# RECORD 8d5f6b09f9058fc5a131a22d58a856d18ce849c7
# TITLE: Information value of property description: A Machine learning approach
# YEAR: 2021.0  VENUE: Journal of Urban Economics  DOI: 10.1016/j.jue.2020.103299
# AUTHORS: Lily Shen; Stephen L. Ross
# PAGES: 42  CANDIDATE TABLE PAGES (1-based): [38, 39, 40, 42]
# ABSTRACT: Abstract This paper employs machine learning to quantify the value of “soft” information contained in real estate property descriptions. Textual descriptions contain information that traditional hedonic attributes cannot capture. A one standard deviation increase in the uniqueness of a property based on this “soft” information leads to a 15% increase in property sale price in a hedonic price model and a 10% increase in a repeat sales price model. The effects in the hedonic model appear to arise through two channels: the unobserved quality of the housing unit, and the market power of the housing unit relative to competing properties. The effects in the repeat sales model appear to be driven entirely by the market power of the unit. Further, an annual hedonic price index ignoring our measure of unobserved quality overstates real estate prices by between 10% to 23% and mistimes the stabilization of housing prices following the Great Recession. Similar, but smaller effects, are observed for the repeat sales price index.


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

Department of Economics Working Paper Series 
365 Fairfield Way, Unit 1063
Storrs, CT 06269-1063
Phone: (860) 486-3022
Fax: (860) 486-4463
http://www.econ.uconn.edu/
This working paper is indexed in RePEc, http://repec.org 
Information Value of Property Description: 
A Machine Learning Approach
by 
Lily Shen
Clemson University
Stephen Ross 
University of Connecticut
Working Paper 2019-20R 
December 2019, Revised September 2020


===== PAGE 38 (score 92, numeric-paren 78, table-words 5, var-words 4) =====

Table 3: Compare Covariates with and without Unique Measures
Dependent Variable: Ln(price)
Full Sample
Repeat Sale
(1)
(2)
(3)
(4)
MLS Area-Year Unique Score (standard)
0.149***
0.099***
(0.034)
(0.020)
Age
-0.013***
-0.012***
-0.014***
-0.014***
(0.001)
(0.001)
(0.002)
(0.002)
AgexAge
0.0001***
0.0001***
0.0001***
0.0001***
(0.0000)
(0.0000)
(0.0000)
(0.0000)
Bed
-0.047**
-0.042**
-0.014
-0.013
(0.021)
(0.020)
(0.021)
(0.021)
Bath
0.247***
0.245***
0.320***
0.315***
(0.018)
(0.016)
(0.035)
(0.033)
Ranch
-0.260***
-0.244***
-0.036
-0.038
(0.021)
(0.020)
(0.042)
(0.041)
Renovated
0.253***
0.225***
0.359***
0.345***
(0.032)
(0.027)
(0.042)
(0.038)
Sqft
0.009***
0.009***
-0.001
-0.001
(0.001)
(0.001)
(0.001)
(0.001)
Large Lot
0.012
0.004
0.066
0.077
(0.028)
(0.025)
(0.070)
(0.073)
Pool
0.049***
0.051***
0.026
0.010
(0.018)
(0.016)
(0.056)
(0.057)
Feature
0.126***
0.090***
0.023
-0.004
(0.031)
(0.030)
(0.083)
(0.078)
Constant
10.338***
8.610***
12.246***
8.696***
(0.081)
(0.359)
(0.265)
(0.521)
Observations
40,851
40,851
7,317
7,317
R-squared
0.760
0.769
0.956
0.958
House Characteristics
Yes
Yes
Yes
Yes
LocationxYear FE
Yes
Yes
Yes
Yes
Property FE
No
No
Yes
Yes
Cluster
Area-Year
Area-Year
Area-Year
Area-Year
Note: The results displayed in columns (1) and (2) are using the full sample, and results displayed in
columns (3) and (4) are based on the repeat sale sample.
Columns (2) and (4) include the control for
property uniqueness. The unique score coefficients are standardized to show the effect of the increase in one
standard deviation. All models include MLS Area by Year fixed effects. Robust standard errors are clustered
at the MLS Area by Year level, shown in parentheses (*** p<0.01, ** p<0.05, * p<0.1).
36


===== PAGE 39 (score 34, numeric-paren 17, table-words 8, var-words 1) =====

Table 4: House Uniqueness vs. Market Power
Full Sample
Repeat Sale
Dependent Variables: Ln(Price)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
MLS Area-Year Unique Score (standard)
0.149***
0.099***
(0.034)
(0.020)
MLS Area Competitor Unique Score (standard)
0.125***
0.069
0.090***
0.084*
(0.032)
(0.071)
(0.020)
(0.047)
MLS Area Non-Competitor Unique Score (standard)
0.112***
0.059
0.068***
0.006
(0.028)
(0.061)
(0.021)
(0.048)
Observations
40,851
40,851
40,851
40,851
7,317
7,317
7,317
7,317
R-squared
0.769
0.767
0.767
0.767
0.958
0.958
0.957
0.957
House Characteristics
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
LocationxYear FE
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Property FE
No
No
No
No
Yes
Yes
Yes
Yes
Cluster
Area-Year
Area-Year
Area-Year
Area-Year
Area-Year
Area-Year
Area-Year
Area-Year
Note: The dependent variable this table is Ln(sale price). Columns (1) and (5) repeat the estimates from the previous table for the hedonic and
repeat sales models respectively. Columns (2) and (6) present estimates using the new MLS area measure of uniqueness based on competing units,
and Columns (3) and (7) present estimates using uniqueness based on non-competing units. Finally, in Columns (4) and (8), we run a horse race
between the uniqueness measure based on competing units and based on non-competing units. All of the unique score coefficients are standardized
to show the effect of the increase in one standard deviation. All models include MLS area by year fixed effects. Robust standard errors are clustered
at the MLS Area by Year level, shown in parentheses (*** p<0.01, ** p<0.05, * p<0.1).
37


===== PAGE 40 (score 44, numeric-paren 18, table-words 12, var-words 2) =====

Table 5: Are the Uniqueness Effects Sensitive to Alternative Models?
Panel A: Full Sample Analysis
Dependent Variable:
Ln(Price)
Price Spread
(1)
(2)
(3)
(4)
(5)
MLS Area-Year Unique Score (standard)
0.149***
0.149***
0.136***
0.122***
-0.761
(0.034)
(0.032)
(0.026)
(0.019)
(0.616)
Observations
40,851
40,851
40,851
40,851
40,851
R-squared
0.769
0.784
0.811
0.886
0.238
House Characteristics
Yes
Yes
Yes
Yes
Yes
Keywords
No
No
Yes
Yes
No
Transaction Characteristics
No
Yes
Yes
Yes
No
LocationxYear FE
Yes
Yes
Yes
Yes
Yes
Agent FE
No
No
No
Yes
No
Property FE
No
No
No
No
No
Cluster
Area-Year
Area-Year
Area-Year
Area-Year
Area-Year
Panel B: Repeat Sale Analysis
Dependent Variable:
Ln(Price)
Price Spread
(1)
(2)
(3)
(4)
(5)
MLS Area-Year Unique Score (standard)
0.099***
0.098***
0.080***
0.050***
-0.297
(0.020)
(0.020)
(0.015)
(0.018)
(0.592)
Observations
7,317
7,317
7,317
4,117
7,317
R-squared
0.958
0.961
0.968
0.989
0.684
House Characteristics
Yes
Yes
Yes
Yes
Yes
Keywords
No
No
Yes
Yes
No
Transaction Characteristics
No
Yes
Yes
Yes
No
LocationxYear FE
Yes
Yes
Yes
Yes
Yes
Agent FE
No
No
No
Yes
No
Property FE
Yes
Yes
Yes
Yes
Yes
Cluster
Area-Year
Area-Year
Area-Year
Area-Year
Area-Year
Note: Panel A shows hedonic estimates using the full sample, and Penal B displays repeat sale results based
on the repeat sales sample. The dependent variable in all specifications in Column (1)–(4) is Ln(sale price),
and the dependent variable in Column (5) is Listing-Sale Spread in thousand dollars. Column (1) presents
the standard hedonic model with just the uniqueness control added, Column (2) presents estimates for a
model that also adds transaction characteristics, Column (3) also adds the dummy variables for the presence
of specific keywords, and Column (4) adds agent fixed effects. Column (5) presents the model on the listing-
sales price spread. The unique score coefficients are standardized to show the effect of the increase in one
standard deviation. Robust standard errors are clustered at the Area-Year level, shown in parentheses (***
p<0.01, ** p<0.05, * p<0.1).
38


===== PAGE 41 (score 13, numeric-paren 0, table-words 4, var-words 5) =====

Table 6: Summary Statistics of Hedonic Controls by Year
Year:
2010
2011
2012
2013
VARIABLES
Mean
Std. Dev.
Mean
Std. Dev.
Mean
Std. Dev.
Mean
Std. Dev.
MLS Area-Year Unique Score
0.75
0.07
0.78
0.06
0.78
0.05
0.78
0.05
MLS Area-Year Unique Score (Mean Diff)
0.0013
0.0269
0.0034
0.0262
-0.0003
0.0262
-0.0011
0.0242
Ln(Sale Price)
11.84
1.39
11.85
1.44
12.01
1.32
12.34
1.17
Age
42.62
30.32
42.85
30.20
45.03
30.09
45.33
30.02
Fireplace (d)
0.99
1.05
1.03
1.06
1.05
1.05
1.07
0.99
Pool (d)
0.05
0.21
0.06
0.23
0.05
0.22
0.05
0.23
Bed
3.61
1.02
3.62
1.04
3.61
1.03
3.63
1.06
Ranch (d)
0.47
0.50
0.44
0.50
0.43
0.50
0.42
0.49
Bath
2.64
1.30
2.70
1.33
2.72
1.29
2.75
1.26
Renovated (d)
0.00
0.01
0.03
0.18
0.06
0.24
0.08
0.27
Sqft (100)
20.95
11.61
22.56
12.99
22.95
13.71
23.44
13.58
Year:
2014
2015
2016
VARIABLES
Mean
Std. Dev.
Mean
Std. Dev.
Mean
Std. Dev.
MLS Area-Year Unique Score
0.79
0.04
0.78
0.05
0.78
0.06
MLS Area-Year Unique Score (Mean Diff)
-0.0021
0.0254
-0.0016
0.0267
-0.0008
0.0272
Ln(Sale Price)
12.45
1.12
12.47
1.10
12.49
1.05
Age
46.85
31.35
47.60
32.02
48.53
31.94
Fireplace (d)
1.10
1.06
1.08
1.06
1.03
1.06
Pool (d)
0.05
0.22
0.05
0.22
0.05
0.21
Bed
3.61
1.05
3.59
1.02
3.60
1.02
Ranch (d)
0.44
0.50
0.42
0.49
0.44
0.50
Bath
2.73
1.29
2.71
1.28
2.68
1.26
Renovated (d)
0.08
0.27
0.09
0.28
0.10
0.30
Sqft (100)
23.75
14.51
23.53
13.72
23.06
13.86
Notes: This table displays the summary statistics of the control variables by transaction year. The means
for each variable are shown in separate rows, and means and standard deviation for each year are shown in
separate columns.
39


===== PAGE 42 (score 53, numeric-paren 45, table-words 4, var-words 0) =====

Table 7: Coefficient of Year FE
Dependent Variable: Ln (Price)
Model:
Hedonic Index
Repeat Sale Index
Model 1
Model 2
Diff. M1- M2
Model 1
Model 2
Diff. M1- M2
(1)
(2)
(3)
(4)
(5)
(6)
2011 (d)
-0.057*
-0.177***
0.120
-0.012
-0.068**
0.056
(0.030)
(0.026)
(0.033)
(0.028)
P(Chi2)
0.000
0.009
2012 (d)
0.068
-0.060**
0.008
0.063*
0.020
0.043
(0.044)
(0.025)
(0.035)
(0.033)
P(Chi2)
0.004
0.020
2013 (d)
0.362***
0.220***
0.142
0.287***
0.250***
0.037
(0.054)
(0.029)
(0.047)
(0.039)
P(Chi2)
0.003
0.057
2014 (d)
0.463***
0.304***
0.159
0.360***
0.323***
0.037
(0.066)
(0.034)
(0.046)
(0.037)
P(Chi2)
0.006
0.066
2015 (d)
0.494***
0.352***
0.142
0.484***
0.453***
0.031
(0.080)
(0.057)
(0.049)
(0.048)
P(Chi2)
0.019
0.130
2016 (d)
0.553***
0.444***
0.109
0.604***
0.575***
0.029
(0.091)
(0.065)
(0.072)
(0.063)
P(Chi2)
0.115
0.206
Control Unique
No
Yes
No
Yes
R-squared
0.509
0.571
0.947
0.949
Notes: This table displays the year indicator coefficient estimates from the indexing models first for the
hedonic models and then for the repeat sale models. The estimates for each year are presented in separate
rows, and each specification is shown in a set of three columns: the first presenting estimates for a model that
does not control for the uniqueness, the second from a model that includes uniqueness, and the third column
presenting the differences between these two model estimates and the corresponding P–values.
Robust
standard errors are clustered at the Zip Code level, shown in parentheses (*** p<0.01, ** p<0.05, * p<0.1).
40