1record_id,eligible,short_ref,country,city,property_type,price_type,n_obs,sample_period,data_source,methods_compared,best_method,metric,best_value,hedonic_value,test_set,features_ai,incremental_ai_gain,incremental_ai_metric,marginal_effects,notes,confidence205c8066a4f1ebf0565508520c4002c90671161c4,True,Tarasov & Dessoulavy-Sliwinski (2025),Poland,Warsaw,residential apartments (secondary market),transaction price per m2 (PLN/m2),10827.0,2021,"Register of Property Prices (RCN), Warsaw City Office",random forest;XGBoost,XGBoost,MAPE,8.99,,k-fold,none,,,SHAP,"Table 2, 5-fold CV validation results: XGBoost MAE 1,009.91 PLN/m2, RMSE 1,308.49 PLN/m2, MAPE 0.0899 (=8.99%); RF MAE 1,030.99, RMSE 1,328.74, MAPE 0.0918. No OLS/hedonic benchmark estimated. SHAP waterfall/violin/heatmap plus SHAP partial-dependence plots of usable area, building age, etc.",high30726a6cc581fee39cc268e4d32d8292b296bebc9,True,"Dewi, Ulama & Susilaningrum (2024)",Indonesia,Surabaya,single-family residential houses,"asking/selling price from property advertisements, re-verified with sellers (million IDR)",548.0,July-September 2023 (properties transacted or offered in the past year),"property advertisement websites, re-verified with sellers (purposive sample, >=5 houses per sub-district)",OLS hedonic (multiple regression analysis);artificial neural network (backpropagation),artificial neural network,MAPE,31.04,34.24,holdout,none,,,coefficients,"Table 4, 20% holdout test data (80:20 split, same partition for both models): ANN MAE 2.98 x 10^2, RMSE 3.70 x 10^2, MAPE 31.04%; MRA MAE 3.28 x 10^2, RMSE 3.96 x 10^2, MAPE 34.24% (MAE/RMSE presumably in million IDR; mean price 1.935 billion IDR). Hedonic regression coefficients reported in Table 3 (adj. R2 0.91 on training data). Authors note ANN errors remain high, model only to support appraisers.",medium40795e1f2569ab3c9b87c0451960aa54739093319,True,Amri & Tularam (2012),Australia,"Bathurst, NSW",houses (residential sales),sale price,7849.0,1992-2004,Council/cadastre sales records GIS-matched with land value and ABS census social data,stepwise multiple regression (MR);MR with social indicators (MRsoc);neural network (MLP);ANFIS neuro-fuzzy,neural network,R2,0.4536,0.3694,k-fold,none,,,coefficients,"Table 2 (10-fold CV): adjusted R2 NN 0.4536, ANFIS 0.3815, MR 0.3694; MR with social data 0.42 (text). Share of predictions within 10% of actual: NN 31%, ANFIS 28%, MR 28%. Dataset lacks structural attributes (land value + location + social variables only), which the authors cite as the reason for low R2. n_obs = 7,849 GIS-matched unique sales (counts in Table 2 imply ~4,200 used in the comparison). MR coefficients reported (Eq. 2-3).",medium50c59335ca1b171d1a7095c8fce95805f7852118d,True,Cui et al. (2024),China,"Tai'an (central district), Shandong","pre-owned (second-hand) housing, community-level average","asking price per m2 (RMB/m2, Anjuke listings, community average)",1177.0,2023,Anjuke.com listings (web crawler) + Baidu Street View imagery + Amap POI,OLS hedonic price model (HPM);GWR;MGWR,MGWR,R2,0.613,0.346,none,street-view CNN,,,coefficients,"Table 1 (in-sample fit, 1,177 communities, macro-scale model): HPM R2 0.346 (adj 0.334), GWR 0.582 (0.521), MGWR 0.613 (0.560); mountain-view subsample (475 communities): HPM 0.419 (0.392), GWR 0.893 (0.841), MGWR 0.911 (0.881). AI features = street-view element indices from semantic segmentation (building/flat/mountain view indices) plus a CNN mountain-view satisfaction classifier (66.7% accuracy on an 80/20 image split) and Grad-CAM; the price models themselves are not out-of-sample tested and no with/without-AI-feature comparison is printed. Local MGWR coefficients for street-view variables are mapped (Figure 9).",medium61076844f99e9f35c7768a4e64f2c167ab26b5afd,True,Ezzeldin et al. (2025),Saudi Arabia,Riyadh,residential units (apartments and villas),listing price (log),2997.0,Q3 2024,"Aqar listing platform (sa.aqar.fm), scraped and georeferenced","semi-log OLS hedonic regression;artificial neural network (MLP, 5-fold CV tuning);random forest (500 trees)",random forest,R2,0.9065,0.8034,holdout,none,,,coefficients,"70/30 split (2,097 train / 900 test), same split for all three models. Test-set metrics as printed in the abstract/Tables 3, 5, 6: RF R2 0.9065, RMSE 0.1911, MAE 0.1353; ANN R2 0.8968, RMSE 0.2008, MAE 0.1450; hedonic R2 0.8034, RMSE 0.2673, MAE 0.1970 (dependent variable is log price, so RMSE/MAE are in log units). Hedonic coefficients reported (Table 2) plus RF impurity-based feature importance; no SHAP/PDP computed for the ML models.",high71462a807ace2b4c97b7b8368ffbe3a1c694cd0f7,True,Solovev & Pröllochs (2021),Germany,Berlin,apartments (rental listings with floor plans),"asking rent (rent per m2 and total rent, log)",9174.0,mid-2017 to end-2019 (crawled end of 2019),ImmobilienScout24.de listings (web crawler),log-linear OLS hedonic;log-linear OLS + floor-plan CNN score;CatBoost;CatBoost + floor-plan CNN score;deep neural network;deep neural network + floor-plan CNN score,deep neural network + floor-plan CNN score,MAE,0.0205,0.1565,k-fold,floor-plan,-10.56,"MSE (% change, DNN with vs without floor plans, rent per m2)",coefficients,"Table 2 (rent per m2, 5-fold CV out-of-sample): benchmark log-linear OLS MSE 0.04984 / MAE 0.1565; OLS + FloorPlanSentiment MSE 0.04400 / MAE 0.1347 (MSE -11.71%, MAE -13.93%); CatBoost 0.00140/0.0262 -> 0.00134/0.0258; DNN 0.00142/0.0214 -> 0.00127/0.0205 (MSE -10.56%, MAE -4.21%). Units are log rent per m2. Floor-plan score comes from a two-stage CNN (Xception) trained on historical prices; explanatory OLS regressions report significant FloorPlanSentiment coefficients. Table 3 gives analogous results for total rent.",high823c746a3cf2275575136860a11ab1116e38d2228,True,Nor & Hussein (2025),Somalia,Baidoa,residential properties,"residential property price (USD, survey-reported)",118.0,,"Baidoa Housing Survey (primary survey of agents, owners, housing authorities; stratified random sample)","OLS hedonic regression;artificial neural network (MLP, 5-fold CV)",artificial neural network,MAE,1200.0,,holdout,none,,,coefficients,"Table 5/7: ANN test MAE $1,200 and test MSE 2,500,000 (USD); OLS MAE $1,500 reported in Table 7 but without stating whether it is on the same test set (OLS R2 = 0.742, adj. 0.728, in-sample), so hedonic_value is left null; the paper claims a ~20% MAE/MSE reduction for the ANN. Train/validation/test split with 5-fold CV mentioned; sizes not printed. Hedonic coefficients are reported; no SHAP/PDP for the ANN. Very small sample (118).",low92580ac3af513b0ee5f2e19ed10bd991f2a8c8015,True,Takizawa (2023),Japan,Osaka Prefecture,family-oriented rental apartments (3LDK-type units),"monthly rent (excluding common expenses, Yen)",15323.0,"2015 (LIFULL HOME'S snapshot, September 2015)",LIFULL HOME'S dataset (listings + floor-plan images),"OLS hedonic (baseline, general attributes);OLS hedonic + graph features (GF model);OLS hedonic + GCN floor-plan value (FPV model)",OLS hedonic + GCN-derived floor plan value (FPV model),RMSE,11280.0,,k-fold,floor-plan,,,coefficients,"10-fold 'semi-cross-validation' on test folds: FPV model mean RMSE 11,280 Yen and mean adjusted R2 0.79, reported as significantly better (Bonferroni) than the baseline hedonic model and the graph-feature model; baseline model RMSE/R2 shown only in Figure 11 boxplots, not printed as numbers, so hedonic_value and the incremental gain are left null. Floor-plan value is a GCN embedding of the access graph extracted from floor-plan images and entered as one regressor. Regression coefficients of the FPV model (Tables 5-6) and integrated-gradient analysis of the GCN are reported.",medium102b0979784a826ff5181dbb43492b0239ca3062e8,True,"Gillespie, Lyons & Hynes (2023)",Ireland,national (coastal properties within 5 km of the coast),mixed (houses and apartments),listing price (log); transaction-price subsample robustness,100030.0,2014-2019,"daft.ie listings; Property Price Register for ~50,000 matched transactions; 10 m DEM viewsheds",OLS hedonic with spatial fixed effects (6 levels of SFE);OLS with GIS-only view measures;OLS with NLP-only sea-view dummy,OLS hedonic with Small Area + micro-market x year fixed effects,R2,0.89,0.89,none,listing text (rule-based NLP sea-view detection),0.0,R2 (in-sample),coefficients,"Hedonic study: rule-based NLP (NLTK pattern matching on ad text) flags sea views, GIS 3D viewsheds measure breadth/depth. Table 3 preferred specification R2 0.890, RMSE 0.222 (log price, in-sample). Table 4 separates GIS and NLP measures: R2 stays 0.890 and RMSE 0.222 in all six specifications, so the printed fit gain from the NLP feature is 0.000; its contribution is in coefficient identification (sea-view premium 8.2%, up to 15% for the top breadth quartile).",high112cdb6f374417590715ab5db24f593d17616b9ea2,True,"Han, Woo & Lee (2025)",South Korea,Seoul,single-family houses,transaction price (log),13776.0,2017-2019,"Value Up Systems transaction platform (MOU); Google Street View panoramas (233,106) segmented with DeepLab V3+",linear regression (hedonic);random forest;XGBoost;gradient boosting,gradient boosting,R2,0.817,0.66,holdout,street-view CNN,,,SHAP,"Table 2, 20% holdout (2,755 test obs): R2 0.660 (linear regression), 0.318 (RF), 0.765 (XGBoost), 0.817 (gradient boost); MSE 0.056 and MAE 0.169 for GB in log-price units. All models include the street-view segmentation shares (greenery, sky, buildings, etc.); no model without them is reported, so no incremental gain. SHAP global/local plots plus PDP for key variables.",high12321664826269788c8435ead0315a2539fb9837f3,True,"Sa'at, Adi Maimun & Idris (2021)",Malaysia,Johor Bahru,"houses (landed residential, four mukims)",transaction price (RM),3732.0,2009-2018,Valuation and Property Services Department (JPPH) Johor Bahru sales transactions,"hedonic price model (OLS);ANN (feed-forward backpropagation, SPSS)","ANN (one hidden layer, sigmoid activation)",RMSE,0.0047,0.049,unknown,none,,,none,"Table 3: HPM MSE 0.0024 / RMSE 0.0490 vs ANN MSE 0.0020 / RMSE 0.0047 (units not stated; apparently a normalized/log scale, and MSE-RMSE pairs are internally inconsistent). Data split 60/30/10 train/test/validation for the ANN but it is not stated on which subset Table 3 is computed. Table 4 compares forecasts for ten 2018 transactions (errors in %), with 2009-2017 used for fitting. No coefficients or importance reported for either model in the text available. Low confidence given reporting quality.",low133284b13d3d2589a5886f884f24420e24a883d1c3,True,Geerts et al. (2025),Multi: Usa,King County (USA); Flanders (Belgium); Barcelona (Spain); Beijing (China),residential (transactions; Barcelona = listings),sale price (transaction; listing price for Barcelona),20832.0,King County 2014-05 to 2015-05; Flanders 2015-01 to 2023-05; Barcelona 2018-03 to 2018-12; Beijing 2015-01 to 2017-12,Kaggle kc-house-data (King County); proprietary Flanders transactions; Barcelona listings (Lopez-Vargas); Kaggle Lianjia Beijing,"LLM in-context learning (Llama 3.2:3B, Llama 3.1:70B, GPT-4o-mini; zero-shot, few-shot, market-report, hybrid prompts);kNN;LightGBM without coordinates;LightGBM",LightGBM (with coordinates),MAPE,0.1378,,holdout,none,,,SHAP,"Table 2 (MAPE printed as a fraction, 1000 random test examples per dataset from a 60:20:20 split): LGBM 0.1378 King County / 0.2625 Flanders / 0.1556 Barcelona / 0.1056 Beijing; best LLM (GPT-4o-mini, 10 ex. mixed prompt) 0.1861 / 0.3170 / 0.1842 / 0.4125; kNN 0.2105 / 0.3207 / 0.2638 / 0.3810. n_obs = King County full dataset (12914+4349+3569); other datasets larger (Flanders 288448, Barcelona 61348, Beijing 195695). No OLS/hedonic benchmark (explicitly excluded). SHAP top-5 features of LGBM compared with LLM self-explanations; conformal prediction intervals also reported. best_value is the King County column; multi-market study, metrics not pooled.",medium1433b79cf36579ae3f85aa1db50b2d2c31cb4ba23a,True,"Embaye, Zereyesus & Chen (2021)",Uganda,,dwellings of rented households in household surveys,"rent (monthly rental value, USD at 2012 prices)",4882.0,Uganda 2010 & 2012; Tanzania 2014 & 2016; Malawi 2014 & 2016,World Bank LSMS-ISA household surveys,OLS hedonic;OLS with spatial lag (SAR);OLS with spatial error (SEM);ridge;LASSO;regression tree;bagging;random forest;boosting,boosting,other,0.85,1.0,k-fold,none,,,coefficients,"Metric = out-of-sample MSE standardized so that OLS = 1.00 (Tables 9-11). Boosting beats OLS in all six country-years: ratios 0.87/0.88/0.96/0.95/0.86/0.91 with SAR (Table 9) and 0.88/0.85/0.96/0.96/0.87/0.91 without spatial terms (Table 11); best_value = 0.85 is the lowest printed Boosting ratio (Uganda 2012, Table 11), the range is 0.85-0.96. Random forest 0.80-1.00, bagging 0.78-1.28, tree worst (up to 1.60). n_obs = sum of the six samples (242+263+873+1,363+662+1,479). Out-of-sample predictions rely on K-fold cross-validation (fold number not stated), coded k-fold. OLS/LASSO/ridge coefficients reported (Tables 2-4) plus tree-ensemble variable importance.",medium153d555b719dcd66f5b773ee480ecb47ce9229f3fe,True,"Štubňová, Urbaníková, Hudáková & Papcunová (2020)",Slovakia,Nitra,apartments,asking price (online listings),256.0,2019-09 (listings collected 9-13 Sept 2019),TopReality.sk online real estate portal,OLS hedonic (stepwise backward);ANN (Levenberg-Marquardt);ANN (Bayesian regularization);ANN (scaled conjugate gradient),"ANN (Bayesian regularization, 2 hidden layers 10/20 neurons)",MAPE,3.39,8.41,holdout,none,,,coefficients,"60 ANNs (Table 4) vs stepwise regression (Table 6). Best BR-ANN: R2 0.9749, MAE 3561.288, MAPE 3.39%; regression: R2 0.7898, RMSE 11454.91 EUR, MAE 8516.60, MAPE 8.41%. Data split 85:15 train/test (70:15:15 for LM/SCG ANNs); the text does not state explicitly whether the printed indices are computed on the test set alone, hence medium confidence. Regression coefficients reported; no ANN interpretation.",medium1641c05e4c190bfd812f576bb4bc53a0c147331a7d,True,Deppner & Cajias (2024),Germany,Frankfurt am Main,apartments (rental listings),"asking rent (monthly, log-transformed)",9256.0,2019-01 to 2020-03,German multiple listing systems (MLS) apartment rental listings,OLS hedonic;SDM (spatial Durbin);SDEM (spatial Durbin error);random forest;XGBoost,XGBoost,R2,0.8521,0.7881,holdout,none,,,coefficients,"Table 4 Panel A (models with spatial controls): true out-of-sample holdout = Q1 2020 listings predicted from models trained on 2019 data. R2 printed as 85.21% for XGB (spatial tuning; 85.20% non-spatial tuning), 85.13% RFR, 81.62% SDM, 78.81% OLS; MAPE 12.83% XGB vs 15.03% OLS; RMSE 244.06 vs 292.14 EUR/month. Paper's main point is that non-spatial k-fold CV errors of RF/XGB are over-optimistic (CV R2 ~89-91%) vs spatial CV; only OLS/SDM/SDEM coefficients are reported for interpretation, no SHAP/PDP for the ML models.",high174c31e259cf5625db8bd0cbbcae78c68fed2c8164,True,Changro (2026),South Korea,Seongdong-gu (Seoul) and Gangjin-gun (Jeonnam),"benchmark (standard) houses, mostly detached/rowhouse",appraised benchmark price (MOLIT 2024 survey),1670.0,2024,MOLIT benchmark house survey (data.go.kr),OLS hedonic without text;OLS hedonic with text embeddings;XGBoost with text embeddings,XGBoost with text embeddings,R2,,,holdout,listing text embeddings,0.05,"adjusted R2 (OLS, Seongdong-gu 0.508 -> 0.558; Gangjin-gun 0.633 -> 0.751, i.e. +0.118)",SHAP,"80/20 random split; 763 houses in Seongdong-gu, 907 in Gangjin-gun, modelled separately. Only printed metric is adjusted R2 of the OLS models (Table 4, apparently full-sample fit): 0.508/0.558 (Seongdong) and 0.633/0.751 (Gangjin) without/with two PCA components of OpenAI sentence embeddings of the environmental description. XGB 'best fit' is shown only graphically (Figure 3, test set), no numeric metric printed, so best_value is null. SHAP importance and SHAP dependence plot for embedding component 1; OLS coefficients also reported.",medium1853d1b6d48e64b056e88ebabd287c3f06ec3b559d,True,"Tapia, Chavez-Garzon, Pezoa, Suarez-Aldunate & Pilleux (2025)",Chile,Santiago Metropolitan Region,houses and apartments (for-sale listings),"asking price (UF, ln(price) as target)",52039.0,2020-08 (listings collected August 2020),"Coderhub SpA scrape of three Chilean real-estate websites (98,630 raw listings, 52,039 after cleaning; listings with at least one photo)",LightGBM (sequence of feature sets);OLS hedonic;SAR (spatial lag),LightGBM (34 predictors incl. image-derived variables),MAPE,10.07,21.13,k-fold,listing-photo CNN (interior/exterior feature detection),-0.17,"MAPE percentage points (Model 6 -> Final Model with 10 image variables, 5-fold CV; paper reports relative decreases of 1.7% MAPE, 1.2% MAE, 1.3% RMSE)",SHAP,"Table 3 (5-fold CV): Final LightGBM MAPE 10.07%, MdAPE 6.84%, MAE 807.53 UF, RMSE 1373.23 UF, R2 0.9516; base model with 7 variables MAPE 11.59%, R2 0.9412. Table 9 (same 5-fold CV): OLS MAPE 21.13%, R2 0.7103; SAR 20.92%, R2 0.7405. Table 5: final model on an 80/20 holdout MAPE 10.04%, R2 0.9525. features_ai is outside the protocol list: 10 image variables flag which photo types (bathroom, bedroom, living room, balcony, floor plan, pool...) are present, plus five structural variables parsed from the description text. SHAP beeswarm, waterfall and dependence plots; OLS/SAR coefficients also reported (Table 8).",high195bcd886348a12f52aa055354a9647ba5c4fa64c4,True,"Almaas, Oust & Pollestad (2026)",Norway,Oslo,condominiums,transaction price,15702.0,January 2022 - December 2023,"Boligmappa AS transactions matched with 393,914 listing images; crowdsourced condition labels",hedonic linear regression (HLR);XGBoost;each without condition / with computer-vision condition / with human-labelled condition,XGBoost with human-labelled condition (XGBoost with computer-vision condition: MAPE 7.71%),MAPE,7.49,12.41,holdout,interior-photo CNN,-0.32,"MAPE (percentage points, XGBoost)",SHAP,"AI feature = CNN classification of listing interior photos (room type + 1-5 condition). Table 4, stratified 70/15/15 split, test set: HLR MAPE 12.75% (no condition) -> 12.42% (CV condition) -> 12.41% (human), R2 0.672 -> 0.687 -> 0.690; XGBoost MAPE 8.03% -> 7.71% -> 7.49%, R2 0.847 -> 0.855 -> 0.863. incremental_ai_gain recorded for XGBoost (8.03 -> 7.71, -0.32 pp); for HLR -0.33 pp. Coefficients of condition in HLR (Tables 5-7) and SHAP dependence plots for XGBoost.",high205cebf4171ab33d78f1774e47b906013d8103d51a,True,"Upadhye, Bhuravane & Khanche (2026)",United States,"Ames, Iowa",single-family residential (Ames Housing dataset),sale price (USD),,,Ames Housing dataset (Kaggle) plus property text descriptions said to be derived from listing metadata and agent narratives,"linear regression;random forest;XGBoost (structured features only);SmartPrice (XGBoost + 1-D CNN text embeddings, feature-level fusion)",XGBoost + CNN text embeddings (SmartPrice),R2,0.92,0.71,holdout,listing text embeddings,0.03,"R2 (0.89 structured-only XGBoost to 0.92 with CNN text embeddings; RMSE 27,880 to 24,930 USD; MAPE 11.6% to 9.8%)",SHAP,"Table I (test set, 70/15/15 stratified split, 5-fold CV on training set): Linear Regression RMSE 43,920 / MAE 29,210 / R2 0.71 / MAPE 18.7%; RF 31,550 / 20,120 / 0.84 / 13.1%; XGBoost 27,880 / 17,340 / 0.89 / 11.6%; SmartPrice 24,930 / 15,980 / 0.92 / 9.8%. The 'linear regression' baseline is a generic regression, not a hedonic specification with reported coefficients. Caveats: the public Ames dataset contains no free-text descriptions and the paper does not say how its texts were obtained; n_obs and sample period are not printed; only 2 of 5 CV folds are shown; text has tutorial-like wording. Global and local SHAP values reported (feature names anonymised in the plot).",low217b29c36e620db272e042f7b9f343c676a2038678,True,"Lorenz, Willwersch, Cajias & Fuerst (2022)",Germany,Frankfurt am Main,apartments (rental listings),asking rent,52966.0,2013-2019,Empirica Systeme residential rental listings,OLS hedonic;SAR;GAM;XGBoost,XGBoost,R2,0.925,,k-fold,none,,,PDP,"XGB tuned with 5-fold, 5-repetition random CV; 'out of sample rental prediction with XGB yields a R2 of 92.50%', MAPE 11.13%. OLS adjusted R2 0.880, SAR R2 0.885, GAM 0.898 are in-sample fits (Table 2), not on the same test set, so hedonic_value left null (paper itself claims an increase of 'up to 4.50 percentage points'). Interpretability via permutation feature importance, PD plots, ALE plots and interaction (H-statistic) analysis; OLS/SAR/GAM coefficients also reported.",high2284fe25f12d7833b71a83f2c2ae3a06dfc013ffcf,True,Fu et al. (2019),China,Beijing and Shanghai,"residential plots (apartment estates), plot-level average","log selling price (10,000 RMB, FANG.com purchasing records, plot-level)",,2015,"FANG.com house purchasing records + Baidu Street View panoramas (71,332 Beijing / 91,759 Shanghai) + Baidu POI + AutoNavi roads",robust OLS double-log hedonic without view indices;robust OLS hedonic with GVI;robust OLS hedonic with GVI+BVI;robust OLS hedonic with GVI+BVI+SVI,robust OLS hedonic with all three street-view indices,R2,0.7242,0.7242,none,street-view CNN,0.0421,"R2 (Beijing: +0.0421 from adding GVI, BVI, SVI to the baseline hedonic; Shanghai: +0.0297; stepwise contributions GVI 0.0214/0.0184, BVI 0.0032/0.0025, SVI 0.0175/0.0088)",coefficients,"Table 2 full model adjusted R2 = 0.7242 (Beijing) and 0.7511 (Shanghai), in-sample robust OLS; best_value reports Beijing. AI features are PSPNet semantic-segmentation view indices (green, building, sky) aggregated over 800-m road buffers around each plot. Number of residential plots is not printed in the main text (data on figshare). Implicit prices: 1% GVI increase -> +0.3151% price (RMB 39,377) in Beijing, +0.1853% (RMB 21,689) in Shanghai; SVI +0.0899%/+0.0466%. All models are in-sample; no holdout.",high23855f8c9e0508b319a7022588b7485215b722768f,True,"Tymchuk, Pylypenko & Kicha (2023)",Ukraine,Kyiv,apartments in new-build residential complexes,"asking price per m2 (developer listings, log)",169.0,"projects completed 2022-2025 (data year not stated, c. 2023)",lun.ua listing platform (scraped),"OLS hedonic with quantitative regressors (log area, rooms, crime rate, minutes to centre, year);OLS with neural-network text-derived housing class (predictedClass) + log area",OLS with NN text-derived housing class,R2,0.889,0.884,none,listing text embeddings,0.005,"R2 (in-sample, 0.884 -> 0.889; model (3) replaces most quantitative regressors by the text-derived class rather than adding to them)",coefficients,Ukrainian-language paper with English abstract. A small feed-forward network with nnlm-en-dim50 word embeddings classifies the complex description (translated to English) into 4 housing classes; the predicted class enters a log-linear regression. Model (2) conventional OLS: R2 0.884 (adj 0.880); model (3) with predictedClass + logarea: 'explains about 88.9% of variance' (R2 ~0.889). Both are in-sample fits on 169 records; no holdout. Not a pure ML-vs-hedonic prediction comparison.,medium248674ce1937729ab9e1f1084511ae7a29ecdbc635,True,"Yang, Krenz, Qiu & Li (2023)",United Kingdom,Greater London,mixed (houses and flats),transaction price (GBP per m2),49603.0,2017 onwards (Land Registry transactions since 2017),HM Land Registry Price Paid Data + EPC (Chi et al. compilation); Google Street View; online perception survey,OLS hedonic with traditional location attributes (Model 0);OLS hedonic with space-syntax location attributes at R400/R800/R2000/R6000 (Models 1-4),"OLS hedonic, Model 0 (baseline location attributes)",R2,0.496,0.496,none,street-view CNN,,,coefficients,"Hedonic OLS study, not a prediction contest: a BP neural network trained on survey ratings of Google Street View segmentations (PSPNet) predicts 8 subjective-perception scores that enter the hedonic model. Table 5: adjusted R2 (in-sample) 0.496 (Model 0), 0.482-0.494 (Models 1-4). No specification without the SP scores is reported, so no incremental gain is printed; Table 4 gives SP scores alone adj. R2 = 0.342. Standardized coefficients reported (Fig. 8).",high25871e09a43b07cddac0d3c99559043f1bd6c1a614,True,Wang (2026b),United States,"Ames, Iowa",single-family residential (Ames Housing dataset),sale price (USD),2930.0,2006-2010,Ames Housing dataset (Kaggle; Ames assessor's office),random forest,random forest,MAPE,10.43,,k-fold,none,,,SHAP,"Single RF model, Table 1 (5-fold CV, mean +/- sd): R2 0.84 +/- 0.10, RMSE 30,289.59 +/- 7,698.85 USD, MAE 17,811.77 +/- 1,886.38 USD, MAPE 10.43% +/- 1.26%. No OLS/hedonic benchmark. Global SHAP (TreeSHAP on 500 training rows, dependence plots for GrLivArea x OverallQual) plus LIME local attributions on 59 atypical test samples. Short conference paper (CONF-MPCS 2026 proceedings).",high268a99113b012c7dc5288863734eba9456a85fd906,True,"Nor, Audu & Mohamed (2025a)",Somalia,Baidoa,residential properties,monthly rent,90.0,,"local real estate agencies, government records and direct surveys (stratified random sample)",OLS hedonic regression;artificial neural network (MLP),OLS hedonic regression,MAE,29.38,29.38,unknown,none,,,coefficients,"Table 4 comparison: regression MSE 1507.13 / MAE 29 (29.38 in Table 2); ANN MSE 6920.83 / MAE 65.75 (ANN evaluated on a 60/25/15 train/test/validation split, Table 5). The regression metrics appear to be in-sample fit statistics (R2 0.96, adj. R2 0.92), so the two models are not clearly evaluated on the same test set; hedonic_value equals best_value because the hedonic model is the best-performing one. Sample size printed as 90 properties in the data section but Table 5 lists 70+30+18 = 118 samples. Rent currency not stated (likely USD). Marginal effects = regression coefficients.",low278d5f6b09f9058fc5a131a22d58a856d18ce849c7,True,Shen & Ross (2020),United States,"Atlanta, GA (MLS metropolitan market area)",single-family homes,transaction (sale) price,40851.0,2010-2017,Multiple Listing Service (MLS) sales with property descriptions,OLS hedonic;OLS hedonic + text-derived uniqueness score;repeat-sales;repeat-sales + uniqueness score,OLS hedonic + text-derived uniqueness score,R2,0.769,0.76,none,listing text embeddings,0.009,R2,coefficients,"Inferential hedonic study, not a predictive horse race: a doc2vec-style word-embedding uniqueness score of the MLS property description is added to a conventional hedonic (and a repeat-sales) model. In-sample R-squared rises from 0.760 to 0.769 in the hedonic model (Table 3) and from 0.956 to 0.958 in the repeat-sales model (n = 7,317); one SD of uniqueness = +14.9% sale price. No holdout evaluation; R2 as fraction.",high289227a7e5ee91510694ed7aa9358e1709d0798b32,True,d'Amato (2004),Italy,Bari (Carrassi-Poggiofranco / Picone zone),"residential (apartments implied, not specified)","transaction price (EUR, time-adjusted to September 2004)",69.0,transactions up to September 2004 (start not printed),"Real Estate Market Observatory, Polytechnic University of Bari",linear multiple regression analysis (MRA);rough set theory with value tolerance relation (RST),linear MRA (hedonic),MAPE,11.0,11.0,holdout,none,,,coefficients,"Early rule-based AI (rough sets) vs MRA. RST rules derived from 19 transactions and tested on the other 50; MRA fitted on all 69 (R2 0.86, adj. 0.85) and evaluated on the same 50 (in-sample for MRA). Text: RST 18% MAPE vs MRA 11% MAPE, valuation variation 15.31%; MRA judged superior. Tables 1 and 5 are images not captured in the text layer. RST gives no marginal prices.",medium299544f4ac839e433b51f55da090f58729fbbf9919,True,"Hong, Choi & Kim (2020)",South Korea,Seoul (Gangnam district),apartments,transaction price,16601.0,2006-2017,"Apartment transaction records, Gangnam district (approx. 40% of all 39,564 transactions)",OLS hedonic;random forest,random forest,MAPE,5.482407,19.60567,holdout,none,,,none,"Random 90/10 train/test split, averaged over 10 trials (Table 5). MAPE in %; same table reports R-squared 0.976198 (RF) vs 0.726056 (OLS) and COD 5.48 vs 19.57; hitting rates within 5%: 72.0% vs 17.4%. Only RF variable-importance (Figure 4) is reported, no SHAP/PDP. Text states 16,601 samples in data section but 16,061 in results section.",high309bd166858991ed89e221a8fb56169533ad4f2ece,True,Wang et al. (2026a),China,Shanghai (central urban area),second-hand flat-type apartments in slab blocks,listing price per m2 (yuan),7634.0,2024,Beike (ke.com) second-hand housing listings; Baidu street-view panoramas (semantic segmentation view indices),gradient boosted decision trees (GBDT),GBDT,R2,0.47,,holdout,street-view CNN,,,PDP,"Single GBDT model (grid search + 5-fold CV on training data, evaluated on a held-out test set): R2 = 0.47, MAE = 7,052.82, RMSE = 12,968.90 (yuan; mean price 60,178.85 yuan, i.e. per m2). No OLS/hedonic benchmark and no ablation of the street-view variables. Street-view view indices (sky, tree, grass, building, road, sidewalk, land diversity) from semantic segmentation of Baidu panoramas; relative importance and partial dependence plots reported.",high31a072fb43e67fcc553e457777f0c68324ace19eb7,True,Zhong (2024),United States,,houses,"listed/transaction price (unspecified, Kaggle)",545.0,,"Kaggle 'Housing Price Prediction' dataset (Harish Kumar, 2023)",OLS multiple linear regression (log price);random forest,OLS multiple linear regression,RMSE,0.2032158,0.2032158,unknown,none,,,coefficients,"Table 3: MLRM RMSE 0.2032158 / MAPE 0.0103682 vs RF RMSE 0.2094694 / MAPE 0.0107626 (dependent variable is log housing price, so RMSE is in log units and MAPE is a fraction of log price). The OLS hedonic is itself the best model, so hedonic_value equals best_value. No train/test split is described (544 obs after IQR outlier removal), so test-set design is unknown. Multiple R2 of the OLS = 0.697 (adjusted 0.6896). Conference proceedings, very small sample.",medium32a483774defcbadf7676aaa742f594af28562b13a,True,"Lenaers, Boudt & De Moor (2023)",Belgium,"Belgium (national, private dataset)",residential properties (rentals),"monthly gross rent (agent estimates, indexed to 2019 EUR)",18935.0,2019-01-01 to 2022-05-04,Private dataset from a Belgian real estate player (agent rent estimates) + OpenStreetMap POI features,linear regression (OLS);random forest;XGBoost,XGBoost,MAPE,10.33,11.59,holdout,none,,,SHAP,"Table II, test set: XGBR MAE 104.68 EUR, RMSE 210.71 EUR, MAPE 10.33%; RFR 110.07/221.95/10.83%; LR 120.82/246.47/11.59%. Models tuned with TPE and 10-fold CV on the train set, evaluated on a holdout test set. Caveat: the logarithm of the indexed asking price is included as a predictor, which inflates accuracy for all models. Interpretation via SHAP feature-importance and SHAP summary plots (asking price and cadastral income dominant).",high33b922ab81201881b47779769df9e76de174e13df7,True,"Oliveira, Reyes-Bueno, Gonzalez & da Silva (2024/2025)",Brazil,Fortaleza,apartments,listed prices and actual sales (log price),43585.0,2017 and 2021,"Real Estate Market Observatory, Local Finance Secretary of Fortaleza",MRA (OLS);spatial error regression;MARS;M5P;random forest;quantile random forest;GBM;LightGBM;CatBoost;XGBoost;deep learning,XGBoost,MAPE,8.96,13.9,holdout,none,,,coefficients,"Appendix Table B, stratified random split 32,688 train / 10,897 test: MAPE 8.96% XGBoost, 9.05% RF, 9.25% GBM, 13.90% MRA, 11.02% SER; R2 0.96 for RF/GBM/LightGBM/XGBoost vs 0.88 MRA and 0.94 SER; RMSE in BRL (XGBoost 116,633.77; GBM lowest 113,640.60; MRA 197,400.28); COD 8.98% for RF and XGBoost. XGBoost is the only model with no residual spatial autocorrelation (Moran's I). Coefficients reported for MRA/SER only; ML models are not explained (authors defer interpretation to another study).",high34ba6d3043dcfb0e7c6ad45ae35c0629afbc728c5f,True,Haque (2003),United States,"College Station, Texas",single-family homes (structurally homogeneous subsample),"sale price per square foot, adjusted to 2000 values (MLS)",145.0,1997-2000,"MLS sales via local broker, merged with Brazos County Appraisal District",OLS hedonic regression;Jordan-Elman recurrent ANN,Jordan-Elman recurrent ANN,R2,0.8627,,holdout,none,,,coefficients,"ANN trained on 130 obs and evaluated on 15 arbitrarily separated obs: R2 0.7788 (training) and 0.8627 (evaluation set). Hedonic OLS R2 = 0.622 is an in-sample fit on all 145 obs, not the same test set, so hedonic_value is null. Only inputs are age and four siding-material dummies; hedonic coefficients give implicit siding premiums. Conference paper (ASEE), tiny sample.",medium35bc5c2c3c86f1f45909dcce75c8f09acff914dd28,True,"An, Song, Jang & Ahn (2025)",South Korea,Busan,condominiums (apartments),transaction price (log),61152.0,2015,Song et al. (2021) hedonic dataset of condominium transactions,OLS hedonic;spatial lag regression (SLR);random forest;deep neural network,random forest,R2,0.941,,holdout,none,,,SHAP,"70/30 train-test split. RF baseline RMSE 0.080 (log price) and R2 0.941 printed in appendix Table 8 (tuned RF 0.079/0.942). OLS and SLR accuracy appear only graphically in Fig. 3 (OLS R2 roughly 0.67, RMSE roughly 0.19 by eye), so hedonic_value is left null. Compares SHAP ranks with Gini importance and with OLS/SLR coefficient signs (Tables 5-6).",medium36c911b3fd22af6e787ff69e7f09ef0ceca9cbebd6,True,Marimon & Vergara (2026),Philippines,Metro Manila,residential listings (not specified),listing price (log),1007.0,"listings scraped May 2024, January 2025, March 2025",Lamudi listings; Google Street View (PSPNet/ADE20K segmentation for Green View Index); OpenStreetMap POIs,"linear hedonic regression (overall, LCZ-based and city-based models)",linear hedonic regression (overall model),R2,0.547,0.547,holdout,street-view CNN,,,coefficients,"Hedonic OLS with a CNN-derived Green View Index averaged in 400 m buffers; 80/20 split (805 train / 202 test). Table 1: training R2 0.495, testing R2 0.547, testing MSE 0.446 (log price). No model without GVI is reported. GVI coefficient 0.107 (standardised), p<0.001; effects vary by local climate zone and city.",high37d8361358e99098ffa7a8408974e7bea4a38e7010,True,"Liu, Chen, Orford, Tian & Zou (2024)",United Kingdom,Cardiff,"mixed (detached, semi-detached, terraced, flat/maisonette)",transaction price (log),16297.0,2001-2007,"HM Land Registry (with floor area), OS street network",XGBoost;random forest;gradient boosting decision tree,XGBoost,R2,0.76,,holdout,none,,,SHAP,"Four XGBoost models (closeness/betweenness at 600 m, 3 km, 6 km, 10 km radii); Table 2 reports R2 0.76/0.76/0.76/0.74, MAE 0.18-0.19 and RMSE 0.28-0.30 in log-price units on a 10% random holdout (5-fold CV for tuning). RF and GBDT were said to perform worse but their metrics are not printed; no OLS benchmark. SHAP dependence plots give nonlinear marginal effects of accessibility.",high38de48aa9a311d709ef0a33453d7042001c2e721a6,True,"Fan, Teo & Wan (2021)",Singapore,Bukit Panjang (Singapore),public housing (HDB) resale flats,transaction price (log),893.0,2010-2018,government open data portal (resale transactions); text noise-complaint records scored with SentimentR,OLS hedonic with lagged noise-sentiment index;IV/DID variants,OLS hedonic with block-level noise sentiment,R2,0.919,0.919,none,text sentiment,,,coefficients,"Hedonic/causal study, not a prediction contest: dictionary-based sentiment (SentimentR, NLTK-style) of 527 noise complaints builds a block-level noise index; Table 6 log-price regressions on 893 matched resale transactions, R2 0.919 (in-sample, with controls and fixed effects). Coefficient -0.030 (1 scale-point increase in noise sentiment -> about 3% lower price). No specification without the sentiment feature reports fit, so no incremental gain.",high39doi_10_1002_isaf_287,True,Kagie & van Wezel (2007),Netherlands,"Six submarkets: Groningen, Apeldoorn, Eindhoven, Amsterdam, Rotterdam, Zeeland","existing houses (all types, incl. apartments)",transaction price (EUR),27735.0,2004,"NVM (Dutch brokers association) transactions of existing houses, 83 explanatory variables (+ StatLine demographic data in extension)",linear regression (untransformed OLS);CART;LSBoost (boosted regression trees);LADBoost,"boosted regression trees (LSBoost/LADBoost, depth-5 trees)",MAPE,9.38,,holdout,none,,,PDP,"Metric printed as mean relative error (MRE, %) and MAE in EUR; values are averages over 100 random 90/10 train/test splits, reported per region (Table 3). Best_value = Amsterdam (largest region, n = 8,490): LADBoost MRE 9.38%, MAE 29,320 EUR, printed improvement vs linear regression 44.00% (MRE) and 32.84% (MAE); other regions: Groningen 11.32% (+32.01%), Apeldoorn 9.31%, Eindhoven 8.06%, Rotterdam 10.91%, Zeeland 14.83%, improvements 32-44% in MRE. Linear-regression totals per region are not printed (only per house type in Table 17), so hedonic_value is null. n_obs = sum of the six regional datasets (3387+2402+3730+8490+7510+2216). Interpretation via relative importance and partial dependence plots; adding demographic data improves MRE a further 4-9%. Working paper EI 2006-17, published 2007 in Intell. Syst. Account. Finance Manag.",medium40ee4a73efc5ad67b4400db23b09eeacbeb4812f86,True,"Kim, Choi & Lee (2025)",South Korea,Seoul (Gangnam District),apartments (condominiums),transaction price (ten-thousands KRW),15162.0,2006-2017,"Ministry of Land, Infrastructure and Transport (MOLIT) apartment transaction records",multiple linear regression (OLS);decision tree;random forest;XGBoost;CatBoost;kNN,random forest,MAPE,5.928,12.227,k-fold,none,,,SHAP,"5-fold CV (Section 5 and Appendix Table A1): random forest R2 0.970, MAPE 0.05928 (=5.928%), RMSE 8114.391, MAE 4044.365; multiple linear regression R2 0.892, MAPE 0.12227 (=12.227%), RMSE 20852.602, MAE 14353.423 (RMSE/MAE in ten-thousands KRW). 10-fold CV also reported (RF R2 0.972, MLR R2 0.799). SHAP (TreeSHAP) and permutation feature importance for both MLR and RF, plus temporal sub-period analysis; MLR coefficients also reported.",high41fe9b5a96e7007d52e9bebe9eaea418541aa3b090,True,"Tajani, Di Liddo, Ranieri & Anelli (2021)",Italy,"Milan, Turin, Rome, Naples, Catania (five separate samples)",residential units (rented),rent (EUR per m2 per month),1100.0,second half of 2019,rented properties collected with local market operators and real estate agents (220 per city),"evolutionary polynomial regression (EPR, genetic-algorithm AVM)",evolutionary polynomial regression (EPR),R2,0.8869,,k-fold,none,,,coefficients,"Single-method AVM (genetic-programming polynomial regression) estimated city by city; n_obs = 5 x 220. Coefficient of determination printed per city: Milan 88.69%, Turin 79.31%, Rome 86.87%, Naples 74.92%, Catania 67.06% (average about 80%); best_value is the Milan model. RMSE/MAPE/MaxAPE are mentioned but their values are not printed in the text layer; a ten-fold cross-validation is said to confirm stability without printed CV metrics. Marginal contributions derived from the polynomial terms (percentage variation of rent per variable).",medium42