(12) United States Patent
`Khedkar et ai.
`
`111111
`
`1111111111111111111111111111111111111111111111111111111111111
`US006609118Bl
`
`(10) Patent No.:
`(45) Date of Patent:
`
`US 6,609,118 Bl
`Aug. 19, 2003
`
`(54) METHODS AND SYSTEMS FOR
`AUTOMATED PROPERTY VALUATION
`
`(75)
`
`Inventors: Pratap Shankar Khedkar,
`Philadelphia, PA (US); Piero Patrone
`Bonissone, Schenectady, NY (US);
`David Clarence Golibersuch,
`Schenectady, NY (US)
`
`(73) Assignee: General Electric Company,
`Schenectady, NY (US)
`
`( *) Notice:
`
`Subject to any disclaimer, the term of this
`patent is extended or adjusted under 35
`U.S.c. 154(b) by 0 days.
`
`(21) Appl. No.: 09/337,284
`
`(22) Filed:
`
`Jun. 21, 1999
`
`Int. CI? ................................................ G06F 17/00
`(51)
`(52) U.S. CI. ........................... 705/400; 705/10; 705/30;
`705/35; 705/36; 705/37; 705/38; 705/26;
`705/27; 707/10; 707/100; 707/200
`(58) Field of Search ............................ 705/400, 10, 30,
`705/35, 36, 37, 38, 26, 27; 707/200, 100,
`10
`
`(56)
`
`References Cited
`
`U.S. PATENT DOCUMENTS
`5,361,201 A * 11/1994 Jost et al. ................... 364/401
`5,857,174 A * 1/1999 Dugan ........................... 705/1
`6,178,406 B1 * 1/2001 Cheetham et al. ............ 705/10
`
`FOREIGN PATENT DOCUMENTS
`20002/0035520 A1 * 3/2002
`
`........... G06F/17/60
`
`WO
`
`OlliER PUBLICATIONS
`
`Valentine, Automated Valuation Models Speed the Appraisal
`Process. [retrieved on Apr. 9, 2002]. Retrieved from the
`Internet: <URL:
`http//www.banking.com/aba/mortgage_
`0199.asp>.
`
`O'Rourke, Automated Valuation Models-threat and oppor(cid:173)
`tunity. Appraisal today [online], Sep. 1998 [retrieved on Apr.
`9,2002]. Retrieved from the Internet:<URL: http//www.ap(cid:173)
`praisaltoday.com/ avms.htm>.
`O'Rourke, Where Have We Been and Where Are We Going?
`O'Rourke's predictions for 1999. Appraisal Today [online],
`Jan. 1999 [retrieved on Apr. 9, 2002]. Retrieved from the
`Internet: <URL:
`http://www.appraisaltoday.com/
`sample1.htm>.
`Case Shiller Weiss, [retrieved on Apr. 9, 2002] Retrieved
`from the Internet: <URL: http://www.cswcasa.com/>.
`Automated valuation models spped the appraisal process
`(Http://www.banking.com/aba/mortgate).*
`Automated valuation models-threat and Opportunity
`(Appraisal Today, Oct. 1999). *
`
`* cited by examiner
`
`Primary Examiner-James P. Trammell
`Assistant Examiner-C Owen Sherr
`(74) Attorney, Agent, or Firm-Christoper L. Bernard,
`PLLC
`
`(57)
`
`ABSTRACT
`
`The present invention is a method and system for automat(cid:173)
`ing the process for valuing a property that produces an
`estimated value of a subject property, and a quality assess(cid:173)
`ment of the estimated value, that is based on the fusion of
`multiple processes for valuing a property. In one
`embodiment, three processes for valuing a subject property
`are fused. The first process, called LOCVAL, uses the
`location and living area to provide an estimate of the subject
`property's value. The second process, called AIGEN, is a
`generative artificial intelligence method that trains a fuzzy(cid:173)
`neural network using a subset of cases from a case-base, and
`produces a run-time system to provide an estimate of the
`subject property's value. The third process, calledAICOMP,
`uses a case based reasoning process similar to the sales
`comparison approach to determine an estimate of the subject
`property's value.
`
`26 Claims, 18 Drawing Sheets
`
`10
`
`F-'
`
`Done before use
`
`22
`
`24
`
`Sales price
`LIVing_Area
`
`$/sq. ft. ~
`Q
`Address - I Geocodlng ~ t~~~~~uede
`
`Grid
`
`$ Reliability
`Value
`
`!
`
`!
`
`18
`
`20
`
`

`

`u.s. Patent
`
`Aug. 19,2003
`
`Sheet 1 of 18
`
`US 6,609,118 Bl
`
`FIG. 1
`
`10
`~
`
`Done before use
`
`22
`
`Sales price
`Living_Area
`
`$1
`sq.
`
`ft
`
`•
`
`. A
`Address ~I Geocoding f--- Latitude Y
`
`.
`
`.
`
`Longitude
`
`Whe~~sed for estimating
`
`26 ~ $/sq. ft. Deviation
`;2
`. I
`Latitude
`'\
`--I
`..---~_--. Address
`Geocodlng-Longitude
`28
`t:
`Subject
`14
`Property
`
`Living_Area - - - - - - - - - - . (
`
`$ Reliability
`Value
`
`!
`
`20
`
`!
`
`18
`
`

`

`u.s. Patent
`
`Aug. 19,2003
`
`Sheet 2 of 18
`
`US 6,609,118 Bl
`
`FIG. 2
`
`30
`~
`
`I
`g °V-
`~OWle~}-----+~ Fuzzy ___ B_u_ild __ -----.~ IX~
`~ ~~
`
`32
`
`I
`
`34
`
`r------+o
`o
`
`0
`
`Refine
`
`NN + Fuzzy = Adaptive Knowledge Bases
`
`36
`
`

`

`u.s. Patent
`
`Aug. 19,2003
`
`Sheet 3 of 18
`
`US 6,609,118 Bl
`
`('
`
`FIG. 3
`~
`
`Inputs
`
`IF-part
`
`Rules THEN-part
`
`------,
`
`42'-.. @)
`
`Living_Area
`
`43'-.. @
`
`Lot_Size
`
`44'-..
`@
`
`Location_ Value
`
`0
`®
`
`"\
`Output 38
`~
`
`40
`
`/
`
`Property
`---+ $ Value
`
`Total_Rooms
`
`45'-.. e
`46'-.. e
`47"\.. @
`
`Dev _Prevailing
`
`Bed/Bath
`
`direct connections
`
`

`

`I--"
`~
`00
`I--"
`I--"
`-..\1:;
`Q
`CJ\
`-..CJ\
`rJ'l
`
`e
`
`~
`
`'""'" 00
`o ....,
`~ .....
`'Jl =(cid:173)~
`
`C
`N
`'""'" ~~
`~
`~
`
`8
`
`~ = .....
`~ .....
`~
`•
`rJl
`d •
`
`(1700,$1,$2)
`
`6 ~
`I ,
`
`T
`
`$
`
`f2(LA, LS, LV, ... )
`
`.'
`
`:'
`
`1700
`A
`, ,
`:
`1---------------------\' --------------------t ------------~;
`I then I
`
`$2
`L2:.
`A
`I ,
`small:
`
`I
`
`$1
`A
`, ,
`small :
`
`I
`
`I and I
`
`I and I
`
`I
`
`large
`
`@]
`
`f1 (LA, LS, LV, ... )
`
`~ i
`, , , , ,
`,
`(PV)
`
`Property Value
`
`__ ~i-_--_--------~~~IJ--~-__________ ~~~IJ~---~~~_~~~ ...
`~
`
`1
`
`1
`
`I
`
`~
`
`,
`
`Lot Size Loc Value
`
`(LV)
`
`(LS)
`
`Living Area
`
`(LA)
`
`~
`
`48
`
`FIG. 4
`
`

`

`u.s. Patent
`
`Aug. 19,2003
`
`Sheet 5 of 18
`
`US 6,609,118 Bl
`
`FIG. 5
`
`50
`~
`
`51
`
`Initial Comp
`Retrieval
`
`Transaction Characteristics:
`60
`• Location
`~ Date of Sale
`62
`
`Main House Characteristics:
`64 - - - . Living Area
`• Lot Size
`66
`68~ # Bedrooms
`70~ # Bathrooms
`
`Additional Characteristics:
`Age
`Quality
`Condition
`Fireplaces
`Pool
`
`..-~ Typical $/sq. ft.
`
`Compute
`Similarity Measure
`
`52
`
`53
`
`54
`
`Apply
`Modification
`Rules
`
`Final
`Comp
`Selection
`
`Aggregate
`Selected Comps
`
`55
`
`56
`
`Output: Estimated Property Value
`Reliability: Comp similarity values
`Justification: Comparables used
`
`

`

`u.s. Patent
`
`Aug. 19,2003
`
`Sheet 6 of 18
`
`US 6,609,118 Bl
`
`86 i
`
`months since
`date of sale
`
`88
`
`;
`
`distance from
`subject
`
`90
`
`r' ;
`Ivmg area
`
`92
`
`;
`
`lot size
`
`FIG. 6
`
`80
`~
`
`3
`
`6
`
`9
`
`12 months
`
`.25
`
`.50
`
`.75
`
`1.0 miles
`
`75%
`
`94% 106%
`
`125%
`
`1
`
`0
`
`1
`
`0
`
`1
`
`0
`
`1
`
`o
`
`

`

`u.s. Patent
`US. Patent
`
`Aug. 19, 2003
`Aug. 19,2003
`
`Sheet 7 0f 18
`Sheet 7 of 18
`
`US 6,609,118 Bl
`US 6,609,118 B1
`
`0
`0
`0
`
`.
`
`0
`C\J
`0
`
`0
`ex:>
`·
`0
`
`0
`0
`0r-
`
`·
`
`+
`<.0
`<0
`
`L()
`
`~
`
`Ct)
`
`C\I
`
`0r-
`
`0
`0
`0
`
`O
`0
`·
`0
`
`0
`0
`0
`
`·
`
`L()
`0
`0
`
`0
`L()
`0
`
`·
`
`0
`O
`·
`or-
`
`0
`0
`0
`
`0
`0
`0
`
`·
`
`L()
`0
`0
`
`·
`
`0
`L()
`0
`
`0
`0
`0r-
`
`·
`
`0
`C\I
`·
`0
`
`%
`
`FIG.7
`
`LO
`0
`0
`
`0
`.
`<.0
`0
`
`0
`0
`0r-
`
`0
`Ct)
`O
`
`LO
`0
`0
`
`0
`<.0
`·
`0
`
`0
`0
`0r-
`
`·
`
`0
`LO
`·
`O
`
`LO
`0
`0
`
`0
`0
`0
`
`0
`0
`0r-
`
`·
`
`0
`<.0
`·
`O
`
`L()
`0
`0
`
`·
`
`0
`0
`0
`
`0
`0
`0
`
`0
`CO
`O
`
`0
`C\I
`0
`
`·
`
`0
`0
`0
`
`0
`0
`0
`
`0
`0
`0
`
`·
`
`+ <.0
`
`nmmmmmm
`Subject’s W-I-I-I-—nmwmmmm
`
`or-
`
`C\I
`
`C't)
`
`"Q"
`
`L()
`
`~
`
`en
`~
`(,)
`.......
`(])
`..c
`:::J
`en
`
`en
`E
`0
`0
`"C
`(])
`en
`
`~
`
`~
`
`en
`~en
`_(I)
`E
`(])
`.0
`0
`ro
`sh
`0
`~ ~
`CO
`cu
`"'0
`0...
`(])
`en
`E
`0
`=#:
`()
`
`2 .
`
`0 a
`
`Q Eo 0
`
`

`

`I--"
`~
`00
`I--"
`I--"
`-..\1:;
`Q
`CJ\
`-..CJ\
`rJ'l
`
`e
`
`'""'" 00
`o ....,
`00
`~ ....
`'JJ. =(cid:173)~
`
`C
`N
`'""'" ~~
`~
`~
`
`8
`
`~ = .....
`~ .....
`~
`•
`rJl
`d •
`
`1.00
`
`0.95
`
`0.75
`
`0.35
`
`0.15
`
`0.95
`
`0.70
`
`0.30
`
`O.OS
`
`0.00
`
`0.00
`
`0.00
`
`0.00
`
`5+
`
`1.00
`
`0.90
`
`0.45
`
`0.10
`
`0.00
`
`0.00
`
`0.00
`
`0.00
`
`4.5
`
`0.95
`
`1.00
`
`0.85
`
`0.40
`
`0.05
`
`0.00
`
`0.00
`
`0.00
`
`4
`
`0.70
`
`0.90
`
`1.00
`
`0.85
`
`0.80
`
`1.00
`
`0.20
`
`0.05
`
`0.05
`
`0.00
`
`3.5
`
`0.75
`
`0.25
`
`0.10
`
`0.01
`
`3
`
`84
`
`r
`
`0.30
`
`0.50
`
`0.05
`
`0.10
`
`0.20
`
`0.45
`
`0.80
`
`1.00
`
`0.70
`
`0.25
`
`0.05
`
`2.5
`
`0.00
`
`0.00
`
`0.05
`
`0.15
`
`OAO
`
`0.75
`
`1.00
`
`0.60
`
`0.20
`
`2
`
`0.00
`
`0.00
`
`0.00
`
`0.05
`
`0.10
`
`0.20
`
`0.70
`
`1.00
`
`0.75
`
`1.5
`
`0.00
`
`0.00
`
`0.00
`
`0.00
`
`0.01
`
`0.05
`
`0.10
`
`0.60
`
`1.00
`
`1
`
`----
`
`5+
`
`4.S
`
`4
`
`3.S
`
`3
`
`2.S
`
`2
`
`1.S
`
`1
`
`Comparable
`
`Subject
`
`FIG. 8
`
`

`

`I--"
`~
`00
`I--"
`I--"
`-..\1:;
`Q
`CJ\
`-..CJ\
`rJ'l
`
`e
`
`~
`
`'""" 00
`o ....,
`~ .....
`'JJ. =(cid:173)~
`
`C
`N
`'""" ~~
`~
`~
`
`8
`
`~ = .....
`~ .....
`~
`•
`rJl
`d •
`
`170
`
`0.768333
`
`Similarity Measure (Sum of Weighted Preference/Sum of Weights) =
`
`0.0417
`
`0.056
`
`0.75
`
`2.5->2
`
`I
`
`.
`
`0.0556
`
`0.056
`
`1.00
`
`0%
`
`2
`
`3
`
`2.5
`
`3
`
`# Bathrooms
`
`# Bedroom
`
`0.0367
`
`0.111
`
`0.33
`
`175%
`
`35000
`
`20000
`
`Lot Size
`
`0.2633
`
`0.333
`
`0.79
`
`90%
`
`1800
`
`2000
`
`Living Area
`
`0.2222
`
`0.222
`
`1.00
`
`0.2 miles
`
`0.2 miles
`
`0.1489
`
`0.222
`
`0.67
`
`6 months
`
`6 months
`
`X
`
`X
`
`Distance
`
`date of sale
`Months since
`
`Preference
`
`Weight Weighted
`
`Subject Comparable Comparison Preference
`
`Attribute
`
`160
`
`150
`
`140
`
`130
`
`120
`
`110
`
`100
`
`FIG. 9
`
`

`

`I--"
`~
`00
`I--"
`I--"
`-..\1:;
`Q
`CJ\
`-..CJ\
`rJ'l
`
`e
`
`'""'" 00
`o ....,
`'""'" c
`~ .....
`'JJ. =(cid:173)~
`
`C
`N
`'""'" ~~
`~
`~
`
`8
`
`~ = .....
`~ .....
`~
`•
`rJl
`d •
`
`$ 10000 for a pool
`(Luxury>Excellent>Good>Average>Fair>Poor)
`
`Pool
`
`Quality (.02 * sale price) for each level of difference:
`
`max of 10% of sale price
`W= .5
`if(Age_subject + Age_comp)/2<15 then w = 1 else
`if(Age_subject + Age_comp)/2<8 then w = 2 else
`if(Age_subject + Age_comp)/2<6 then w = 3 else
`if(Age_subject + Age_comp)/2<4 then w = 4 else
`
`w * (Age_comp-Age_subject) * (Sale_Price_comp/1000)
`
`Effective Year Built
`
`(subject -comp) * 2000
`
`Fireplaces
`
`see figure 11
`
`Bathrooms
`
`(subject -comp) * 1
`
`Lot Area
`
`(subject -comp) * (22 + (Sales_Price_of_comp * .00003)
`
`Living Area
`
`y'
`200
`
`FIG. 10
`
`

`

`I--"
`~
`00
`I--"
`I--"
`-..\1:;
`Q
`CJ\
`-..CJ\
`rJ'l
`
`e
`
`'""" 00
`'""" o ....,
`'"""
`~ ....
`'JJ. =(cid:173)~
`
`C
`N
`'""" ~~
`~
`~
`
`8
`
`~ = .....
`~ .....
`~
`•
`rJl
`d •
`
`0.00
`
`'@*S
`
`'@*5
`
`'@*5
`
`'@*S
`
`'@*-5
`
`0.00
`
`'@*-5
`
`-3.00
`
`3.00
`
`0.00
`
`'@*-5
`
`-5.S0
`
`-3.00
`
`6.00
`
`3.00
`
`0.00
`
`'@*-S
`
`-7.S0
`
`-5.00
`
`-2.50
`
`8.00
`
`5.S0
`
`2.S0
`
`0.00
`
`N/A
`
`N/A
`
`N/A
`
`N/A
`
`S+
`
`N/A
`
`N/A
`
`N/A
`
`N/A
`
`4.5
`
`-7.00
`
`-4.50
`
`-2.00
`
`N/A
`
`N/A
`
`N/A
`
`4
`
`-6.50
`
`-9.00
`
`N/A
`
`3.5
`
`3
`
`212
`
`I
`
`~
`210
`
`214.../
`
`'@*5
`
`10.00
`
`7.00
`
`4.50
`
`2.00
`
`0.00
`
`-4.00
`
`-2.25
`
`N/A
`
`N/A
`
`8.50
`
`6.S0
`
`3.00
`
`2.00
`
`0.00
`
`-6.00
`
`-3.S0
`
`-1.00
`
`N/A
`
`N/A
`
`N/A
`
`8.50
`
`6.S0
`
`4.S0
`
`1.S0
`
`0.00
`
`-8.00
`
`-S.OO
`
`-3.00
`
`-1.50
`
`N/A
`
`N/A
`
`N/A
`
`N/A
`
`9.00
`
`7.00
`
`4.00
`
`1.00
`
`0.00
`
`5+
`
`4.5
`
`4
`
`3.5
`
`3
`
`2.S
`
`2
`
`1.S
`
`1
`
`2.5
`
`2
`
`1.S
`
`1
`
`Subject Comp
`
`FIG. 11
`
`

`

`I--"
`~
`00
`I--"
`I--"
`-..\1:;
`Q
`CJ\
`-..CJ\
`rJ'l
`
`e
`
`'""" 00
`o ....,
`'""" N
`~ .....
`'JJ. =(cid:173)~
`
`C
`N
`'""" ~~
`~
`~
`
`8
`
`~ = .....
`~ .....
`~
`•
`rJl
`d •
`
`195750
`
`10000
`
`3500
`
`2800
`
`2000
`
`2000
`
`-5000
`
`5450
`
`175000
`
`No
`
`Average
`
`Average
`
`89
`
`0
`
`3
`
`2
`
`25000
`
`1800
`
`175000
`
`Yes
`
`Average
`
`Good
`
`93
`
`1
`
`3
`
`2.5
`
`20000
`
`2000
`
`?
`
`Adjusted Price =
`Pool
`
`Condition
`
`Quality
`
`Eft Year Built
`
`SFR Fireplaces
`
`SFR Bedrooms
`
`SFR Total Baths
`
`Lot Area
`
`Living Area
`
`Sale Price
`
`~
`
`~
`
`236
`
`234
`
`232
`
`230
`
`228
`
`226
`
`224
`
`222
`
`Adjustment
`
`Comparable
`
`Subject
`
`Attribute
`
`~
`
`220
`
`FIG. 12
`
`

`

`u.s. Patent
`
`Aug. 19,2003
`
`Sheet 13 of 18
`
`US 6,609,118 Bl
`
`FIG. 13
`
`240
`~
`
`Create ordinal comp score for:
`-Similarity
`-Net adjustment
`-Gross Adjustment
`
`2
`V 24
`
`Threshold on similarity measure until 4
`or more comps are obtained
`(& at least 1 of different sign)
`
`244
`
`V
`
`Compute total comp score V 246
`
`Divide comps by sign of net adj. V 24 7
`
`Order each group by overall score V- 2 48
`
`Within each group, select up to V 24
`
`four comps (when available)
`
`9
`
`

`

`I--"
`~
`00
`I--"
`I--"
`-..\1:;
`Q
`CJ\
`-..CJ\
`rJ'l
`
`e
`
`'""'" 00
`o ....,
`'""'" ~
`~ .....
`'JJ. =(cid:173)~
`
`C
`N
`'""'" ~~
`~
`~
`
`8
`
`~ = .....
`~ .....
`~
`•
`rJl
`d •
`
`27
`
`14
`
`21
`
`10
`
`13
`
`18
`
`17
`
`8
`
`7
`
`9
`
`2
`
`8
`
`3
`
`5
`
`6
`
`7
`
`1
`
`4
`
`11300
`
`4410
`
`9261
`
`5670
`
`6099
`
`6160
`
`8191
`
`4186
`
`5924
`
`9
`
`4
`
`6
`
`1
`
`3
`
`8
`
`7
`
`5
`
`2
`
`9310
`
`3546
`
`-5261
`
`-948
`
`3139
`
`6150
`
`5686
`
`3586
`
`1344
`
`9
`
`8
`
`7
`
`6
`
`5
`
`4
`
`3
`
`2
`
`1
`
`0.44
`
`0.48
`
`0.53
`
`0.58
`
`0.64
`
`0.67
`
`0.78
`
`0.88
`
`0.95
`
`-
`
`-
`
`331-018
`
`431-023
`
`873-005
`
`847-984
`
`685-046
`
`305-006
`
`093-011
`
`306-018
`
`113-012
`
`Rank
`
`Rank
`
`Value
`
`Total
`
`Gross Adjust G.A.
`
`Rank
`
`N.A.
`
`Value
`
`Rank
`
`Value
`
`Net Adjust
`
`Score
`
`Score
`
`Comparable
`
`~
`
`250
`
`FIG. 14
`
`

`

`I--"
`~
`00
`I--"
`I--"
`-..\1:;
`Q
`CJ\
`-..CJ\
`rJ'l
`
`e
`
`'""'" 00
`o ....,
`'""'" Ul
`~ .....
`'JJ. =(cid:173)~
`
`C
`N
`'""'" ~~
`~
`~
`
`8
`
`~ = .....
`~ .....
`~
`•
`rJl
`d •
`
`199900
`
`757640
`
`116580
`
`122880
`
`153270
`
`177760
`
`187150
`
`Price
`
`3.83
`
`0.58
`
`0.64
`
`0.78
`
`0.88
`
`0.95
`
`Final estimate = 757640/3.83 =
`
`201000
`
`192000
`
`196500
`
`202000
`
`197000
`
`Price
`
`Total
`
`847-984
`
`685-046
`
`093-011
`
`306-008
`
`113-012
`
`Weighted
`
`Score
`
`Adjusted
`
`Comparable
`
`~
`
`260
`
`FIG. 15
`
`

`

`u.s. Patent
`
`Aug. 19,2003
`
`Sheet 16 of 18
`
`US 6,609,118 Bl
`
`FIG. 16
`
`260
`~
`
`262
`
`l..
`
`264
`
`l..
`
`266
`
`l..
`
`270
`
`l..
`
`268
`
`l..
`
`comps found
`
`1
`
`0
`
`I 4
`
`3
`
`similarity
`
`1
`
`0
`
`I
`
`.25
`
`I
`9
`
`I
`12
`
`6
`
`t::
`
`.75
`
`.5
`
`1
`
`atypicality
`1
`
`I ~
`
`.5
`
`1
`
`1.5
`
`2
`
`deviation of comparables values
`
`0
`
`1
`
`0
`
`I ~I
`
`6%
`
`9%
`
`12%
`
`15%
`
`3%
`span of comparables values
`1
`
`0
`
`10% 20% 30%
`
`~I
`
`40%
`
`

`

`I--"
`~
`00
`I--"
`I--"
`-..\1:;
`Q
`CJ\
`-..CJ\
`rJ'l
`
`e
`
`'""'" 00
`o ....,
`'""'"
`-..J
`~ .....
`'JJ. =(cid:173)~
`
`C
`N
`'""'" ~~
`~
`~
`
`8
`
`~ = .....
`~ .....
`~
`•
`rJl
`d •
`
`0.32
`
`0.00
`
`0.80
`
`0.80
`
`1.00
`
`1.00
`
`1.00
`
`0.70
`
`1.00
`
`0.15
`
`Value
`
`Cont.
`
`13
`
`15
`
`8.11
`
`18
`
`12
`
`9.33
`
`7.24
`
`19
`
`8.57
`
`6.32
`
`4.24
`
`3.85
`
`2.83
`
`4.5
`
`3.24
`
`2.89
`
`2.05
`
`5.67
`
`2.24
`
`2.02
`
`1.34
`
`1.97
`
`0.81
`
`0.17
`
`0.73
`
`0.29
`
`0.66
`
`0.94
`
`0.38
`
`1.42
`
`0.77
`
`0.82
`
`0.74
`
`0.74
`
`0.90
`
`0.95
`
`0.85
`
`0.71
`
`0.94
`
`0.63
`
`11
`
`12
`
`19
`
`12
`
`15
`
`14
`
`24
`
`11
`
`35
`
`3
`
`7.8
`
`-13.9
`
`3.1
`
`5.2
`
`5.2
`
`-1.6
`
`0.5
`
`17.3
`
`-2
`
`-9.8
`
`Span
`
`Oev.
`
`Found
`
`Comps
`
`Comps
`
`Atyp.
`
`Simil.
`
`Comps
`
`Error
`
`FIG. 17
`
`

`

`I--"
`~
`00
`I--"
`I--"
`-..\1:;
`Q
`CJ\
`-..CJ\
`rJ'l
`
`e
`
`'""'" 00
`o ....,
`'""'" 00
`~ .....
`'JJ. =(cid:173)~
`
`C
`N
`'""'" ~~
`~
`~
`
`8
`
`~ = .....
`~ .....
`~
`•
`rJl
`d •
`
`5%
`
`I FU~ON I
`
`$$
`
`HIGH
`
`AICOMP
`
`Esitmate
`Fused $
`
`Pool, AicCond, and Heating
`Age, Eft_Age, Quality, Condition, Fireplaces,
`optional property attributes if available:
`AICOMP can use the following
`
`7-9%
`
`$$
`
`AIGEN
`
`10-12%
`
`$$
`
`LOW
`
`Reliability
`
`Relative
`
`Error
`
`Estimator
`
`Required
`
`Inputs
`
`at Run Time
`Computation
`
`~
`
`300
`
`FIG. 18
`
`

`

`US 6,609,118 Bl
`
`1
`METHODS AND SYSTEMS FOR
`AUTOMATED PROPERTY VALUATION
`
`BACKGROUND OF THE INVENTION
`
`5
`
`2
`reliability is limited, an explanation is provided as to the
`limitations of the estimate. These characteristics allow for a
`determination of the suitability of the estimate within a
`given business application context.
`
`BRIEF DESCRIPTION OF THE DRAWINGS
`
`FIG. 1 shows a schematic of the LOCVAL system.
`FIG. 2 shows a schematic of the AIGEN system.
`FIG. 3 shows a schematic of the architecture of the
`AIGEN system.
`FIG. 4 shows a schematic of the fuzzy interference
`process for the AIGEN system.
`FIG. 5 shows a flow chart for the AICOMP system.
`FIG. 6 shows the preference criteria for four attributes.
`FIG. 7 shows the reflexive asymmetric relations for the
`bedrooms attribute.
`FIG. 8 shows the reflexive asymmetric relations for the
`bathrooms attribute.
`FIG. 9 shows an example of the similarity measure
`computation.
`FIG. 10 illustrates sample adjustment rules.
`FIG. 11 shows an example of adjustments to a compara(cid:173)
`ble's prices as a function of the number of different bath(cid:173)
`rooms between the subject property and the comparable
`property.
`FIG. 12 shows a sample computation of a comparable's
`30 adjusted price.
`FIG. 13 is a flow chart of the method for selecting the best
`comparables.
`FIG. 14 shows a comparison of comparablcs based on a
`number of different factors.
`FIG. 15 shows a sample computation of the final estimate.
`FIG. 16 shows the membership functions for a variety of
`characteristics.
`FIG. 17 illustrates that the estimated error decreases as the
`40 number of comparables increases.
`FIG. 18 shows a schematic of the Fusion system.
`
`DETAILED DESCRIPTION OF THE
`INVENTION
`
`The present invention relates generally to property valu(cid:173)
`ation and more particularly to automated property valuation.
`Property valuation is a process of determining a dollar
`estimate of a property's value for given market conditions. 10
`The value of a property changes with market conditions.
`Consequently, a property's value is often updated to reflect
`changes in market conditions, including for example, recent
`real estate transactions.
`Property valuations have many applications. For example, 15
`many financial institutions grant new mortgages to
`homebuyers, and purchase mortgage packages, which can
`contain hundreds of mortgages, on the secondary market as
`investments. Property valuations are usually necessary to
`grant most new mortgages, as well as to evaluate mortgage 20
`packages that may be available for purchase. By way of
`further example, property valuations are also used to guide
`buyers and sellers with making purchasing decisions, and
`are needed for a variety of insurance purposes.
`The current process for valuing properties usually 25
`requires an on-site visit by a human appraiser, can take
`several days, and can cost hundreds of dollars per subject
`property. The process usually used by appraisers is a sales
`comparison approach, which consists of finding compa(cid:173)
`rabIes (i.e., recent sales that are comparable to the subject
`property, using for example sales records), contrasting the
`subject property with the comparables, adjusting the com(cid:173)
`parables' sales prices to relled the dilIerences thereof from
`the subject property, using for example, heuristics and
`personal experience, and reconciling the comparables' 35
`adjusted sales prices to derive an estimate for the subject
`property, using any reasonable averaging method.
`The human appraisal process is slow and expensive for
`multiple appraisals, which are often required by banks to, for
`example, update their loan and insurance portfolios, verify
`risk profiles of servicing rights, or evaluate default risks for
`securitized mortgage packages. Consequently, the appraisal
`process for multiple valuations is currently estimated, to a
`lesser degree of accuracy, by sampling techniques.
`
`45
`
`An automated property valuation for a subject property is
`obtained by combining the outputs of the three property
`valuation estimators, which are referred to as LOCVAL,
`AIGEN, and AICOMP. These three estimators, which are
`described below, each provide an independent estimate of a
`50 value for a subject property using different methodologies.
`
`SUMMARY OF THE INVENTION
`
`Thus, there is a particular need to automate the valuation
`process. The present invention is a method and system for
`automating the valuation process that produces an estimated
`value of a subject property that is based on the fusion of
`multiple processes for valuing property. In one embodiment,
`three processes for valuing a subject property are fused. The
`first process, called LOCVAL, uses the location and living
`area to provide an estimate of the subject property's value. 55
`The second process, called AIGEN, is a generative artificial
`intelligence method that trains a fuzzy-neural network using
`a subset of cases from a case-base, and produces a run-time
`system to provide an estimate of the subject property's
`value. The third process, called AICOMP, uses a case based 60
`reasoning process similar to the sales comparison approach
`to determine an estimate of the subject property's value.
`The fusion of LOCVAL, AIGEN and AICOMP provides
`a better estimate than anyone method alone. Fusion also
`provides a way of assessing the quality or reliability of the 65
`fused estimate. If reliability is high, the fused estimate is
`more accurate than any of the individual estimates. If
`
`LOCVAL
`
`Referring to FIG. 1, the LOCVAL system 10, which can
`be implemented in the form of a computer program, takes as
`an input a valid, geocoded address 12 and a living area 14
`(in sq. ft.) for the subject property 16, and outputs two
`values, the locational_ value 18, which is an estimated value
`of the subject property, and deviation13 I from_prevailing 20,
`which is the standard deviation for properties within the
`selected geographic region. If either input is missing, or
`clearly out-of-range, LOCVAL 10 does not provide an
`output.
`The output is based the values of all properties within a
`certain geographic region, for example a neighborhood, city,
`county or state. In this regard, all known, filtered historical
`sales 22 in a geographic region are used to construct a
`
`

`

`3
`smooth surface 24 spanning a geographic region that rep(cid:173)
`resents a dollar/sq. ft. value 26 and deviation 28 at every
`point of longitude and latitude within the selected geo(cid:173)
`graphic region. A smoothing function 24 is derived using
`radial basis functions that drop off exponentially with dis-
`tance and a "space" constant of about 0.15-0.2 miles. The
`smoothing function 24 is described as the weighted sum of
`radial basis functions (all of the same width), each situated
`at the site of a sale within the past one year and having an
`amplitude equal to the sales price. Consequently, based on 10
`the inputs of a valid, geocoded address 12 and a living area
`14 for the subject property 16, an estimate 18 of the subject
`property's value and corresponding reliability 20 can be
`obtained.
`
`5
`
`AlGEN
`
`Referencing FIG. 2, the AlGEN system 30 is a generative
`system based on a combination of fuzzy logic systems 32
`and neural networks 34. The AlGEN system 30 is a network(cid:173)
`based implementation of fuzzy inference based on a system
`that implements a fuzzy system as a five-layer neural
`network 34 so that the structure of the network 34 can be
`interpreted in terms of high-level rules. The neural network
`34 is trained automatically from data 36.
`FIG. 3 shows the architecture 38 used to output an
`estimate of the subject property's value 40. The output may
`comprise linear functions of variables that do not necessarily
`occur in the input (i.e., segment the input space on a proper
`subset of the total variable set only and use a cylindrical
`projection of that segmentation for the whole space).
`FIG. 4 shows a schematic 48 for the fuzzy inference
`process, where the rules have the following form:
`
`Rule 1: IF x is A, and y is B, THEN z is f,(x,y)
`
`(1) 35
`
`Rule 2: IF x is A2 and y is B2 THEN z is f2(X,y)
`
`(2)
`
`The two variables x and y take on real values. The predicate
`Ai against which x is matched is a fuzzy set rather than a
`crisp value or an interval. All the sets Ai and Bi above are
`fuzzy sets. The IF part is referred to as the antecedent or
`precondition of the rule, and the THEN part is the conse(cid:173)
`quent or postcondition.
`The IF/THEN rules are used to map inputs to outputs by
`a fuzzy logic inference system that works in several steps.
`First, the inputs are matched against the fuzzy sets Ai and Bi.
`Second, the degree of applicability Wi of each rule is
`determined by multiplying together the degrees to which
`that rule's antecedent clauses match the given input. Third,
`the outputs recommended by each rule are determined by
`evaluating f;(x,y) on the input. Finally, the output is defuzzi(cid:173)
`fied by combining the outputs of all rules by a normalized,
`weighted sum, where the weight of a rule is its degree of
`applicability Wi'
`The specific form used for f(x,y) is a linear function of the
`inputs, such that the general rule is:
`
`(3)
`
`Such a rule is referred to as a TSK-type rule. A special case
`of equation (3) is when all cij except CiO are 0, in which case
`each rule recommends a fixed number. The inference pro(cid:173)
`cedure with TSK-type rules yields:
`
`(4)
`
`where Wi is the weight of Rule i, computed as a weighted
`sum.
`
`US 6,609,118 Bl
`
`4
`The antecedent fuzzy membership functions are given by:
`
`flA (x)~l/(l +( (x-c )/a )2b)
`
`(5)
`
`The membership function given by equation (5) is centered
`symmetrically around c, has a width controlled by a, and has
`a curvature controlled by b. For b=oo, a crisp interval [c-a,
`c+a] is obtained. For b---;.oo, the set A tends to a non-fuzzy
`interval [c-a, c+a]. For b=O, the membership function (i.e.,
`curve) does not bend at all. Although an initial value of b=2
`is preferably used, the selection of the initial value ofb is not
`critical as the system will change the value of b if required
`by the data. Consequently, the tuning of the value of b is not
`critical.
`The weight Wi of each rule is obtained by multiplying the
`15 fA(x) of the two clauses in that rule's IF part. The choice of
`granularity (i.e., how many fuzzy functions per axis) is
`governed by the tradeoff between simplicity and accuracy. A
`high number of rules leads to a more "folded" surface and
`is preferably avoided unless it is necessary for fitting the
`20 data. Preferably, up to two membership functions per input
`dimension are used.
`The system's architecture is based on the number of
`membership functions assigned to each input dimension. For
`example, if six inputs are used, and two membership func-
`25 tions are assigned to four of the inputs, the network will have
`six input units, eight units in the first layer (which come from
`the two membership functions for each of the four
`variables), sixteen in the next two layers (which come from
`every combination of one membership function from each of
`30 the four inputs, i.e., 2x2x2x2=16 rules), and one summation
`unit to produce the output in the output layer. Each of the
`sixteen rules has a TSK-type consequent which depends on
`all six inputs. Since each antecedent membership function
`has three degrees of freedom (a, b, and c), and each
`consequent has seven coefficients, there are one hundred
`thirty-six degrees of freedom (8x3+16x7=136).
`Once the architecture is constructed, the parameters are
`initialized in a reasonable manner instead of randomly as in
`neural networks. For example, the membership functions
`40 can be spaced at uniform distances over the axis so as to
`cover the range of the data points. The consequent linear
`functions are initialized to zero.
`A variant of the gradient descent technique is used to train
`the network 34 based on training data in order to minimize
`45 the mean squared error between the network's 34 outputs
`and the desired answers, when presented with the data points
`in the training set. The training of the network 34 includes
`several steps. First, a sample point in a training data set is
`presented to the network 34 and the output is computed.
`50 Second, the error between the network's 34 output and the
`desired answer is computed. Third, holding the IF-part
`parameters fixed, the optimal values of the THEN -part
`parameters are solved using a least-mean-squares optimiza(cid:173)
`tion method. A recursive Kalman filter method is preferably
`55 used. Fourth, the effect of the IF-part parameters on the error
`is computed using derivatives of the functions implemented
`by intermediate layers. Fifth, using the information
`obtained, the IF-part parameters are changed by small
`amounts so that the error at the output is reduced. Finally, the
`60 above steps are repeated several times using the entire
`training set, until the error is sufficiently small. Repeating
`the above steps (i.e., training) is stopped when the error
`becomes fixed or decreases very slowly. The resulting
`network 34 is interpreted as a fuzzy rulebase, with each
`65 parameter in the network 34 having a definite meaning in
`terms of the fuzzy sets or consequent functions. Notably,
`learning speed is very fast compared to the conventional
`
`

`

`US 6,609,118 Bl
`
`5
`neural net paradigm. Additional data, if available, can be
`used to further train the network 34 using the same
`backpropagation-type algorithm. The resulting surface is
`very well-behaved and provably smooth. The rule base is
`extremely compact, so a large number of models of the 5
`network 34 can be stored easily.
`Different models for the same problem can be obtained by
`changing the inputs to the neuro-fuzzy network, or by
`varying its architecture. For example, one could make the
`dollars/sq. ft. value for the property the dependent variable, 10
`use a network to compute this value, and then multiply this
`value by the living area to generate the predicted price. By
`way of another example, one could make the logarithm of
`the sale price the dependent variable (i.e., the output of a
`network), or one could use a different combination of 15
`property attributes as the inputs to a network. The choice of
`which model to deploy depends on the evaluation or error
`metric. Preferably, models which estimate a property's
`value, directly, or which estimate a property's value using
`the sale price per unit of building area, are used.
`The AlGEN system 30 uses a network 34 comprising six
`inputs 42-47 and one output 40. Four of the inputs 42-45 are
`used to partition the input space into sixteen overlapping
`sets, and give rise to sixteen fuzzy rules. The remaining two
`inputs 46, 47 are used only by the consequents of the rules. 25
`Each of the sixteen fuzzy rules provide an output (i.e., a
`prediction based on the six input variables), and the outputs
`are then combined using interpolative reasoning to produce
`an estimate of the subject property's value 40. The overall
`map from six inputs 42-47 to one output 40 is a nonlinear, 30
`differentiable map that is constructed by melding together
`sixteen hyperplanes in a seven dimensional space. For
`example, the form of a rule can be: IF lot_size is small and
`livin~area is small and locational_value is high ... WEN
`price is f( ), where f( ) is a linear function of the six input 35
`variables.
`The data set of property values used to train the AlGEN
`system 30 is preferably restricted to a certain price range to
`eliminate obvious outliers. Consequently, the system 30 is
`preferably not used to estimate the value of a property that 40
`is outside the restricted price range. The system 30 will
`output a price estimate if given a property that is outside the
`price range, however, the output will be bound by the range
`and a warning will be issued to the user of the system.
`Preferably, one system is used for the entire price range, as 45
`opposed to separate systems for each price range.
`The sixteen fuzzy rules have one hundred twelve degrees
`of freedom in the consequent. This is a large share of the
`dimensionality of the parameter space, which uses a variant
`of the Kalman filtering algorithm to train the parameters in 50
`the consequent. Preferably the consequent is partially (four
`inputs=eighty parameters) trained using the interleaved
`backpropagation process described above, followed by a
`final batch phase where all one hundred twelve consequent
`parameters are retrained while holding the antecedent 55
`parameters constant. The training set is preferably error-free
`and more or less randomly distributed so as not to bias the
`system 30. The specific size of the training set is not
`significant. Preferably, the training data set includes at least
`ten times the number of degrees of freedom (e.g., lOx136= 60
`1,360). Preferably, only a small part of the total data is used
`to avoid overfitting of the data.
`Inputs to the system 30 are based on seven attributes of a
`property: total_rooms 45, num_bedrooms, num_baths,
`living_area 42, lot_size 43, locational_value 44, and 65
`deviation_from_prevailing 46. The number of bedrooms
`and bathrooms are combined to produce a bedrooms/
`
`6
`bathrooms ratio 47 that is fed along with the other five
`values to the system 30. Of the six inputs, total_rooms 45,
`livin~area 42, lot_size 43, and locational_value 44 are
`used to partition the space into sixteen fuzzy regions. The
`output 40 is the dollar value of the house. Alternatively, the
`system 30 can produce the 10glo(sale_price) as an output. In
`this embodiment, the inputs are: 10glo(total_rooms), 10glo
`(lot_size), 10glo(living_area), 10glo(locational_value),
`bedroomslbathrooms, and deviation_from_prevailing. The
`first four inputs are used to partition the input space.
`Alternatively, the system 30 can produce the sale_price per
`sq