1# RECORD bc5c2c3c86f1f45909dcce75c8f09acff914dd282# TITLE: Toward transparent and accurate housing price appraisal: Hedonic price models versus machine learning algorithms3# YEAR: 2025.0 VENUE: Financial Innovation DOI: 10.1186/s40854-025-00874-w4# AUTHORS: S. An; Yena Song; H. Jang; Kwangwon Ahn5# PAGES: 29 CANDIDATE TABLE PAGES (1-based): [22, 25]6# ABSTRACT: The nonlinearity of hedonic datasets demands flexible automated valuation models to appraise housing prices accurately, and artificial intelligence models have been employed in mass appraisal to this end. However, they have been referred to as “black-box” models owing to difficulties associated with interpretation. In this study, we compared the results of traditional hedonic pricing models with those of machine learning algorithms, e.g., random forest and deep neural network models. Commonly implemented measures, e.g., Gini importance and permutation importance, provide only the magnitude of each explanatory variable’s importance, which results in ambiguous interpretability. To address this issue, we employed the SHapley Additive exPlanation (SHAP) method and explored its effectiveness through comparisons with traditionally explainable measures in hedonic pricing models. The results demonstrated that (1) the random forest model with the SHAP method could be a reliable instrument for appraising housing prices with high accuracy and sufficient interpretability, (2) the interpretable results retrieved from the SHAP method can be consolidated by the support of statistical evidence, and (3) housing characteristics and local amenities are primary contributors in property valuation, which is consistent with the findings of previous studies. Thus, our novel methodological framework and robust findings provide informative insights into the use of machine learning methods in property valuation based on the comparative analysis.789===== PAGE 1 (score 7, numeric-paren 0, table-words 0, var-words 7) =====1011Toward transparent and accurate housing 12price appraisal: Hedonic price models 13versus machine learning algorithms14Sihyun An1, Yena Song2, Hanwool Jang3 and Kwangwon Ahn4* 15Introduction16Real estate is an essential sector of social and economic systems, and real estate prices 17fluctuate across time and place, which can affect economic stability (Glaeser et al. 2014; 18Tchuente and Nyawa 2022). Capturing dynamic patterns in real estate prices requires 19flexible valuation models that utilize sophisticated algorithms to precisely capture the 20nonlinearity of housing price datasets. Moreover, housing prices are affected by vari-21ous attributes (Anderson and West 2006; Czembrowski and Kronenberg 2016; An et al. 222024); therefore, the interpretability of model outcomes has been accentuated to rein-23force our understanding of the pattern in property prices. This approach enables us to 24Abstract 25The nonlinearity of hedonic datasets demands flexible automated valuation models 26to appraise housing prices accurately, and artificial intelligence models have been 27employed in mass appraisal to this end. However, they have been referred to as “black-28box” models owing to difficulties associated with interpretation. In this study, we 29compared the results of traditional hedonic pricing models with those of machine 30learning algorithms, e.g., random forest and deep neural network models. Com-31monly implemented measures, e.g., Gini importance and permutation importance, 32provide only the magnitude of each explanatory variable’s importance, which results 33in ambiguous interpretability. To address this issue, we employed the SHapley Addi-34tive exPlanation (SHAP) method and explored its effectiveness through comparisons 35with traditionally explainable measures in hedonic pricing models. The results dem-36onstrated that (1) the random forest model with the SHAP method could be a reliable 37instrument for appraising housing prices with high accuracy and sufficient interpreta-38bility, (2) the interpretable results retrieved from the SHAP method can be consolidated 39by the support of statistical evidence, and (3) housing characteristics and local ameni-40ties are primary contributors in property valuation, which is consistent with the find-41ings of previous studies. Thus, our novel methodological framework and robust find-42ings provide informative insights into the use of machine learning methods in property 43valuation based on the comparative analysis.44Keywords: Hedonic price model, Importance measure, Machine learning, Housing 45price appraisal46Open Access47© The Author(s) 2025. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits 48use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original 49author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third 50party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the mate-51rial. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or 52exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://53creativecommons.org/licenses/by/4.0/.54RESEARCH55An et al. Financial Innovation (2025) 11:141 56https://doi.org/10.1186/s40854-025-00874-w57Financial Innovation58*Correspondence: 59k.ahn@yonsei.ac.kr601 Division of Finance and AI, 61Hankuk University of Foreign 62Studies, Yongin, South Korea632 Department of Geography, 64Chonnam National University, 65Gwangju, South Korea663 Department of Finance, 67Accounting and Risk, Glasgow 68Caledonian University, Glasgow, 69UK704 Department of Industrial 71Engineering, Yonsei University, 72Seoul, South Korea737475===== PAGE 22 (score 28, numeric-paren 0, table-words 6, var-words 13) =====7677Page 22 of 2978An et al. Financial Innovation (2025) 11:141 79used to represent the traditional approach. In addition, the RF model was used for this 80comparative analysis because of its robustness.81The benchmark dataset includes housing sale prices as the target variable and com-82prises 80 explanatory variables divided into a training set with 1,460 observations and a 83test set with 1,459 observations. Prior to developing the property valuation models, we 84conducted exploratory data analysis and applied feature engineering to refine the data-85set and extract informative variables. During this process, we observed that some vari-86ables had many missing values. For example, the PoolQC variable (indicating the pool 87quality) had missing values in 99.5% of the total observations, and the MiscFeature vari-88able (indicating the miscellaneous feature) was missing in approximately 1,406 of the 89observations (representing approximately 96.3%). In addition, variables, such as Utilities 90(type of available utilities), exhibited identical values across 1,459 observations, thereby 91indicating little variability with a minimal impact on housing price dynamics. Thus, we 92eliminated these variables from the dataset because they were deemed noise rather than 93meaningful contributors to appraising housing prices.94In addition, we note that there are still missing values in the preprocessed dataset. The 95benchmark dataset comprises both qualitative and quantitative variables; thus, the miss-96ing values were imputed by the missing-indicator method (Zhuchkova and Rotmistrov 972022) for the qualitative variables and the mean imputation (Silva-Ramírez et al. 2011; 98Khan and Hoque 2020) method for the quantitative variables. The missing-indicator 99method imputes missing values in a qualitative variable by a numerical value indicating 100the absence, which has been widely used in previous studies (Trevizo and Lopez 2016; 101Table 9 Comparisons of two permutation importance measurements102The average score is the mean value of the importance score, which is obtained by repeating the measurement of 103permutation importance that shuffles the data points in variables over 10 iterations104Variables105Shuffling method (10 iterations)106Omitting 107method108OLS109SLR110OLS111SLR112Rank113Average score114Rank115Average score116Rank117Rank118Higher degree11911200.060412111220.059912311241125Year12621270.048412821290.048813021312132Parking13331340.035413531360.035413731383139Area14041410.021314241430.021214441454146Households14761480.009814961500.009715151525153CBD15451550.013215651570.012615861596160Floor16171620.007516371640.007416571667167Greenspace16881690.005517081710.005617281738174Waterfront17591760.001917791780.001917991809181Subway182111830.0010184111850.00091861018710188Top school189121900.0008191121920.00081931119411195Spring196101970.0010198101990.00102001220112202Fall203152040.0004205152060.00042071320813209Winter210142110.0006212132130.00062141521515216Pop. density217132180.0006219142200.00052211422214223Heating type224162250.0002226162270.00022281622916230Bus count231172320.0001233172340.00012351723617237238239===== PAGE 23 (score 7, numeric-paren 0, table-words 3, var-words 1) =====240241Page 23 of 29242An et al. Financial Innovation (2025) 11:141 243 244Weiss et al. 2017). For the quantitative variables, the mean imputation method is a gen-245eral approach for handling missing values (Silva-Ramírez et al. 2011) with low compu-246tational costs (Khan and Hoque 2020) because it replaces the missing values with the 247average value for the considered variable. These processes preserve most of the obser-248vations across the entire dataset for training valuation models. Finally, the training set 249contained a total of 1,460 observations with 68 variables.250Considering the large number of housing variables in the refined dataset compared 251with our hedonic dataset, we conducted principal component analysis, which is a gen-252eral feature engineering technique in the modeling process that effectively reduces the 253dimensionality of variable sets (Noori et al. 2010; Wang et al. 2016; Zhang et al. 2018). 254Here, we considered two sets of principal components determined by two points where 255explainable variances over 90% and 95% of the total variance were achieved. Figure 6 256shows the cumulative explainable variance across varying numbers of principal compo-257nents. In particular, 43 principal components explained 90.23% of the variance, and 51 258principal components explained 95.31% of the variance.259Accordingly, we compared the traditional approach (i.e., the OLS regression model) 260against the sophisticated RF model in terms of the RMSE value. Three cases were inves-261tigated: (1) using the full variable set, (2) using the 43 principal components set, and (3) 262using the 51 principal components set. Table 10 shows the results obtained on the Kag-263gle housing dataset. The RF model consistently achieved lower RMSE values than those 264of the OLS-based regression model for all the cases. In our analysis, the case without 265feature engineering exhibited the highest accuracy in the assessment of housing prices. 266For each case, the RF model demonstrated an average increase in accuracy of 18% com-267pared with the OLS-based regression model. These results are well aligned with the 268Fig. 6 Cumulative explainable variance of principal components269Table 10 RMSE results obtained on the Kaggle housing dataset270The numbers in parentheses indicate each case: (1) Full variable set, (2) 43 principal components, and (3) 51 principal 271components272OLS-based regression273RF274(1)275(2)276(3)277(1)278(2)279(3)280RMSE2810.17492820.18612830.18962840.14512850.15312860.1532287288289===== PAGE 25 (score 20, numeric-paren 0, table-words 8, var-words 2) =====290291Page 25 of 29292An et al. Financial Innovation (2025) 11:141 293 294SLRs in terms of both R2 and RMSE values because of their ability to mitigate biased 295estimates (López-Laborda et al. 2021). These results are consistent with those of pre-296vious studies (Han and Jang 2013; López-Laborda et al. 2021). However, the RF model 297consistently demonstrates the highest predictive power in all the cases, emphasizing that 298RF-based valuation models can be considered a desirable approach for precisely apprais-299ing housing prices.300Additional accuracy metrics and processing time301We use the additional statistical indices, mean absolute error (MAE) and mean absolute 302percentage error (MAPE), to further assess the predictive power of the models. These 303indices are commonly employed metrics in model validation (Ho et al. 2021). We train 304each model using the complete set of hedonic variables to ensure comparability. We also 305examine the processing time of the ML and DL models, as convergence speed is a key 306factor in evaluating the computational efficiency of AI-based systems (Yates and Islam 3072021; Abed et al. 2022).308Table 12 presents the predictive power of the HPMs and AI models. Our findings 309demonstrate that the RF model consistently outperforms the other models across all 310the evaluation metrics, including the RMSE, R2, MAE, and MAPE. Furthermore, the RF 311model requires approximately 13% of the processing time of DNN models, underscoring 312its computational efficiency. These findings confirm the superiority of the RF model in 313terms of precision and computational efficiency.314According to Yates and Islam (2021), the computational complexity of the RF model 315primarily depends on the number of trees. Correspondingly, a smaller ensemble (i.e., 316a reduced number of tree estimators) consumes fewer computational resources (Yates 317and Islam 2021). Nevertheless, the fine-tuning experiment for RF in Table 8 shows that 318the RF configuration with fewer trees can achieve predictive power comparable to those 319with a larger number of trees. As such, our resultant measures in Table 12 show that the 320RF model can be a more efficient tool for housing price valuation with fewer trees while 321maintaining comparable predictive accuracy.322Abbreviations323AI 324Artificial intelligence325DL 326Deep learning327DNN 328Deep neural network329GLM 330Generalized linear model331Table 12 Model validation with additional accuracy measures and processing times332A lower value of the reported metric indicates that the model has greater predictive power, excluding R2. The “Processing 333time (sec.)” column shows the convergence time of the model training process, measured in seconds. The row “DNNs” 334indicates the case of the DNN model with min–max scaling335RMSE336R2337MAE338MAPE339Processing 340time (sec.)341OLS3420.1943430.6533440.1433450.010346−347SLR3480.1943490.6513500.1433510.010352−353RF3540.0813550.9393560.0533570.00435814.340359DNN3600.1863610.6813620.1373630.009364112.669365DNNs3660.1273670.8503680.0893690.006370116.708371372373===== PAGE 26 (score 11, numeric-paren 0, table-words 0, var-words 11) =====374375Page 26 of 29376An et al. Financial Innovation (2025) 11:141 377GMSE 378Generalized mean-squared error379HPM 380Hedonic price model381MAE 382Mean absolute error383MAPE 384Mean absolute percentage error385ML 386Machine learning387OLS 388Ordinary least squares389ReLU 390Rectified linear unit391RF 392Random forest393RMSE 394Root mean square error395SHAP 396SHapley Additive exPlanation397SLR 398Spatial lag regression399Acknowledgements400The authors would like to thank the editors and the anonymous reviewers for their invaluable comments and 401suggestions.402Author contributions403All the authors contributed equally.404Funding405This work was supported by (i) the National Research Foundation of Korea grant funded by the Korea government (MSIT) 406(RS-2025-16067531: Kwangwon Ahn) and (ii) Hankuk University of Foreign Studies Research Fund (0f 2025: Sihyun An).407Availability of data and materials408The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable 409request.410Declarations411Competing interests412The authors declare that they have no competing interests.413Received: 7 July 2023 Accepted: 6 November 2025414References415Abed M, Imteaz MA, Ahmed AN, Huang YF (2022) Modelling monthly pan evaporation utilising random forest and deep 416learning algorithms. Sci Rep 12(1):13132417Ahn K, Jang H, Song Y (2020) Economic impacts of being close to subway networks: a case study of Korean metropolitan 418areas. Res Transp Econ 83:100900419Ali O, Abdelbaki W, Shrestha A, Elbasi E, Alryalat MA, Dwivedi YK (2023) A systematic literature review of artificial intel-420ligence in the healthcare sector: benefits, challenges, methodologies, and functionalities. J Innov Knowl 8(1):100333421Alonso Robisco A, Carbó JM (2022) Measuring the model risk-adjusted performance of machine learning algorithms in 422credit default prediction. Financ Innov 8(1):70423An S, Ahn K, Bae J, Song Y (2024) Economic impacts of a subway system: exploring local contexts in a metropolitan area. 424Res Transp Bus Manag 56:101188425An S, Jang H, Kim H, Song Y, Ahn K (2023) Assessment of street-level greenness and its association with housing prices in 426a metropolitan area. Sci Rep 13(1):22577427Anderson ST, West SE (2006) Open space, residential property values, and spatial context. Reg Sci Urban Econ 42836(6):773–789429Anselin L (2013) Spatial econometrics: methods and models, vol 4. Springer, Heidelberg430Asif MR, Bording TS, Barfod AS, Grombacher DJ, Maurya PK, Christiansen AV, Auken E, Larsen JJ (2021) Effect of data pre-431processing on the performance of neural networks for 1-D transient electromagnetic forward modeling. IEEE Access 4329:34635–34646433Bailey JR, Lauria D, Lindquist WB, Mittnik S, Rachev ST (2022) Hedonic models of real estate prices: GAM models; environ-434mental and sex-offender-proximity factors. J Risk Financ Manag 15(12):601435Bax D, Chasomeris MG (2019) Listing price estimation of apartments: a generalised linear model. J Econ Financ Sci 43612(1):1–11437Breiman L (2001) Random forests. Mach Learn 45(1):5–32438Calle ML, Urrea V (2011) Letter to the editor: stability of random forest importance measures. Brief Bioinform 12(1):86–89439Chen Y, Ye Y, Liu X, Yin C, Jones CA (2025) Examining the nonlinear and spatial heterogeneity of housing prices in urban 440Beijing: an application of GeoShapley. Habitat Int 162:103439441Chien CY, Hsu SW, Lee TL, Sung PS, Lin CC (2020) Using artificial neural network to discriminate Parkinson’s disease from 442other parkinsonisms by focusing on putamen of dopamine transporter SPECT images. Biomedicines 9(1):12443Chin T, Chau K (2003) A critical review of literature on the hedonic price model. Int J Hous Sci Appl 27(2):145–165444Chuan CH, Sun R, Tian S, Tsai WHS (2024) EXplainable artificial intelligence (XAI) for facilitating recognition of algorithmic 445bias: an experiment from imposed users’ perspectives. Telemat Inform 91:102135446Czembrowski P, Kronenberg J (2016) Hedonic pricing and different urban green space types and sizes: insights into the 447discussion on valuing ecosystem services. Landsc Urban Plan 146:11–19448