1# RECORD 1462a807ace2b4c97b7b8368ffbe3a1c694cd0f72# TITLE: Integrating Floor Plans into Hedonic Models for Rent Price Appraisal3# YEAR: 2021.0 VENUE: The Web Conference DOI: 10.1145/3442381.34499674# AUTHORS: K. Solovev; Nicolas Pröllochs5# PAGES: 10 CANDIDATE TABLE PAGES (1-based): [4, 7]6# ABSTRACT: Online real estate platforms have become significant marketplaces facilitating users’ search for an apartment or a house. Yet it remains challenging to accurately appraise a property’s value. Prior works have primarily studied real estate valuation based on hedonic price models that take structured data into account while accompanying unstructured data is typically ignored. In this study, we investigate to what extent an automated visual analysis of apartment floor plans on online real estate platforms can enhance hedonic rent price appraisal. We propose a tailored two-staged deep learning approach to learn price-relevant designs of floor plans from historical price data. Subsequently, we integrate the floor plan predictions into hedonic rent price models that account for both structural and locational characteristics of an apartment. Our empirical analysis based on a unique dataset of 9,174 real estate listings suggests that current hedonic models underutilize the available data. We find that (1) the visual design of floor plans has significant explanatory power regarding rent prices – even after controlling for structural and locational apartment characteristics, and (2) harnessing floor plans results in an up to 10.56 % lower out-of-sample prediction error. We further find that floor plans yield a particularly high gain in prediction performance for older and smaller apartments. Altogether, our empirical findings contribute to the existing research body by establishing the link between the visual design of floor plans and real estate prices. Moreover, our approach has important implications for online real estate platforms, which can use our findings to enhance user experience in their real estate listings.789===== PAGE 1 (score 12, numeric-paren 0, table-words 0, var-words 12) =====1011Integrating Floor Plans into Hedonic Models for12Rent Price Appraisal13Kirill Solovev14kirill.solovev@wirtschaft.uni-giessen.de15University of Giessen16Germany17Nicolas Pröllochs18nicolas.proellochs@wi.jlug.de19University of Giessen20Germany21ABSTRACT22Online real estate platforms have become significant marketplaces23facilitating users’ search for an apartment or a house. Yet it remains24challenging to accurately appraise a property’s value. Prior works25have primarily studied real estate valuation based on hedonic price26models that take structured data into account while accompanying27unstructured data is typically ignored. In this study, we investigate28to what extent an automated visual analysis of apartment floor29plans on online real estate platforms can enhance hedonic rent30price appraisal. We propose a tailored two-staged deep learning ap-31proach to learn price-relevant designs of floor plans from historical32price data. Subsequently, we integrate the floor plan predictions33into hedonic rent price models that account for both structural and34locational characteristics of an apartment. Our empirical analysis35based on a unique dataset of 9,174 real estate listings suggests that36current hedonic models underutilize the available data. We find37that (1) the visual design of floor plans has significant explanatory38power regarding rent prices – even after controlling for structural39and locational apartment characteristics, and (2) harnessing floor40plans results in an up to 10.56% lower out-of-sample prediction41error. We further find that floor plans yield a particularly high gain42in prediction performance for older and smaller apartments. Alto-43gether, our empirical findings contribute to the existing research44body by establishing the link between the visual design of floor45plans and real estate prices. Moreover, our approach has important46implications for online real estate platforms, which can use our47findings to enhance user experience in their real estate listings.48CCS CONCEPTS49• Applied computing →Electronic commerce; Economics; •50Information systems →Decision support systems; • Computing51methodologies →Modeling methodologies; Neural networks.52KEYWORDS53Online real estate platforms, hedonic price models, floor plans,54visual analytics, image sentiment55ACM Reference Format:56Kirill Solovev and Nicolas Pröllochs. 2021. Integrating Floor Plans into He-57donic Models for Rent Price Appraisal. In Proceedings of the Web Conference58This paper is published under the Creative Commons Attribution 4.0 International59(CC-BY 4.0) license. Authors reserve their rights to disseminate the work on their60personal and corporate Web sites with the appropriate attribution.61WWW ’21, April 19–23, 2021, Ljubljana, Slovenia62© 2021 IW3C2 (International World Wide Web Conference Committee), published63under Creative Commons CC-BY 4.0 License.64ACM ISBN 978-1-4503-8312-7/21/04.65https://doi.org/10.1145/3442381.3449967662021 (WWW ’21), April 19–23, 2021,Ljubljana, Slovenia. ACM, New York, NY,67USA, 10 pages. https://doi.org/10.1145/3442381.344996768169INTRODUCTION70Online real estate platforms have become significant marketplaces71for the real estate industry. These online platforms provide an72overview of available properties in a specific location and facilitate73users’ search for an apartment or a house [39]. In 2011, one of the74biggest online real estate providers, Zillow, reported 24 million75unique visitors. Today, this number has risen to 196 million and76stays on the upwards trend, indicating continual interest from77prospective home buyers and tenants [40]. While the number of78people referring to online real-estate platforms continues to grow,79there is a rising need for tools and recommendation algorithms that80allow for an accurate appraisal of real estate prices [e. g. 39].81Prior works have primarily studied real estate valuation based on82hedonic price models [25, 34, 37]. Hedonic pricing theory suggests83that the valuation of housing values or rents can be viewed as a84weighted sum of its features [37]. Specifically, the price and rent85are determined by the attributes and characteristics of the dwelling86unit and the surrounding neighborhood. Hedonic models provide a87high degree of flexibility when selecting the attributes, which can88roughly be divided into structural and locational attributes [26].89The structural factors describe an apartment itself (e. g. number90of rooms, amenities, parking, pool), while the locational attributes91are composed of external features affecting the price. For instance,92previous studies [e. g. 30, 33] have examined the impact of the93distance to transport hubs and shopping centers on real estate94prices. Altogether, the hedonic approach provides scholars and95practitioners with a framework of assessing the value of real estate96properties, where the final price entails a precise valuation of a97diverse feature package.98While structural and locational attributes are of great importance99for real estate price models, modern online real estate marketplaces100provide additional information that has been neglected in previ-101ous works. A particularly relevant feature for housebuyers and102prospective tenants may be real estate floor plans. According to103rightmove.co.uk, 90 % of home-buyers think that a floor plan is an104essential part of the decision making process when finding an apart-105ment or house. A real estate floor plan is a schematic illustration106that provides a top-down view of the property. While information107about floor plans may be partially provided in structured form (e. g.,108in floor plan filings), the actual floor plan images may contain price-109relevant hidden information. For example, the relationship between110rooms and spaces, as well as the overall layout are typically only111available from the floor plan image itself. Information on floor plans112is typically also not available in textual form as the vast majority113arXiv:2102.08162v1 [cs.LG] 16 Feb 2021114115116===== PAGE 4 (score 25, numeric-paren 2, table-words 2, var-words 15) =====117118WWW ’21, April 19–23, 2021, Ljubljana, Slovenia119Kirill Solovev and Nicolas Pröllochs120in hedonic pricing models – even after controlling for structured121apartment characteristics such as size and location.1223123DATA1243.1125Apartment Listings126The data for this study was extracted from ImmobilienScout24.de,127the largest German online real estate aggregator. At the end of1282019, ImmobilienScout24 contained 91,415 active rent listings across129Germany, with approximately 44 million visitors per month.130We implemented a Python-based web crawler to download and131store information about apartments in the city of Berlin in a data-132base format. The crawler was active for three weeks at the end133of 2019, during which it gathered a total number of 15,604 obser-134vations. Each observation represents an apartment in Berlin from135an active or archived listing on ImmobilienScout24. Our dataset136includes apartment listings between mid-2017 and end of 2019 (i. e.,137spans two and a half years). For each apartment, the crawler col-138lected a raw image showing the floor plan of the apartment, the139monthly rent price, the monthly rent price per m2, the city district140in which the apartment is located, and 36 fields with categorical141and numerical information about the apartment (e. g., apartment142size, etc.)1433.2144Variable Definitions145Dependent variables: The target variables in our study are two146measures of the monthly costs of an apartment:147• Rpms: The monthly rent price (in e) per m2148• Rent: The total monthly rent price (in e) of an apartment149Explanatory variables: Our hedonic rent price model accounts150for the following apartment characteristics from previous works:151• Area: The total size of the apartment in m2152• Rooms: The total number of rooms in the apartment153• Floor: The floor at which the apartment is located154• TopFloor: The number of floors of the building in which155the apartment is located156• CityDistance: The distance (in km) from the center of the157city to the apartment158• YearBuilt: The construction year of the building in which159the apartment is located160Control variables: Our dataset further includes 12 locational161dummies that provide information about the city district, and a162set of 30 additional control variables in the form of categorical163dummies that provide fine-grained information about contractual164and structural characteristics of an apartment (e. g. whether the165apartment offers access to a parking lot).1663.3167Summary Statistics168Dependent variables: Table 1 shows summary statistics, whereas169Fig. 3 plots the complementary cumulative distributions (CCDFs)170for the dependent variables. The total rent prices (Rent) in our171data range from e230 per month to almost e3000 with a standard172deviation of 476.80. The rent per m2 (Rpms) ranges from 5.38 to17327.80 with a standard deviation of 4.32. Fig. 1 provides an overview174of the monthly apartment costs across different districts in Berlin.175Evidently, there is a higher concentration of apartments with high176monthly costs near the city center of Berlin.177Explanatory variables: Fig. 2 plots the distributions of the178explanatory variables in our dataset. The size of the apartments179(Area) ranges from 16.00 to 178.08 square meters with a standard180deviation of 26.35. The number of rooms ranges from 1 to 6, with181most of the apartments having either 2 or 3 rooms. The average182distance between an apartment and the city center (CityDistance)183is 7.99 km. The variables TopFloor and Floor have the same mini-184mum of 0 and a maximum of 27 and 26, respectively. The houses in185our sample have been built between the years 1830 and 2020. Most186buildings are relatively new with a median construction year of1871985.188(a) 𝑹𝒑𝒎𝒔189100190Rent per m2 (Rpms)1910.119211931019410195CCDF (%)196519730198(b) 𝑹𝒆𝒏𝒕1991002000.12011202102031000204Rent205CCDF (%)2062002073500208Figure 3: Complementary cumulative distribution functions209(CCDFs) for Rpms and Rent.210Variable211Mean Median212Min213Max214Std. dev215Dependent variables216Rpms (in e )21711.7321811.002195.3822027.802214.32222Rent (in e )223847.12224720.15225230.19 2,988.00226476.84227Explanatory variables228Area (in m2)22971.3523067.4123116.00232178.0823326.35234Rooms2352.442362.002371.002386.002390.89240Floor2412.892423.002430.0024426.002452.29246TopFloor2475.092485.002490.0025027.002512.21252CityDistance (in km)2537.992547.512550.3725617.542574.26258YearBuilt259197026019852611830262202026343.76264Table 1: Summary statistics of key variables.2653.4266Cross-correlations267Fig. 4 provides an overview of cross-correlations between the in-268dependent variables. Unsurprisingly, we see a positive correlation269between the size of an apartment and the number of rooms (cor-270relation of 0.81). Likewise, we find a positive correlation of 0.35271between Floor and TopFloor. We observe a negative correlation272of −0.14 between CityDistance and TopFloor, indicating that in273Berlin, there is a higher concentration of tall buildings in areas274close to the city center. The correlation between YearBuilt and275TopFloor (0.16) suggests that newer buildings tend to have more276floors, while the correlation between the construction year and277the distance from the city center (0.18) highlights that there has278been more construction activity near the edges of Berlin in recent279years. All remaining correlations are fairly small. Importantly, the280281282===== PAGE 5 (score 18, numeric-paren 0, table-words 2, var-words 13) =====283284Integrating Floor Plans into Hedonic Models for Rent Price Appraisal285WWW ’21, April 19–23, 2021, Ljubljana, Slovenia286variance inflation factors of all independent variables in our later287analysis are below the critical threshold of 4. Hence there is no288evidence that multicollinearity impedes the validity of our findings.2891.02900.52910.0292-1.0293-0.5294Correlation295YearBuilt296CityDistance297TopFloor298Floor299Rooms300Area301YearBuilt302CityDistance303TopFloor304Floor305Rooms306Area30713080.1830913100.16311-0.1431213130.07314-0.063150.3531613170.083180.04319-0.033200.033211322-0.01323-0.13324-0.023250.013260.813271328Figure 4: Cross-correlations.3293.5330Extraction of Floor Plans from Images331We perform three main preprocessing steps to transform the raw332floor plans into a format that allows for further calculations. First,333we manually remove erroneous entries containing images that334are not floor plans. These include interior and exterior photos,335promotional materials, and presentation slides. Second, we exclude336technically correct but unrelated floor plans. For example, if there337is a floor plan for the parking space, it is removed, so the neural338network does not learn unrelated information. These filtering steps339reduce our dataset to 9,174 observations.340Once the data set is reduced to contain only the relevant entries,341we modify them for the use in neural networks. We crop every342picture until only the plan remains. Subsequently, we scale them343with proportions preserved, filling empty space with pure black to344prevent unnecessary calculations. All floor plans are then adjusted345to follow the Xception guidelines [7] of min-max normalized color346images with 299 by 299 pixels in size. For this purpose, we use an347ImageMagick script that scales images without shifts in proportions348to avoid potential loss of information and to lighten the computa-349tional burden. The prepared images are then vectorized, min-max350normalized, and concatenated to the data frame. The process is351illustrated in Fig. 5.3524353METHODOLOGY354We propose a two-stage approach to determine the hedonic price355of apartment floor plans. As a first step, we compute adjusted rent356prices by regressing the monthly apartment costs on the structural357and locational variables (Stage 1). This allows us to extract the358variance in rent that is unexplained by these variables and prevents359our model from learning information that is already available from360known features. Second, we use convolutional neural networks361(CNN) and the adjusted rent prices to predict the sentiment of362the floor plans, i. e. the apartment rent price after controlling for363locational and structural characteristics of an apartment (Stage 2).364Figure 5: Example of cropping and scaling of floor plans.3654.1366Computing Adjusted Rent Prices (Stage 1)367Stage (1) performs a hedonic regression to calculate the variance in368the apartment costs that is unexplained by the structural and loca-369tional attributes of an apartment. This approach follows previous370works [2, 9, 10, 27], where regression residuals are used as a proxy371for the proportion of variance of the dependent variable that is un-372explained by known determinants. Let 𝑦𝑖denote the logarithmized373rent/rent per m2 belonging to listing 𝑖= 1, . . . ,𝑛. We then estimate374a linear model via ordinary least squares (OLS) regressing 𝑦𝑖on the375structural and locational attributes 𝑐𝑖1, . . . ,𝑐𝑖𝐽. Formally, the model376is described by377𝑦𝑖= 𝛽0 +378𝐽∑︁379𝑗=1380𝛽𝑗𝑐𝑖𝑗+ 𝜀𝑖,381(1)382with intercept 𝛽0 and error term 𝜀. The model residuals ˜𝑦𝑖= 𝑦𝑖−^𝑦𝑖383then represent adjusted rent prices that are used in a neural network384model in Stage (2). Importantly, the calculation of the adjusted prices385ensures that the neural network does not merely learn to predict386easily observable structural apartment characteristics.3874.2388Floor Plan Sentiment (Stage 2)389Stage (2) uses convolutional neural networks (CNN) to predict the390sentiment of the floor plans. The input to the model is a vector rep-391resentation of the preprocessed floor plan images (299×299 pixels).392The variable to predict is the adjusted rent price from Stage (1), i. e.,393˜𝑦𝑖. We implement our learning task using Xception [7] developed394by Google, which is the state-of-the-art in image classification [20].395Specifically, we employ a pretrained Xception model trained on a396large dataset of images from ImageNet [7] and employ a transfer397learning strategy (i. e., learn a new task through the transfer of398knowledge from a related task that has already been learned) to399transfer the pre-trained Xception model to our floor plan prediction400task.401Architecture customization: The pretrained Xception model402uses a softmax layer as the final layer to make categorical predic-403tions. As our goal is to use floor plan images to predict a continuous404variable (˜𝑦𝑖), we modify the architecture by removing the softmax405layer from the pre-trained model. Further, we apply the global av-406erage pooling operation on the output layer. We then append a407set of fully connected layers with ReLu activation function, fol-408lowed by an output layer consisting of a single neuron (with linear409activation function). The latter ensures that our model produces410411412===== PAGE 7 (score 23, numeric-paren 0, table-words 7, var-words 5) =====413414Integrating Floor Plans into Hedonic Models for Rent Price Appraisal415WWW ’21, April 19–23, 2021, Ljubljana, Slovenia416Control variables: Our regression model controls for 30 addi-417tional structural and locational apartment characteristics (see Fig. 1).418The corresponding estimates are omitted from Fig. 7 for the sake of419brevity. In short, more preferable contractual clauses (e.g. non-coal420heating system, the allowance of pets or recent renovation) result421in higher rent prices. We further observe location effects, i. e. city422district dummies that are statistically significant.423Goodness-of-fit: We calculated the adjusted-𝑅2 for each model,424resulting in relatively high values of 0.622 for the baseline hedonic425model without floor plans. If we additionally include floor plans, the426adjusted-𝑅2 increases to 0.697. Evidently, our model specification427accounts for a large proportion of the variations in the depen-428dent variable and FloorPlanSentiment has significant explana-429tory power. This observation is supported by the difference in the430AIC scores for the model with and without FloorPlanSentiment.431For both dependent variables, the difference is greater than 10, in-432dicating strong support for the corresponding candidate models433[5]. Therefore, the model that incorporates the floor plans is to be434preferred.435Analysis of total rent (𝑹𝒆𝒏𝒕): The estimates for the regres-436sion with the total rent (𝑅𝑒𝑛𝑡) as dependent variable are omitted437for brevity. In short, the coefficients are consistent with those438for the rent price per m2 (Rpms). Also here, the coefficient for439the FloorPlanSentiment variable is statistically significant at the4401 % statistical significance level. The regression results again sug-441gest that the sentiment of the floor plans explains a considerable442amount of the variations in rent prices after controlling for struc-443tural and locational characteristics of an apartment. The coefficient444for FloorPlanSentiment is again among the variables with the445largest effect sizes. In other words, floor plans contain price relevant446information in online real estate listings.4475.2448Prediction Performance449Next, we examine whether floor plans are useful regarding out-of-450sample prediction of rent prices. As mentioned in Sec. 2.1, hedo-451nic analysis is not limited to linear regression and can be used in452combination with machine learning methods. In the following, we453compare the predictive value of incorporating floor plans as an addi-454tional predictor for three models: log-linear OLS regression (acting455as a benchmark), boosted decision tree ensembles (CatBoost), and456deep neural networks. To evaluate the predictive value of floor457plans, we split our dataset into training and test sets and com-458pare the effect of incorporating floor plans on the out-of-sample459prediction performance of the models. Specifically, we use 5-fold460cross-validation to calculate mean squared error (MSE) and mean461absolute error (MAE).462Prediction of rent per m2: Table 2 reports the out-of-sample463prediction performance for the rent per m2 (𝑅𝑝𝑚𝑠). Incorporating464floor plans yields improvements for all considered out-of-sample465performance metrics. A log-linear hedonic regression model yields466an MSE of 0.0484 and an MAE of 0.1565. Incorporating FloorPlanSen-467timent reduces the MSE by 11.71 % and the MAE by 13.93 %. Hedonic468models trained with machine learning methods yield further im-469provements. A hedonic model trained with neural networks yields470the lowest out-of-sample prediction error (MSE of 0.0014 and MAE471of 0.0214). Also here, incorporating floor plans yields a substantially472lower error. The MSE is reduced by 10.56 %, whereas the MAE is473reduced by 4.21 %. For the decision tree ensembles (CatBoost), we474find a reduction of 4.29 % in MSE and a reduction in MAE of 1.53 %.475We also conducted paired 𝑡-tests to assess the statistical significance476of the reductions in prediction error between the machine learning477models with vs. without floor plans. We find that incorporating478floor plan yields statistically significantly reduced prediction errors479for all considered machine learning methods, i. e., log-linear regres-480sion (𝑝< 0.001), CatBoost (𝑝< 0.05), and deep neural networks481(𝑝< 0.001).482Method483Floor Plans484MSE485MAE486Benchmark: Log-Linear Regression487✗4880.049844890.1565490Hedonic models491Log-Linear Regression (with floor plans)492✓4930.044004940.1347495CatBoost496✗4970.001404980.0262499CatBoost (with floor plans)500✓5010.001345020.0258503Deep Neural Networks504✗5050.001425060.0214507Deep Neural Networks (with floor plans)508✓5090.001275100.0205511Table 2: Out-of-sample prediction performance for rent per512m2 (Rpms). The lowest values of mean squared error (MSE)513and mean absolute error (MAE) are highlighted in bold.514Prediction of total rent: Table 3 reports the performance for515the prediction of the total rent price (Rent). All considered models516show a significant reduction of the prediction error when including517floor plans. The MSE of log-linear regression, decision tree ensem-518ble, and deep neural networks reduce by 6.05 %, 4.27 %, and 5.08 %.519We also observe a lower MAE for log-linear regression (−3.92 %), de-520cision tree ensemble (−2.03 %), and deep neural networks (−1.97 %).521Analogous to the analysis of the rent per m2, paired 𝑡-tests show522that the reductions in prediction error between the models with vs.523without floor plans are statistically significant for each considered524machine learning method. Altogether, these results demonstrate525that floor plans not only feature significant explanatory power but526also serve as highly relevant features for predictive purposes.527Method528Floor Plans529MSE530MAE531Benchmark: Log-Linear Regression532✗5330.050785340.1529535Hedonic models536Log-Linear Regression (with floor plans)537✓5380.047715390.1459540CatBoost541✗5420.001175430.0246544CatBoost (with floor plans)545✓5460.001125470.0241548Deep Neural Networks549✗5500.001185510.0203552Deep Neural Networks (with floor plans)553✓5540.001125550.0199556Table 3: Out-of-sample prediction performance for rent557(Rent). The lowest values of mean squared error (MSE) and558mean absolute error (MAE) are highlighted in bold.5595.3560Sensitivity Analysis & Robustness Checks561Analysis on data subsets: We repeated our analysis on multiple562subsets of the data. While the increase in explanatory power stem-563ming from the floor plans remained significant for every tested564subset, we observed a statistically significant more pronounced565effect for smaller apartments, as well as for older houses.566567568===== PAGE 8 (score 11, numeric-paren 0, table-words 0, var-words 11) =====569570WWW ’21, April 19–23, 2021, Ljubljana, Slovenia571Kirill Solovev and Nicolas Pröllochs572Model specification: We repeated our analysis using a number573of alternative model specifications, including, but not limited to574CNN models (Xception, VGG16, ResNet101V2, DenseNet, Efficient-575Net), alternative input image formats (color, gray scale, preserved576ratios, fill methods), alternative parameter configurations of the577neural network (model concatenation, input concatenation, out-578put concatenation, depths), and alternative variable specifications579(normalization methods, log-specifications). Additionally, we tested580the inclusion of the images directly alongside the structured data581and as a standalone feature. In total, we have tested more than58250 various permutations yielding robust findings, i. e. significant583improvements in explanatory power and prediction performance584when incorporating floor plans into hedonic price models for rent585price appraisal.586Additional robustness checks: We conducted additional checks587to validate the robustness of our hedonic regression model: (1) We588calculated variance inflation factors for all independent variables589in our hedonic regression models and found that all remain be-590low the critical threshold of four. (2) We tested alternative model591specifications in which naturally correlated variables such as size592and number of rooms of an apartment are iteratively added one by593one and tested for statistical significance after each iteration (i. e.,594stepwise regression). (3) We controlled for outliers in the depen-595dent variables. (4) We added quadratic terms for each explanatory596variable to our hedonic regression models. In all cases, our results597are robust and consistently support our findings.5985.4599Exemplary Apartment Listings600We now explore apartment listings for which floor plans are par-601ticularly informative for rent price appraisal on online real estate602platforms. Fig. 8 shows four exemplary floor plans with particularly603pronounced differences between the predicted rent prices of the604hedonic model with vs. without floor plans. The floor plans (a) and605(b) yielded upward price adjustments, while the floor plans (c) and606(d) yielded downward price adjustments.607Fig. 8 suggests that incorporating floor plans yields higher rent608price predictions in cases in which floor plans convey relatively609complex information. A possible reason is that these floor plans are610particularly informative because they convey information that is611not clear from the structured data alone. For instance, floor plan612(a) in Fig. 8 has an uncluttered entrance, opening up to the entire613building and a comparatively high number of windows, which may614facilitate higher valuation. Floor plan (b) in Fig. 8 shows a spacious615apartment with a large and open-spaced living room, as well as616two separated and self-contained private sections, each outfitted617with personal bathrooms. This may be highly price-relevant infor-618mation that is only available from the apartment floor plan. On the619contrary, floor plan (c) seems to lack features that warrant a higher620rent price. The apartment it is separated into relatively small rooms621connected by a very long corridor. The only point of entrance to a622single balcony is through the room farthest away from the entrance.623Likewise, floor plan (d) has only a tiny bathroom and four rooms624that are very elongated. The apartment thus exhibits potentially625unfavorable characteristics that are not directly observable from626from the structured data. Altogether, our exploratory analysis sug-627gests that floor plans contain hidden information that is highly628price-relevant – even after accounting for structural and locational629characteristics of an apartment.6306631DISCUSSION632Research implications: Our empirical findings contribute to the633existing research body by quantifying the hedonic value of floor634plans and establishing the link between the visual design of floor635plans and real estate prices. Prior research [15, 28] suggests that636real estate markets tend to contain a variety of non-observable or637hidden characteristics that are not taken into account by conven-638tional valuation methods. Our findings show that there is indeed639an underutilization of the available data in current hedonic models.640Specifically, we show that floor plans contain price relevant infor-641mation in online real estate listings. This suggests that there are642hidden features in floor plans, such as the relative size and position-643ing between the rooms, that are relevant even after controlling for644structural and locational characteristics. To the best of our knowl-645edge, our paper is the first study to demonstrate that harnessing646floor plans can enhance real estate appraisal on online real estate647platforms.648Practical implications: From a practical perspective, our find-649ings are particularly relevant for online real estate platforms. Cur-650rently, the decision of whether or not to upload a floor plan for651an apartment is typically left to the discretion of the user. Based652on our finding that floor plans contain price-relevant information653that helps users to make an informed decision, real estate platforms654should consider making them a mandatory part of the listings and655show them more prominently. This would allow for greater mar-656ket transparency for both renters and landlords, as well as home657buyers and sellers. In the backend of online real estate platforms,658floor plans could also be used as an additional price predictor. For659example, online real estate platforms could implement a model660similar to ours after the author has filled all fields to recommend an661appropriate rent price. Alternatively, the proposed prediction could662be listed alongside the price set by the author, such that the renter663or home buyer has a clearer picture of the value of the property.664Our model could also suggest links to similar lots based not only on665the price but on the information from the floor plans, providing a666better market overview and diminishing the problem of information667asymmetry. While some valuation projects, such as Zillow, have668started taking images into account, floor plans remain severely un-669derutilized. Ultimately, our study equips real estate investors with670the clear benefit of improved risk assessment, allowing them to671enhance their portfolios and reduce the possibility of misvaluation,672which indirectly improves the quality of financial markets [16].673Limitations and future research directions: Our work pro-674vides several avenues for future research that could further enhance675the accuracy of recommendation systems on online real estate plat-676forms. First, it would be interesting to extend our hedonic model by677performing textual analysis (e. g., sentiment analysis [22, 23, 31])678of apartment descriptions in real estate listings. Second, while our679study was conducted on the rental market of Berlin, the underlying680deep learning approach can easily be applied to listings from other681marketplaces and locations. Third, future research could extend682our study by studying the hedonic value of different floor plan683