# RECORD 1462a807ace2b4c97b7b8368ffbe3a1c694cd0f7
# TITLE: Integrating Floor Plans into Hedonic Models for Rent Price Appraisal
# YEAR: 2021.0  VENUE: The Web Conference  DOI: 10.1145/3442381.3449967
# AUTHORS: K. Solovev; Nicolas Pröllochs
# PAGES: 10  CANDIDATE TABLE PAGES (1-based): [4, 7]
# 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.


===== PAGE 1 (score 12, numeric-paren 0, table-words 0, var-words 12) =====

Integrating Floor Plans into Hedonic Models for
Rent Price Appraisal
Kirill Solovev
kirill.solovev@wirtschaft.uni-giessen.de
University of Giessen
Germany
Nicolas Pröllochs
nicolas.proellochs@wi.jlug.de
University of Giessen
Germany
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 ap-
proach 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. Alto-
gether, 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.
CCS CONCEPTS
• Applied computing →Electronic commerce; Economics; •
Information systems →Decision support systems; • Computing
methodologies →Modeling methodologies; Neural networks.
KEYWORDS
Online real estate platforms, hedonic price models, floor plans,
visual analytics, image sentiment
ACM Reference Format:
Kirill Solovev and Nicolas Pröllochs. 2021. Integrating Floor Plans into He-
donic Models for Rent Price Appraisal. In Proceedings of the Web Conference
This paper is published under the Creative Commons Attribution 4.0 International
(CC-BY 4.0) license. Authors reserve their rights to disseminate the work on their
personal and corporate Web sites with the appropriate attribution.
WWW ’21, April 19–23, 2021, Ljubljana, Slovenia
© 2021 IW3C2 (International World Wide Web Conference Committee), published
under Creative Commons CC-BY 4.0 License.
ACM ISBN 978-1-4503-8312-7/21/04.
https://doi.org/10.1145/3442381.3449967
2021 (WWW ’21), April 19–23, 2021,Ljubljana, Slovenia. ACM, New York, NY,
USA, 10 pages. https://doi.org/10.1145/3442381.3449967
1
INTRODUCTION
Online real estate platforms have become significant marketplaces
for the real estate industry. These online platforms provide an
overview of available properties in a specific location and facilitate
users’ search for an apartment or a house [39]. In 2011, one of the
biggest online real estate providers, Zillow, reported 24 million
unique visitors. Today, this number has risen to 196 million and
stays on the upwards trend, indicating continual interest from
prospective home buyers and tenants [40]. While the number of
people referring to online real-estate platforms continues to grow,
there is a rising need for tools and recommendation algorithms that
allow for an accurate appraisal of real estate prices [e. g. 39].
Prior works have primarily studied real estate valuation based on
hedonic price models [25, 34, 37]. Hedonic pricing theory suggests
that the valuation of housing values or rents can be viewed as a
weighted sum of its features [37]. Specifically, the price and rent
are determined by the attributes and characteristics of the dwelling
unit and the surrounding neighborhood. Hedonic models provide a
high degree of flexibility when selecting the attributes, which can
roughly be divided into structural and locational attributes [26].
The structural factors describe an apartment itself (e. g. number
of rooms, amenities, parking, pool), while the locational attributes
are composed of external features affecting the price. For instance,
previous studies [e. g. 30, 33] have examined the impact of the
distance to transport hubs and shopping centers on real estate
prices. Altogether, the hedonic approach provides scholars and
practitioners with a framework of assessing the value of real estate
properties, where the final price entails a precise valuation of a
diverse feature package.
While structural and locational attributes are of great importance
for real estate price models, modern online real estate marketplaces
provide additional information that has been neglected in previ-
ous works. A particularly relevant feature for housebuyers and
prospective tenants may be real estate floor plans. According to
rightmove.co.uk, 90 % of home-buyers think that a floor plan is an
essential part of the decision making process when finding an apart-
ment or house. A real estate floor plan is a schematic illustration
that provides a top-down view of the property. While information
about floor plans may be partially provided in structured form (e. g.,
in floor plan filings), the actual floor plan images may contain price-
relevant hidden information. For example, the relationship between
rooms and spaces, as well as the overall layout are typically only
available from the floor plan image itself. Information on floor plans
is typically also not available in textual form as the vast majority
arXiv:2102.08162v1  [cs.LG]  16 Feb 2021


===== PAGE 4 (score 25, numeric-paren 2, table-words 2, var-words 15) =====

WWW ’21, April 19–23, 2021, Ljubljana, Slovenia
Kirill Solovev and Nicolas Pröllochs
in hedonic pricing models – even after controlling for structured
apartment characteristics such as size and location.
3
DATA
3.1
Apartment Listings
The data for this study was extracted from ImmobilienScout24.de,
the largest German online real estate aggregator. At the end of
2019, ImmobilienScout24 contained 91,415 active rent listings across
Germany, with approximately 44 million visitors per month.
We implemented a Python-based web crawler to download and
store information about apartments in the city of Berlin in a data-
base format. The crawler was active for three weeks at the end
of 2019, during which it gathered a total number of 15,604 obser-
vations. Each observation represents an apartment in Berlin from
an active or archived listing on ImmobilienScout24. Our dataset
includes apartment listings between mid-2017 and end of 2019 (i. e.,
spans two and a half years). For each apartment, the crawler col-
lected a raw image showing the floor plan of the apartment, the
monthly rent price, the monthly rent price per m2, the city district
in which the apartment is located, and 36 fields with categorical
and numerical information about the apartment (e. g., apartment
size, etc.)
3.2
Variable Definitions
Dependent variables: The target variables in our study are two
measures of the monthly costs of an apartment:
• Rpms: The monthly rent price (in e) per m2
• Rent: The total monthly rent price (in e) of an apartment
Explanatory variables: Our hedonic rent price model accounts
for the following apartment characteristics from previous works:
• Area: The total size of the apartment in m2
• Rooms: The total number of rooms in the apartment
• Floor: The floor at which the apartment is located
• TopFloor: The number of floors of the building in which
the apartment is located
• CityDistance: The distance (in km) from the center of the
city to the apartment
• YearBuilt: The construction year of the building in which
the apartment is located
Control variables: Our dataset further includes 12 locational
dummies that provide information about the city district, and a
set of 30 additional control variables in the form of categorical
dummies that provide fine-grained information about contractual
and structural characteristics of an apartment (e. g. whether the
apartment offers access to a parking lot).
3.3
Summary Statistics
Dependent variables: Table 1 shows summary statistics, whereas
Fig. 3 plots the complementary cumulative distributions (CCDFs)
for the dependent variables. The total rent prices (Rent) in our
data range from e230 per month to almost e3000 with a standard
deviation of 476.80. The rent per m2 (Rpms) ranges from 5.38 to
27.80 with a standard deviation of 4.32. Fig. 1 provides an overview
of the monthly apartment costs across different districts in Berlin.
Evidently, there is a higher concentration of apartments with high
monthly costs near the city center of Berlin.
Explanatory variables: Fig. 2 plots the distributions of the
explanatory variables in our dataset. The size of the apartments
(Area) ranges from 16.00 to 178.08 square meters with a standard
deviation of 26.35. The number of rooms ranges from 1 to 6, with
most of the apartments having either 2 or 3 rooms. The average
distance between an apartment and the city center (CityDistance)
is 7.99 km. The variables TopFloor and Floor have the same mini-
mum of 0 and a maximum of 27 and 26, respectively. The houses in
our sample have been built between the years 1830 and 2020. Most
buildings are relatively new with a median construction year of
1985.
(a) 𝑹𝒑𝒎𝒔
100
Rent per m2 (Rpms)
0.1
1
10
10
CCDF (%)
5
30
(b) 𝑹𝒆𝒏𝒕
100
0.1
1
10
1000
Rent
CCDF (%)
200
3500
Figure 3: Complementary cumulative distribution functions
(CCDFs) for Rpms and Rent.
Variable
Mean Median
Min
Max
Std. dev
Dependent variables
Rpms (in e )
11.73
11.00
5.38
27.80
4.32
Rent (in e )
847.12
720.15
230.19 2,988.00
476.84
Explanatory variables
Area (in m2)
71.35
67.41
16.00
178.08
26.35
Rooms
2.44
2.00
1.00
6.00
0.89
Floor
2.89
3.00
0.00
26.00
2.29
TopFloor
5.09
5.00
0.00
27.00
2.21
CityDistance (in km)
7.99
7.51
0.37
17.54
4.26
YearBuilt
1970
1985
1830
2020
43.76
Table 1: Summary statistics of key variables.
3.4
Cross-correlations
Fig. 4 provides an overview of cross-correlations between the in-
dependent variables. Unsurprisingly, we see a positive correlation
between the size of an apartment and the number of rooms (cor-
relation of 0.81). Likewise, we find a positive correlation of 0.35
between Floor and TopFloor. We observe a negative correlation
of −0.14 between CityDistance and TopFloor, indicating that in
Berlin, there is a higher concentration of tall buildings in areas
close to the city center. The correlation between YearBuilt and
TopFloor (0.16) suggests that newer buildings tend to have more
floors, while the correlation between the construction year and
the distance from the city center (0.18) highlights that there has
been more construction activity near the edges of Berlin in recent
years. All remaining correlations are fairly small. Importantly, the


===== PAGE 5 (score 18, numeric-paren 0, table-words 2, var-words 13) =====

Integrating Floor Plans into Hedonic Models for Rent Price Appraisal
WWW ’21, April 19–23, 2021, Ljubljana, Slovenia
variance inflation factors of all independent variables in our later
analysis are below the critical threshold of 4. Hence there is no
evidence that multicollinearity impedes the validity of our findings.
1.0
0.5
0.0
-1.0
-0.5
Correlation
YearBuilt
CityDistance
TopFloor
Floor
Rooms
Area
YearBuilt
CityDistance
TopFloor
Floor
Rooms
Area
1
0.18
1
0.16
-0.14
1
0.07
-0.06
0.35
1
0.08
0.04
-0.03
0.03
1
-0.01
-0.13
-0.02
0.01
0.81
1
Figure 4: Cross-correlations.
3.5
Extraction of Floor Plans from Images
We perform three main preprocessing steps to transform the raw
floor plans into a format that allows for further calculations. First,
we manually remove erroneous entries containing images that
are not floor plans. These include interior and exterior photos,
promotional materials, and presentation slides. Second, we exclude
technically correct but unrelated floor plans. For example, if there
is a floor plan for the parking space, it is removed, so the neural
network does not learn unrelated information. These filtering steps
reduce our dataset to 9,174 observations.
Once the data set is reduced to contain only the relevant entries,
we modify them for the use in neural networks. We crop every
picture until only the plan remains. Subsequently, we scale them
with proportions preserved, filling empty space with pure black to
prevent unnecessary calculations. All floor plans are then adjusted
to follow the Xception guidelines [7] of min-max normalized color
images with 299 by 299 pixels in size. For this purpose, we use an
ImageMagick script that scales images without shifts in proportions
to avoid potential loss of information and to lighten the computa-
tional burden. The prepared images are then vectorized, min-max
normalized, and concatenated to the data frame. The process is
illustrated in Fig. 5.
4
METHODOLOGY
We propose a two-stage approach to determine the hedonic price
of apartment floor plans. As a first step, we compute adjusted rent
prices by regressing the monthly apartment costs on the structural
and locational variables (Stage 1). This allows us to extract the
variance in rent that is unexplained by these variables and prevents
our model from learning information that is already available from
known features. Second, we use convolutional neural networks
(CNN) and the adjusted rent prices to predict the sentiment of
the floor plans, i. e. the apartment rent price after controlling for
locational and structural characteristics of an apartment (Stage 2).
Figure 5: Example of cropping and scaling of floor plans.
4.1
Computing Adjusted Rent Prices (Stage 1)
Stage (1) performs a hedonic regression to calculate the variance in
the apartment costs that is unexplained by the structural and loca-
tional attributes of an apartment. This approach follows previous
works [2, 9, 10, 27], where regression residuals are used as a proxy
for the proportion of variance of the dependent variable that is un-
explained by known determinants. Let 𝑦𝑖denote the logarithmized
rent/rent per m2 belonging to listing 𝑖= 1, . . . ,𝑛. We then estimate
a linear model via ordinary least squares (OLS) regressing 𝑦𝑖on the
structural and locational attributes 𝑐𝑖1, . . . ,𝑐𝑖𝐽. Formally, the model
is described by
𝑦𝑖= 𝛽0 +
𝐽∑︁
𝑗=1
𝛽𝑗𝑐𝑖𝑗+ 𝜀𝑖,
(1)
with intercept 𝛽0 and error term 𝜀. The model residuals ˜𝑦𝑖= 𝑦𝑖−^𝑦𝑖
then represent adjusted rent prices that are used in a neural network
model in Stage (2). Importantly, the calculation of the adjusted prices
ensures that the neural network does not merely learn to predict
easily observable structural apartment characteristics.
4.2
Floor Plan Sentiment (Stage 2)
Stage (2) uses convolutional neural networks (CNN) to predict the
sentiment of the floor plans. The input to the model is a vector rep-
resentation of the preprocessed floor plan images (299×299 pixels).
The variable to predict is the adjusted rent price from Stage (1), i. e.,
˜𝑦𝑖. We implement our learning task using Xception [7] developed
by Google, which is the state-of-the-art in image classification [20].
Specifically, we employ a pretrained Xception model trained on a
large dataset of images from ImageNet [7] and employ a transfer
learning strategy (i. e., learn a new task through the transfer of
knowledge from a related task that has already been learned) to
transfer the pre-trained Xception model to our floor plan prediction
task.
Architecture customization: The pretrained Xception model
uses a softmax layer as the final layer to make categorical predic-
tions. As our goal is to use floor plan images to predict a continuous
variable (˜𝑦𝑖), we modify the architecture by removing the softmax
layer from the pre-trained model. Further, we apply the global av-
erage pooling operation on the output layer. We then append a
set of fully connected layers with ReLu activation function, fol-
lowed by an output layer consisting of a single neuron (with linear
activation function). The latter ensures that our model produces


===== PAGE 7 (score 23, numeric-paren 0, table-words 7, var-words 5) =====

Integrating Floor Plans into Hedonic Models for Rent Price Appraisal
WWW ’21, April 19–23, 2021, Ljubljana, Slovenia
Control variables: Our regression model controls for 30 addi-
tional structural and locational apartment characteristics (see Fig. 1).
The corresponding estimates are omitted from Fig. 7 for the sake of
brevity. In short, more preferable contractual clauses (e.g. non-coal
heating system, the allowance of pets or recent renovation) result
in higher rent prices. We further observe location effects, i. e. city
district dummies that are statistically significant.
Goodness-of-fit: We calculated the adjusted-𝑅2 for each model,
resulting in relatively high values of 0.622 for the baseline hedonic
model without floor plans. If we additionally include floor plans, the
adjusted-𝑅2 increases to 0.697. Evidently, our model specification
accounts for a large proportion of the variations in the depen-
dent variable and FloorPlanSentiment has significant explana-
tory power. This observation is supported by the difference in the
AIC scores for the model with and without FloorPlanSentiment.
For both dependent variables, the difference is greater than 10, in-
dicating strong support for the corresponding candidate models
[5]. Therefore, the model that incorporates the floor plans is to be
preferred.
Analysis of total rent (𝑹𝒆𝒏𝒕): The estimates for the regres-
sion with the total rent (𝑅𝑒𝑛𝑡) as dependent variable are omitted
for brevity. In short, the coefficients are consistent with those
for the rent price per m2 (Rpms). Also here, the coefficient for
the FloorPlanSentiment variable is statistically significant at the
1 % statistical significance level. The regression results again sug-
gest that the sentiment of the floor plans explains a considerable
amount of the variations in rent prices after controlling for struc-
tural and locational characteristics of an apartment. The coefficient
for FloorPlanSentiment is again among the variables with the
largest effect sizes. In other words, floor plans contain price relevant
information in online real estate listings.
5.2
Prediction Performance
Next, we examine whether floor plans are useful regarding out-of-
sample prediction of rent prices. As mentioned in Sec. 2.1, hedo-
nic analysis is not limited to linear regression and can be used in
combination with machine learning methods. In the following, we
compare the predictive value of incorporating floor plans as an addi-
tional predictor for three models: log-linear OLS regression (acting
as a benchmark), boosted decision tree ensembles (CatBoost), and
deep neural networks. To evaluate the predictive value of floor
plans, we split our dataset into training and test sets and com-
pare the effect of incorporating floor plans on the out-of-sample
prediction performance of the models. Specifically, we use 5-fold
cross-validation to calculate mean squared error (MSE) and mean
absolute error (MAE).
Prediction of rent per m2: Table 2 reports the out-of-sample
prediction performance for the rent per m2 (𝑅𝑝𝑚𝑠). Incorporating
floor plans yields improvements for all considered out-of-sample
performance metrics. A log-linear hedonic regression model yields
an MSE of 0.0484 and an MAE of 0.1565. Incorporating FloorPlanSen-
timent reduces the MSE by 11.71 % and the MAE by 13.93 %. Hedonic
models trained with machine learning methods yield further im-
provements. A hedonic model trained with neural networks yields
the lowest out-of-sample prediction error (MSE of 0.0014 and MAE
of 0.0214). Also here, incorporating floor plans yields a substantially
lower error. The MSE is reduced by 10.56 %, whereas the MAE is
reduced by 4.21 %. For the decision tree ensembles (CatBoost), we
find a reduction of 4.29 % in MSE and a reduction in MAE of 1.53 %.
We also conducted paired 𝑡-tests to assess the statistical significance
of the reductions in prediction error between the machine learning
models with vs. without floor plans. We find that incorporating
floor plan yields statistically significantly reduced prediction errors
for all considered machine learning methods, i. e., log-linear regres-
sion (𝑝< 0.001), CatBoost (𝑝< 0.05), and deep neural networks
(𝑝< 0.001).
Method
Floor Plans
MSE
MAE
Benchmark: Log-Linear Regression
✗
0.04984
0.1565
Hedonic models
Log-Linear Regression (with floor plans)
✓
0.04400
0.1347
CatBoost
✗
0.00140
0.0262
CatBoost (with floor plans)
✓
0.00134
0.0258
Deep Neural Networks
✗
0.00142
0.0214
Deep Neural Networks (with floor plans)
✓
0.00127
0.0205
Table 2: Out-of-sample prediction performance for rent per
m2 (Rpms). The lowest values of mean squared error (MSE)
and mean absolute error (MAE) are highlighted in bold.
Prediction of total rent: Table 3 reports the performance for
the prediction of the total rent price (Rent). All considered models
show a significant reduction of the prediction error when including
floor plans. The MSE of log-linear regression, decision tree ensem-
ble, and deep neural networks reduce by 6.05 %, 4.27 %, and 5.08 %.
We also observe a lower MAE for log-linear regression (−3.92 %), de-
cision tree ensemble (−2.03 %), and deep neural networks (−1.97 %).
Analogous to the analysis of the rent per m2, paired 𝑡-tests show
that the reductions in prediction error between the models with vs.
without floor plans are statistically significant for each considered
machine learning method. Altogether, these results demonstrate
that floor plans not only feature significant explanatory power but
also serve as highly relevant features for predictive purposes.
Method
Floor Plans
MSE
MAE
Benchmark: Log-Linear Regression
✗
0.05078
0.1529
Hedonic models
Log-Linear Regression (with floor plans)
✓
0.04771
0.1459
CatBoost
✗
0.00117
0.0246
CatBoost (with floor plans)
✓
0.00112
0.0241
Deep Neural Networks
✗
0.00118
0.0203
Deep Neural Networks (with floor plans)
✓
0.00112
0.0199
Table 3: Out-of-sample prediction performance for rent
(Rent). The lowest values of mean squared error (MSE) and
mean absolute error (MAE) are highlighted in bold.
5.3
Sensitivity Analysis & Robustness Checks
Analysis on data subsets: We repeated our analysis on multiple
subsets of the data. While the increase in explanatory power stem-
ming from the floor plans remained significant for every tested
subset, we observed a statistically significant more pronounced
effect for smaller apartments, as well as for older houses.


===== PAGE 8 (score 11, numeric-paren 0, table-words 0, var-words 11) =====

WWW ’21, April 19–23, 2021, Ljubljana, Slovenia
Kirill Solovev and Nicolas Pröllochs
Model specification: We repeated our analysis using a number
of alternative model specifications, including, but not limited to
CNN models (Xception, VGG16, ResNet101V2, DenseNet, Efficient-
Net), alternative input image formats (color, gray scale, preserved
ratios, fill methods), alternative parameter configurations of the
neural network (model concatenation, input concatenation, out-
put concatenation, depths), and alternative variable specifications
(normalization methods, log-specifications). Additionally, we tested
the inclusion of the images directly alongside the structured data
and as a standalone feature. In total, we have tested more than
50 various permutations yielding robust findings, i. e. significant
improvements in explanatory power and prediction performance
when incorporating floor plans into hedonic price models for rent
price appraisal.
Additional robustness checks: We conducted additional checks
to validate the robustness of our hedonic regression model: (1) We
calculated variance inflation factors for all independent variables
in our hedonic regression models and found that all remain be-
low the critical threshold of four. (2) We tested alternative model
specifications in which naturally correlated variables such as size
and number of rooms of an apartment are iteratively added one by
one and tested for statistical significance after each iteration (i. e.,
stepwise regression). (3) We controlled for outliers in the depen-
dent variables. (4) We added quadratic terms for each explanatory
variable to our hedonic regression models. In all cases, our results
are robust and consistently support our findings.
5.4
Exemplary Apartment Listings
We now explore apartment listings for which floor plans are par-
ticularly informative for rent price appraisal on online real estate
platforms. Fig. 8 shows four exemplary floor plans with particularly
pronounced differences between the predicted rent prices of the
hedonic model with vs. without floor plans. The floor plans (a) and
(b) yielded upward price adjustments, while the floor plans (c) and
(d) yielded downward price adjustments.
Fig. 8 suggests that incorporating floor plans yields higher rent
price predictions in cases in which floor plans convey relatively
complex information. A possible reason is that these floor plans are
particularly informative because they convey information that is
not clear from the structured data alone. For instance, floor plan
(a) in Fig. 8 has an uncluttered entrance, opening up to the entire
building and a comparatively high number of windows, which may
facilitate higher valuation. Floor plan (b) in Fig. 8 shows a spacious
apartment with a large and open-spaced living room, as well as
two separated and self-contained private sections, each outfitted
with personal bathrooms. This may be highly price-relevant infor-
mation that is only available from the apartment floor plan. On the
contrary, floor plan (c) seems to lack features that warrant a higher
rent price. The apartment it is separated into relatively small rooms
connected by a very long corridor. The only point of entrance to a
single balcony is through the room farthest away from the entrance.
Likewise, floor plan (d) has only a tiny bathroom and four rooms
that are very elongated. The apartment thus exhibits potentially
unfavorable characteristics that are not directly observable from
from the structured data. Altogether, our exploratory analysis sug-
gests that floor plans contain hidden information that is highly
price-relevant – even after accounting for structural and locational
characteristics of an apartment.
6
DISCUSSION
Research implications: Our empirical findings contribute to the
existing research body by quantifying the hedonic value of floor
plans and establishing the link between the visual design of floor
plans and real estate prices. Prior research [15, 28] suggests that
real estate markets tend to contain a variety of non-observable or
hidden characteristics that are not taken into account by conven-
tional valuation methods. Our findings show that there is indeed
an underutilization of the available data in current hedonic models.
Specifically, we show that floor plans contain price relevant infor-
mation in online real estate listings. This suggests that there are
hidden features in floor plans, such as the relative size and position-
ing between the rooms, that are relevant even after controlling for
structural and locational characteristics. To the best of our knowl-
edge, our paper is the first study to demonstrate that harnessing
floor plans can enhance real estate appraisal on online real estate
platforms.
Practical implications: From a practical perspective, our find-
ings are particularly relevant for online real estate platforms. Cur-
rently, the decision of whether or not to upload a floor plan for
an apartment is typically left to the discretion of the user. Based
on our finding that floor plans contain price-relevant information
that helps users to make an informed decision, real estate platforms
should consider making them a mandatory part of the listings and
show them more prominently. This would allow for greater mar-
ket transparency for both renters and landlords, as well as home
buyers and sellers. In the backend of online real estate platforms,
floor plans could also be used as an additional price predictor. For
example, online real estate platforms could implement a model
similar to ours after the author has filled all fields to recommend an
appropriate rent price. Alternatively, the proposed prediction could
be listed alongside the price set by the author, such that the renter
or home buyer has a clearer picture of the value of the property.
Our model could also suggest links to similar lots based not only on
the price but on the information from the floor plans, providing a
better market overview and diminishing the problem of information
asymmetry. While some valuation projects, such as Zillow, have
started taking images into account, floor plans remain severely un-
derutilized. Ultimately, our study equips real estate investors with
the clear benefit of improved risk assessment, allowing them to
enhance their portfolios and reduce the possibility of misvaluation,
which indirectly improves the quality of financial markets [16].
Limitations and future research directions: Our work pro-
vides several avenues for future research that could further enhance
the accuracy of recommendation systems on online real estate plat-
forms. First, it would be interesting to extend our hedonic model by
performing textual analysis (e. g., sentiment analysis [22, 23, 31])
of apartment descriptions in real estate listings. Second, while our
study was conducted on the rental market of Berlin, the underlying
deep learning approach can easily be applied to listings from other
marketplaces and locations. Third, future research could extend
our study by studying the hedonic value of different floor plan