Reader or appraiser: where the language model sits in property valuation

Two research groups spent the past year putting a language model inside a property valuation pipeline, in opposite seats. Serikov and Bogdanova gave the LLM a reading job. It extracts ceiling heights and display glazing from Moscow listing texts, then hands those features to gradient boosting, which owns the price. Geerts, Reusens, Baesens, vanden Broucke and De Weerdt gave it the appraising job. The LLM receives the property attributes and writes the price estimate itself, across four housing markets on three continents. Both teams measured carefully, published their tables, and reached conclusions that lock together from opposite directions. The question was never which model. It is which seat.

The pairing rewards a close read because the two studies fail and succeed in mirror image. The reading seat produces a small, statistically significant gain on top of a strong scaffold. The appraising seat produces a model that rivals classical ML at reading hedonic attributes, then falls apart precisely where the classical model earns most of its accuracy. The mirror is location.

The reading seat, priced to two decimals

The Moscow study is the cleanest ablation I have seen in the valuation literature. Start with 18,416 street-retail listings across Moscow and the region, 50 features. Filter to Moscow with complete records, 14,469 observations. Handle outliers by the interquartile rule, log-transform the target (skewness drops from 0.83 to −0.23), train CatBoost and LightGBM with cross-validated hyperparameters, score on a held-out test set, and run paired significance tests between configurations. Then expand the feature space in three steps, so each data modality gets its own price tag.

The LLM in this design is Sber’s GigaChat, and its job is narrow. Listing texts mention things the structured fields never carry, ceiling height, display glazing, a separate entrance, finishing quality. The model reads each description and emits interpretable named features, llm_has_showcase, llm_ceiling_height. Those features then compete with every other column inside the boosting model. They earn rank 8 and 9 in the SHAP importance ordering, behind the distance-to-centre and coordinate features that dominate the model.

The design also gestures at production. The model emits a decomposition of factor contributions alongside each estimate, which the authors frame as the review surface for management action, revaluation, further review, or correction of the input data. An AVM that shows its factors is one an appraisal committee can argue with.

Three honest limits before anyone generalises. The target is listing prices. An asking-price model inherits whatever negotiation margin separates asking from closing. The study covers one segment of one metro. The transferable result is the method and the effect-size hierarchy, with the coefficients staying in Moscow. And I found no reported validation of the extraction step itself against hand-labelled listings, which means extraction error flows into the features unmeasured. The design would sharpen with a few hundred audited descriptions, cheap work against the value of knowing.

The appraising seat, measured on four markets

The Leuven study, accepted at ECML-PKDD 2025, asks the bolder question. Skip the scaffold. Prompt the LLM with a property’s attributes and ask for the price. The team runs GPT-4o-mini, Llama 3.1:70B and Llama 3.2:3B across King County, Flanders, Barcelona and Beijing, tuning the prompt seriously, zero-shot, few-shot with hedonically similar examples, geographically near examples, mixed selections, market reports. Ten mixed in-context examples emerge as the best general strategy. The example-selection detail is worth an engineer’s minute. Geographically near examples typically beat hedonically similar ones as context, spatial correlation doing its usual quiet work, with Beijing the exception, where property characteristics and broad economic trends outrank neighbourhood. Temporal drift is the other lever. In Beijing a prepended market report drops Llama 3.1:70B from a MAPE of 0.4022 to 0.3322, past the kNN baseline at 0.3810, the largest prompt-driven gain in the paper. The baselines matter, a kNN matched to the same examples the LLM saw, LightGBM stripped of coordinates, and full LightGBM as the state of the art.

The result deserves its precision. LLMs beat kNN broadly and land close to the coordinates-blind LightGBM, which is a genuine finding, a pretrained chat model matching a tuned gradient booster at reading hedonic structure from a handful of examples. On King County the best prompt, a market report plus 10 geographically chosen examples, reaches a MAPE of 0.1390 against full LightGBM’s 0.1378, a match the authors call out. Elsewhere the gap stays real, GPT-4o-mini runs roughly 20% behind the SOTA model in Flanders and Barcelona and 35% behind in King County under the standard prompt.

Then there is Beijing.

The authors state the general form plainly, LLMs prioritise hedonic property features effectively but struggle with spatial and temporal reasoning, undervaluing location and time. Their fourth contribution is that sentence. In a market where spatial autocorrelation carries the pricing structure, the appraising seat is simply the wrong chair.

Both papers point at the same map

Put the two effect sizes side by side and the field’s answer assembles itself.

In Moscow, geospatial features bought 17-20% relative error reduction while the LLM’s text features bought 1.6-3.3%. Location out-earned language by between 8 and 13 to 1 depending on the algorithm, in the study built to make the language features shine. In the Leuven data the same hierarchy appears as a failure mode. The LLM appraiser holds its own wherever hedonic attributes carry the price and loses by a factor of 4 in the city where geography does. Two teams, five markets, three continents, two architectures, one ranking of where the signal lives.

For readers of this blog the shape is familiar. It is the deterministic scaffold and the semantic layer wearing appraisal clothes. It answers the seat question the same way the simulation literature did. The language model plays to its strength inside a structure something else controls. Valuation adds a twist. Here the controlling structure includes geography itself, the one input class the LLM demonstrably cannot internalise from prompts.

There is a second reason to prefer the reading seat. Neither paper states it, so take this as my reading of the pair. In the reader architecture an extraction error becomes one noisy feature among 50, a species of damage gradient boosting tolerates well. In the appraiser architecture the identical error is the price. The seat sets the blast radius of a mistake, which governs unattended operation more than any accuracy gap does.

The intervals nobody should lend against

The Leuven paper runs one further test that any risk function should read before an LLM valuation lands anywhere near a credit decision. Ask each model for a 90% prediction interval around its estimate.

The ranking inside this test deserves separate attention. GPT-4o-mini produces the most accurate point estimates of the three models on King County and Flanders. It also produces the worst coverage anywhere, 35.5% and 25.8% against a 90% target, while the slightly less accurate Llama 3.1:70B reaches 57.5% and 51.4%. Point accuracy and uncertainty calibration are separate properties measured on the same runs. Here they rank the models in opposite orders. A procurement process scoring only the first will select the model that fails hardest at the second.

The authors name the behaviour overconfidence, narrow ranges that read as precision and deliver the opposite. Conformal wrapping fixes calibration for the classical model, 85-90.5% coverage, by widening the bands to fit reality. Nothing comparable rescues the LLM intervals in this study. A valuation model that understates its own uncertainty by that margin is a risk instrument pointed backwards. The finding generalises a pattern this blog keeps meeting, fluent output whose confidence signal carries no information.

One genuinely useful softener from the same paper. The LLM self-explanations, which features drove the estimate, agree substantially with the SHAP rankings of the classical models. The authors still advise caution, and agreement with a real attribution method is more than fluency.

Rosen, 52 years on

The foundational text under both papers is Rosen’s 1974 hedonic price theory, the Journal of Political Economy paper arguing that differentiated goods are bundles of attributes and that markets price the attributes implicitly. Every valuation regression since is a machine for recovering those implicit prices. Both 2025-2026 papers describe their inputs in Rosen’s vocabulary, hedonic variables, without ceremony. The pair also sits inside a visible commercial lineage, interpretable boosting for CRE appraisals reaching the Journal of Real Estate Finance and Economics with Deppner and co-authors in 2025 and ML price indices for commercial property with Calainho, van de Minne and Francke in 2024. The Moscow study is that lineage plus a reader.

Read through that lens, my view of the two papers is that neither is really about language models. The Moscow study widens the attribute set Rosen’s machine can see, pulling ceiling heights out of prose and into the bundle. The Leuven study asks whether a language model can replace the machine. The answer is that it reads bundles well and prices space poorly. The implicit market for location, the attribute Rosen’s successors found dominant in nearly every housing dataset, stays legible only to models that treat coordinates as numbers to learn on. Reasoning about place in words substituted for that in neither study. A 52-year-old frame sorts a 2026 architecture question cleanly.

The bottom line

The operating profiles diverge in a direction the accuracy tables hide. Reading runs once per listing at ingestion, batchable overnight, cacheable, repriced only when the text changes. Appraising runs once per valuation with ten worked examples riding in every prompt, a recurring token cost and a latency budget on each request. Neither paper prices this. In my experience it settles more architecture arguments than accuracy does.

For anyone building valuation systems, the two papers compress to a seating chart. Spend first on location data and a model class that can learn spatial structure, since that is where roughly 90% of the measured error reduction came from in the study that measured both. Put the language model in the reading seat, extracting named, interpretable features from listings, inspections and leases, and validate the extraction against a labelled sample before trusting it, the step Moscow skipped. Let it nowhere near the interval. If a use case demands the LLM produce the price itself, the Leuven prompt work says to select in-context examples by geographic proximity plus attribute similarity, add a market report where trends move fast, and expect parity only where hedonic attributes carry the value and geography stays quiet.

The reviewer’s version is shorter. Ask where the language model sits, then ask who prices the location. A team that answers both has read the same three tables this essay did.

Two studies seated a language model in a valuation pipeline, as reader and as appraiser. The reader added a small, significant gain on top of gradient boosting. The appraiser matched classical models on hedonic attributes and collapsed where location carries the signal, with 90% intervals covering as little as 1.2% of reality. Location out-earned language by close to an order of magnitude. The seat and the map decide what the model is worth.