–2022 validation sample, refitted on 2018–2022, and evaluated once on the 2023–2025 future sample.
Inference: City-stratified paired spatial block bootstrap (5 km blocks, 500 replications) with Holm adjustment for multiple comparisons to account for spatial dependence in prediction errors.

3. Key Contributions

Contribution 1: Interior Optimum in Spatial Representation

Contrary to the assumption that "finer is better," the paper identifies a clear interior optimum for spatial resolution. Under complete pooling ( $K=1$ ), the future Median Absolute Percentage Error (MdAPE) is minimized at 5 km.

Unanchored Performance: MdAPE moves from 27.104 (None) $\to$ 22.599 (5 km) $\to$ 24.119 (2 km) $\to$ 26.118 (1 km) $\to$ 27.406 (500 m).
Significance: All deviations from the 5 km benchmark are statistically significant after Holm adjustment. The 500 m grid performs worse than having no sub-city encoding at all.
Robustness: This optimum persists even when $h$ and $K$ are jointly selected on validation data, proving the 5 km result is not an artifact of imposed complete pooling.

Contribution 2: Diminishing Returns in Market Regionalization

Market regionalization (increasing $K$ ) improves prediction accuracy over complete pooling at every resolution, but returns diminish rapidly.

While $K=1$ is dominated at all resolutions, the data do not identify an interior optimum for $K$ ; the gains flatten out, with most improvement captured by $K=3$ or $K=4$ .
This creates an asymmetry with Contribution 1: representation locality has a clear peak, while model locality shows a plateau.

Contribution 3: Partial, Not Complete, Substitution

The two endogenous channels (resolution $h$ and partition $K$ ) are partial substitutes.

Increasing $K$ from 1 to 7 recovers 73.7% of the penalty associated with a 2 km grid, 53.7% of the 1 km penalty, but only 19.6% of the 500 m penalty.
Mechanism: Regionalization addresses market heterogeneity (differences in price function shapes across cities) but cannot remedy cell sparsity (lack of data within a specific grid cell). In fact, splitting the market into smaller blocks exacerbates sparsity for fine grids.

Contribution 4: External Anchors as Dominant Substitutes

An external spatial anchor (lagged assessed land price) substitutes for both endogenous channels.

Accuracy: The anchor reduces the realized regionalization gain from ~1.5–3.7 percentage points (unanchored) to ~0.4–1.0 points (anchored).
Architecture: The presence of an anchor changes the optimal design. It reduces the validation-selected $K$ from 5 to 3 and flattens the resolution frontier entirely; under the anchor, no resolution contrast survives multiplicity correction (the 5 km optimum disappears).
Systemic Impact: The anchor significantly attenuates the spatial persistence of correlated prediction errors (co-error), removing 72% of the penalty associated with 500 m grids.

4. Key Results

Resolution Frontier: The 5 km grid is the robust optimum for unanchored models. Finer grids (1 km, 500 m) suffer from high variance due to thin local support (median transactions per 500 m cell $\approx$ 17 vs. 628 for 5 km).
Regionalization Limits: While splitting markets helps, it cannot rescue an excessively fine grid. The compensation ratio drops monotonically as over-localization worsens.
Anchor Substitution: The external anchor acts as a "local price level" that the model would otherwise have to infer from sparse data. When available, it renders the choice of grid resolution and the degree of market partitioning largely irrelevant to performance.
Spatial Error Dependence: Excessive locality (fine grids) increases spatially correlated errors (errors cluster in neighborhoods). The anchor mitigates this systemic risk, whereas regionalization alone does not fully eliminate it.

5. Significance and Claims

The paper claims to resolve the "spatial question" in AVMs by reframing locality as a partial-pooling problem rather than a simple data inclusion issue.

Theoretical Shift: It challenges the GeoAI and valuation literature's tendency to treat location as a fixed input or to assume finer scales are always superior. It demonstrates that the optimal scale is bounded by the local density of training observations.
Design Implications:
- For Markets with Public Assessment: If a reliable external anchor (like assessed land values) is available, AVMs should be simpler: coarser grids and fewer regional models are sufficient, reducing operational complexity and spatial error risk.
- For Markets without Public Assessment: Locality must be "manufactured" internally. In these cases, both grid resolution and market partitioning matter significantly, and a 5 km resolution is likely the practical ceiling for non-residential data in similar densities.
Limitations: The author notes the 5 km optimum is specific to the data density of Korean commercial real estate and may not transfer directly to residential markets (which are denser) or other learner architectures, though the principle of an interior optimum is expected to hold.

The paper concludes that the relevant question for AVM design is not whether to model location, but how much locality the available information can support, through which channel, and at what degree of market pooling.