US007120599B2
`
`(12) United States Patent
`US 7,120,599 B2
`(10) Patent No.:
`(45) Date of Patent:
`Oct. 10, 2006
`Keyes
`
`(54) METHODS AND SYSTEMS FOR MODELING
`USING CLASSIFICATION AND
`REGRESSION TREES
`
`(75)
`
`Inventor: Tim Kerry Keyes, West Redding, CT
`(US)
`
`(73) Assignee: GE Capital Commercial Finance,
`Inc., Stamford, CT (US)
`
`( * ) Notice:
`
`Subject to any disclaimer, the term of this
`patent is extended or adjusted under 35
`U.S.C. 154(b) by 1208 days.
`
`(21) Appl. No.: 09/746,411
`
`(22)
`
`Filed:
`
`Dec. 21, 2000
`
`(65)
`
`Prior Publication Data
`
`US 2001/0044766 A1
`
`Nov. 22, 2001
`
`Portfolio Management, Spring 1996, 22, 3; ABI/INFORM Global,
`p. 106-114.*
`Gath, I. and Geva, A.B., “Unsupervised Optimal Fuzzy Clustering,”
`IEEE Trans. Pattern Anal. Machine Inte11., v01. PAMI-ll, N0. 7, pp.
`773-781, Jul. 1989.
`Bezdek, James C.; Hathaway, Richard J.; Sabin, Michael J.; and
`Tucker, William T., “Convergence Theory For Fuzzy c-Means:
`Counterexamples and Repairs,” IEEE Trans. Syst., Man, Cybern.,
`v01. SMC-l7, N0. 5, pp. 873-877, Sep./Oct. 1987.
`Dunn, J.C., “A Fuzzy Relative 0fthe ISODATA Process and Its Use
`in Detecting Compact Well-Separated Clusters,” J. Cybernetics,
`v01. 3, N0. 3, pp. 32-57, 1973.
`L. Breiman, J.H. Friedman, R.A. Olshen and C]. Stone (1984),
`Classification and Regression Trees, Chapters 1, 2, 3, 8 and 11,
`Wadsworth, Belmont, CA.
`Brealey, Richard A. and Myers, Stewart, C.Principles OfCorporate
`Finance.
`
`Efron, Bradley and Tibshirani, Robert J., “An Introduction to the
`Bootstrap,” Chapman & Hall publishers.
`
`Related US. Application Data
`
`* cited by examiner
`
`(60) Provisional application No. 60/174,057, filed on Dec.
`30, 1999.
`
`Primary ExamineriLalita M. Hamilton
`(74) Attorney, Agent, or FirmiArmstrong Teasdale LLP
`
`(51)
`
`Int. Cl.
`(2006.01)
`G06F 1 7/60
`(52) US. Cl.
`...................................... 705/36 R; 705/35
`(58) Field of Classification Search .................. 705/35,
`705/36, 37, 36 R
`See application file for complete search history.
`
`(56)
`
`References Cited
`U.S. PATENT DOCUMENTS
`
`6,088,685 A *
`6,792,399 B1 *
`
`.................. 705/36
`7/2000 Kiron et a1.
`................. 703/2
`9/2004 Phillips et a1.
`FOREIGN PATENT DOCUMENTS
`
`GB
`
`WO 02067087
`
`*
`
`8/2002
`
`OTHER PUBLICATIONS
`
`Derman, Emmanuel, “Valuing Models and Modeling Value: A
`Physicist’s Perspective on Modeling on Wall Street”, Jounal 0f
`
`(57)
`
`ABSTRACT
`
`A method of valuation of large groups of assets using
`classification and regression trees is described. The method
`includes defining relevant portfolio segmentations, assess-
`ing performance of the classification and regression tree
`based model against a simple model and ranking all port-
`folio segments based upon performance of the models.
`Iterative and adaptive statistical evaluation of all assets and
`statistical inferences are used to generate the segmentations.
`The assets are collected into a database, grouped by credit
`variable, subdivided by ratings as to those variables and then
`rated individually. The assets are then regrouped and a
`collective valuation is established by cumulating individual
`valuations.
`
`30 Claims, 14 Drawing Sheets
`
`28
`
`DD TIMELINE
`’/’
`
`BEGIN
`
`
`
`
`EXTRAPOLATION
`
`IMPROVES
`
`"BEST“ VALUE
`CONTINUED
`TO IMPROVE
`
`
`
`
`
`TRULIA - EXHIBIT 1007
`
`TRULIA - EXHIBIT 1007
`
`

`

`U.S. Patent
`
`Oct. 10, 2006
`
`Sheet 1 0f 14
`
`US 7,120,599 B2
`
`IIIIIIII
`UNDER—
`WRITTEN
`
`
`
`
`
`
`
`
`
`4
`
`
`20
`
`PORT ION
`
`

`

`U.S. Patent
`
`Oct. 10, 2006
`
`Sheet 2 0f 14
`
`US 7,120,599 B2
`
`TIM INE
`
`
`28
`
`BEGIN
`
`l/f‘
`EL
`DD
`
`
`BID
`6855‘ 56 58 5° 62
`64 E)
`46
`30
`48
`i§§flfififlfifi 42 //”
`._.»”'
`
`84
`
`7o
`
`72
`
`
`
`nul-
`
`
`
`178 82
`\
`
`
`
`
`
`15 @fi
`74
`
`76
`
`EXTRAPOLATION
`IMPROVES
`
`"BEST" VALUE
`CONTINUED
`TO IMPROVE
`
`46
`
`50
`
`54
`
`48
`
`52/
`66 V
`
`56
`
`58
`
`60
`
`52
`
`64
`
`

`

`U.S. Patent
`
`Oct. 10, 2006
`
`Sheet 3 0f 14
`
`US 7,120,599 B2
`
`UNTOUCHED
`
`LOAN
`
`TABLE
`
`‘44
`142
`
`206(208135
`
`SUPERV I SED
`UNSUPERV I SED
`LEARNING > LEARNING > Sgfififigg
`PROCESS
`PROCESS
`
`UNDERWR I T I NC
`CLUSTERS
`TABLE
`
`>
`
`138
`
`ADJ CREDIT ’“ 14°
`BETA ADJ
`p FOR CREDIT > ANALYST >
`SCORE
`TABLE
`
`125
`
`126
`
`CREDIT
`MANUAL
`LOAN LEVEL
`SAMPLE
`100% A > RE—UNDERWRITE > DATA b ANALYST >
`RANDOM B&C
`PROCESS
`ENTRY
`TABLE
`
`130
`
`40
`
`108
`
`B ADJUST
`>55“ CLASS b CREDIT ANALYST b
`
`TABLE
`BETAADJUST
`F1”
`106:
`RULE SET TO
`r».> GROUP “‘4:} DESECRECATE p
`T0 LOAN LEVEL
`PARTIAL
`
`a116
`ELECTRONIC
`UPLOAD To
`
`SUPER A TABLE
`
`PARTIAL
`
`CASH?
`
`11108112
`
`CrSAMPLE
`PAR“;
`
`UNDERWRI E
`TEAM
`
`VALUE * AUTHENTICATION F CASH TABLE >
`ASSETS
`
`14869<FULL
`
`VALUE b
`LOANS
`
`TEAM
`AUTHENTICATION
`
`V
`
`1007CASH
`TABLE
`
`>
`
`100%
`9
`CASH.
`
`85 4
`
`166—]
`
`168
`
`_\
`
`12 _/ PORTFOLIO
`
`T0 FIG. 4
`
`FIG.3
`
`

`

`U.S. Patent
`
`Oct. 10, 2006
`
`Sheet 4 0f 14
`
`US 7,120,599 B2
`
`FROM FIG. 3
`
`I48
`
`152
`
`168
`
`/
`
`p-y CASH FLOW p CASH FLOW >
`
`VALUATION
`
`BRI DOE
`
`BR | DGE
`
`DETERMINISTIC
`
`STOCHASTIC
`
`146
`
`CASH FLOW
`TIMING TABLE /
`
`I50
`
`154
`
`156
`
`GE PURSE
`PREFERENCES
`TRANCHE PRIORTY
`
`V
`BID OPENING
`
`
`
`
`
`
`TRANCHE
`
`> IRR- "P >-’
`MODEL
`A
`OTHER BIDDER
`PREFERNCES
`
`/ AOTHER BIDDER
`
`A
`BID PROCESS
`RULE SET
`
`
`
`MAX EXPECTED
`> IRR SUBJECT >
`T0 NPV>0
`
`MEAN IRR ’ SENIOR MGT. } BID FORMS
`>30?
`SETS BID
`& BID
`
`PARTNERSHIP
`FINANCIAL
`PRO FORMA
`
`V
`
`A
`PARTNER
`ROUND TABLE
`BID PRICE
`
`k
`
`164
`
`FIG.4
`
`

`

`U.S. Patent
`
`Oct. 10, 2006
`
`Sheet 5 0f 14
`
`US 7,120,599 B2
`
`184
`
`188
`
`FIG.5
`
`

`

`U.S. Patent
`
`Oct. 10, 2006
`
`Sheet 6 0f 14
`
`US 7,120,599 B2
`
`212
`
`DEFINE
`
`(/’216
`
`uw
`SAMPLE
`
`218
`
`
`
`220
`
`SET
`ATTRIB.
`
`222
`
`CLASSIFY
`
`EXPERT
`OPINION
`
`214
`
`224-W
`
`VALUE
`
`CLUSTER
`
`226
`
`
`
`FIG.E5
`
`228
`
`km
`
`

`

`U.S. Patent
`
`Oct. 10, 2006
`
`Sheet 7 0f 14
`
`US 7,120,599 B2
`
`230
`
`232
`
`234
`
`DATA
`ACQUISITION
`
`VARIABLE
`SELECTION
`
`HIERARCHICAL
`SEGMENTATION
`
`
`78
`
`
`
`CASH FLOW/
`RISK SCORES
`
`1
`
`UNDER—WRITING
`
`REVIEW
`
`FIG.7
`
`

`

`U.S. Patent
`
`Oct. 10, 2006
`
`Sheet 8 of 14
`
`US 7,120,599 B2
`
`240
`
`K
`
`242
`
`MANUAL
`
`UW
`
`FORM
`
`CLUSTERS
`
`CALCULATE
`
`COUNTS
`
`
`
`
`
`244
`
`246
`
`248
`
`250
`
`252
`
`254
`
`FIG.8
`
`

`

`U.S. Patent
`
`()ct 10,2006
`
`Sheet 9 of 14
`
`US 7,120,599 B2
`
`m_m>4<z<
`
`x:.\\
`
`O¢
`
`
`
`4mooz304mIm<o
`
`zuhh_mgmmoz:I202OH>41&<*
`
`
`
`oz_k_m3mmoz:mmDHa<o*
`
`moom4302x
`
`mpmmm<
`
`
`
`OQFmzo_H<:4<>
`
`0.0Hu
`
`zo_p<:4<>
`
`m02uo_4_ouse
`
`mmomo<>p_4_m<_m<>£pmoo
`
`museum
`2_>Hz_<HmmozsQHmoo
`m02m0_4_ouse
`mesomm
`
`zooz_p_m3mmoz:
`
`mhmmw<nomga2<m
`
`
`
`ohmhmmm<>z<2ooh*
`
`
`
`wmomzommmum§_hIODOZmHoz*
`
`up_m3mmoz:
`
`\l0mm
`
`
`
`
`
`
`

`

`U.S. Patent
`
`()ct 10,2006
`
`Sheet10 0f14
`
`US 7,120,599 B2
`
`
`
`zgoo10HmEMH_HOMJUm
`
`
`
`>p_o<m<omomzowmm3:1
`
`
`
`mH2m2mm_30mmooml
`
`20mmOmz_2mMHmox*
`
`
`
`.aHRxmoma:m2=o\g04oh_Izo_p<o_:o_4o_40LHmoa
`ma4<OFmp<mmomo
`m;458E220_mm<-Aomv
`
`
`
`gnomemmaommOmone
`
`wh<umoon
`
`
`
`_a:4<mMP<4400
`
`one
`
`
`
`
`
`mmhb<§H.2mmOQmmgommOmHZO_H<Q_DO_4*
`
`
`
`
`
`zo_H<oo44<mmz<muoz_4mz<mm_H<I;*
`
`
`
`Wh_z=oz_4mz<mm_H<I;,
`
`oz69a2_d8:mmwooma
`
`
`mm>mu»>uzmhm_wzoomm».3:mm»
`
`
`
`m¢~awhmohmozmmmuz_>mmhumm<mmpmago
`
`moz_4¢:<m
`
`
`
`
`
`mu»moAmmo:mo_hm_KMHo<m<Io>mx
`
`\IE
`OH.®HM_
`
`
`
`om_u_h<mhmHUMAMm
`
`
`
`mgaz<m20oz<m
`
`AmuHmDJQ>mv
`
`
`
`
`
`
`

`

`U.S. Patent
`
`US 7,120,599 B2
`
`Mwmmm\zoum16.
`nwzo_p<004
`
`
`
`
`
`Emma855%_28Aumo_m¢ouN_4<mmamho_omma_2mo.\w\\\032%.:
`
`
`1Jma»?>Hmuaoma0mm
`
`
`Ammo:wIHmm:_HnoH2300m_44mgIo<m*
`
`
`
`H=mhxo_ua=4moo:
`
`
`
`mJWQOEmJuoozwAmmo:nJuno:NAmmo:
`FAmmo:Gum:mM4m<_z<>
`
`
`Aoupo_ommai4<ako<vzo_Ho_oumm
`
`
`
`
`
`
`m4hwo:\\\\\u\.ammomu4m<_m<>10.x;>mmmum_om4mooz.
`
`
`\\\\\\\\\\\\\\\\v\mommm:::_z_2mIHHummhumm<32*
`
`o4_=m“F4_:mm4moozo
`
`4<m_<mmm<040
`
`
`
`<mm<uQAm
`
`
`
`<mm<oz<4
`
`
`
`AmMHmDAQVgnome
`
`
`
`
`
`
`
`
`nu22m0Juoo:mJMQOEwAmmo:nJMQOENAmooz_Ammo:macawmm<4oPmmm<Ammo:wmhumm<35-202mo;mzo_Ho_ommamz_o<mm><mo;mhxo_mgmmz_zmwpmo*
`
`
`um4<_omMzzoozo_Hog<Hmsoo
`
`m459com
`mw\\\mam?25«820.55%5%888%
`
`
`
`
`
`
`4<_Hzma_mumzo_Ho:<Hmsoo
`
`4<HOH
`
`Ha.DHL
`
`wmm
`
`Now
`
`

`

`U.S. Patent
`
`Oct
`
`10,2006
`
`Sheet12 0f14
`
`9US7
`
`120,599 132
`
`0mm;_2,383222mos:l2%..59:8no8:02
`
`
`
`
`
`02.8on82%m3<>>m8uzo%0mm:_H02.39%82.595152?50:5!“518
`
`
`
`0mm:_n29:80.E:t_C.0.TV20:58E:Ic.3gem20:28
`
`0$5F25:322Smoan.IE:Caszzézaéo
`
` oummd._u8»aSflzééoz<3
`oummd._umu»gz<38552;
`
`.
`
`.v
`
`
`
`NH.®HII_8
`
`
`
`
`
`

`

`U.S. Patent
`
`Oct. 10, 2006
`
`Sheet 13 0f 14
`
`US 7,120,599 B2
`
`
`
` NON-
`REVOLVING
`NON—
`REVOLVING
`
`REVOLVING
`
`SHAKER?
`TREE 11
`
`272
`
`
`
`
`
`REVOLVING
`
`SHAKER?
`TREE 16
`
`
`
` 264
`
`2 52
`
`
`NON—
`ZERO
`
`FIG. 13
`
`

`

`U.S. Patent
`
`Oct. 10, 2006
`
`Sheet 14 0f 14
`
`US 7,120,599 B2
`
`(/»76
`
`DATABASE
`
`DATABASE
`
`SERVER
`
`304
`
`COMPUTER
`
`
`
`
`302-#//
`
`300 ———’//fl/
`
`COMPUTER
`
`304
`
`FIG. l4
`
`

`

`US 7,120,599 B2
`
`1
`METHODS AND SYSTEMS FOR MODELING
`USING CLASSIFICATION AND
`REGRESSION TREES
`
`CROSS REFERENCE TO RELATED
`APPLICATIONS
`
`This application claims the benefit of US. Provisional
`Application No. 60/174,057, filed Dec. 30, 1999, which is
`hereby incorporated by reference in its entirety.
`
`BACKGROUND OF THE INVENTION
`
`This invention relates generally to valuation methods for
`financial instruments and more particularly to rapid valua-
`tion of large numbers of financial instruments.
`A large number of assets such as loans, e.g., ten thousand
`loans or other financial
`instruments, sometimes become
`available for sale due to economic conditions, the planned or
`unplanned divestiture of assets or as the result of legal
`remedies. The sale of thousands of commercial loans or
`
`other financial instruments sometimes involving the equiva-
`lent of billions of dollars in assets must sometimes occur
`within a few months. Of course, the seller of assets wants to
`optimize the value of the portfolio, and will sometimes
`group the assets in “tranches.” The term “tranche” as used
`herein is not limited to foreign notes but also includes assets
`and financial instrument groupings regardless of country or
`jurisdiction.
`Bidders may submit bids on all tranches, or on only some
`tranches. In order to win a tranche, a bidder typically must
`submit the highest bid for that tranche. In connection with
`determining a bid amount to submit on a particular tranche,
`a bidder often will engage underwriters to evaluate as many
`loans as possible within a tranche and within the available
`limited time. Up until the time for submitting a bid is about
`to expire, the bidder will evaluate the loans underwritten at
`that time, and then attempt to extrapolate a value to the loans
`that have not then been analyzed by the underwriters.
`As a result of this process, a bidder may significantly
`undervalue a tranche and submit a bid that is not competitive
`or bid higher than the underwritten value and assume
`unquantified risk. Of course, since the objective is to win
`each tranche at a price that enables a bidder to earn a return,
`losing a tranche due to significant undervaluation of the
`tranche represents a lost opportunity. It would be desirable
`to provide a system that facilitates accurate valuation of a
`large number of financial instruments in a short period of
`time and understand the associated probabilities of return for
`a given bid.
`
`BRIEF SUMMARY OF THE INVENTION
`
`In an exemplary embodiment, an iterative and adaptive
`approach is provided wherein a portfolio is divided into
`three major valuations. Full underwriting of a first type of
`valuation of an asset portfolio is performed based upon an
`adverse sample. A second valuation type is efficiently
`sampled from categories of common descriptive attributes,
`and the assets in the selective random sample are fully
`underwritten. The third valuation type is subjected to sta-
`tistically inferred valuation using underwriting values and
`variances of the first and second portions and applying
`statistical inference to individually value each asset in the
`third portion. Clustering and data reduction are used in
`valuing the third portion.
`
`10
`
`15
`
`20
`
`25
`
`30
`
`35
`
`40
`
`45
`
`50
`
`55
`
`60
`
`65
`
`2
`
`As the process proceeds and more assets are underwritten,
`the number of assets in the first and second portions increase
`and the number of assets in the third portion decreases and
`the variance of the valuation of the assets in the third portion
`becomes more and more defined. More specifically,
`the
`assets in the third portion are evaluated by grouping the
`assets into clusters based on similarity to valuations of assets
`in the first and second portions. Hypothetical bids are
`generated using the valuations to determine an optimum bid
`within parameters determined by the bidder. The optimum
`bid is identified through an iterative bid generation process.
`One method for grouping assets based on similarity uses
`a classification and regression tree analysis of asset portfo-
`lios, where the method includes the steps of defining rel-
`evant portfolio segmentations, assessing performance of the
`classification and regression tree based model against a
`simple model and ranking all portfolio segments based upon
`performance of the models.
`
`BRIEF DESCRIPTION OF THE DRAWINGS
`
`FIG. 1 is a flow diagram illustrating a known process for
`valuing a portfolio of assets;
`FIG. 2 is a flow diagram illustrating valuing a portfolio of
`assets in accordance with one embodiment of the present
`invention;
`FIG. 3 is a flow diagram illustrating, in more detail, one
`embodiment of a first portion of a rapid valuation process for
`large asset portfolios that breaks assets into categories of
`variance;
`FIG. 4 is a flow diagram illustrating a second portion of
`a rapid valuation process for a large asset portfolios that
`aggregates from a basis to a tranche or portfolio basis;
`FIG. 5 illustrates a probability distribution for exemplary
`assets whose recovery value is inferred;
`FIG. 6 is a flow diagram of a supervised learning step of
`the process of FIG. 3;
`FIG. 7 is a flow diagram of an unsupervised learning step
`of the process of FIG. 3;
`FIG. 8 is an embodiment of the process for unsupervised
`learning;
`FIG. 9 is an embodiment of the generation 1 (first pass)
`rapid asset valuation process;
`FIG. 10 is a flow diagram of a fuzzy clustering method
`used in the unsupervised learning of FIG. 8;
`FIG. 11 is a pair of tables showing an example of model
`selection and model weighting for a rapid asset evaluation
`process;
`FIG. 12 is a table showing exemplary attributes for a rapid
`asset valuation process; and
`FIG. 13 is a cluster diagram of an exemplary clustering
`method for a rapid asset valuation process; and
`FIG. 14 is a computer network schematic.
`
`DETAILED DESCRIPTION OF THE
`INVENTION
`
`FIG. 1 is a diagram 10 illustrating a known process for
`valuing a large portfolio of assets 12 through an underwrit-
`ing cycle and through to making a bid for purchasing asset
`portfolio 12, for example, in an auction. FIG. 1 is a high
`level overview of a typical underwriting and extrapolation
`process 10 which is not iterative and not automated. In
`diagram 10, underwriters underwrite 14 a number of indi-
`vidual assets from portfolio 12 to generate an underwritten
`first portion 16 and an untouched remainder portion 18.
`
`

`

`US 7,120,599 B2
`
`3
`Before any of the assets are underwritten, first portion 16 is
`zero percent and remainder portion 18 is one hundred
`percent of portfolio 12. As the underwriting process
`progresses, first portion 16 increases and remainder portion
`18 decreases. The objective is to underwrite as many assets
`as possible before a bid is submitted for the purchase of asset
`portfolio. The team of underwriters continues individually
`underwriting 14 until just before a bid must be submitted. A
`gross extrapolation 20 is made to evaluate remainder portion
`18. The extrapolated value 20 becomes the non-underwritten
`inferred value 24. The gross extrapolation generates a valu-
`ation 24 for remainder portion 18. Valuation 22 is simply the
`total of the individual asset values in first portion 16.
`However, valuation 24 is a group valuation generated by
`extrapolation and may be discounted accordingly. Valua-
`tions 22 and 24 are then totaled to produce the portfolio asset
`value 26. Valuation processes are performed on each tranche
`of the portfolio.
`FIG. 2 is a diagram illustrating one embodiment of a
`system 28 for rapid asset valuation. Included in FIG. 2 are
`representations of process steps taken by system 28 in
`valuating asset portfolio 12. System 28 individually evalu-
`ates (“touches”) every asset, except for a very small quantity
`30 of untouched assets considered statistically insignificant
`or financially immaterial. Specifically, all assets in portfolio
`12 other than quantity 30 undergo an iterative and adaptive
`valuation 32 in which the assets in portfolio 12 are indi-
`vidually valued,
`listed individually in tables and then
`selected from the tables and grouped into any desired or
`required groups or
`tranches for bidding purposes
`(as
`described below.) As in diagram 10, underwriters begin a
`full underwrite 14 of individual assets in portfolio 12 to
`produce a fully underwritten first portion 16 of assets.
`Underwriters also underwrite 34 a sample of assets in a
`second portion 36 of portfolio 12, and a computer 38
`statistically infers 40 value for a third portion 42 of portfolio
`12. Computer 38 also repetitively generates 44 tables (de-
`scribed below) showing values assigned to the assets in
`portions 16, 36 and 42 as described below. In one embodi-
`ment, computer 38 is configured as a stand alone computer.
`In another embodiment, computer 38 is configured as a
`server connected to at least one client system through a
`network (shown and described in FIG. 14), such as a
`wide-area network (WAN) or a local-area network (LAN).
`For example, and still referring to FIG. 2, an unsampled
`and non-underwritten portion 46 of a third portion 42 of
`portfolio 12 is subjected to a statistical inference procedure
`40 using fuzzy-C means clustering (“FCM”) and a compos-
`ite High/Expected/Low/Timing/Risk (“HELTR”) score to
`generate two categories 48 and 50. HELTR is defined as
`HiHigh cash flow, EiExpected cash flow, LiLow cash
`flow, TiTiming of cash flow (for example in months: (L6,
`7718, 19736, 37760), and RiRisk assessment of borrower
`(9iboxer used by credit analysts). Category 48 is deemed
`to have sufficient commonality for evaluation as a whole.
`Category 50 is further divided into clusters 52 and 54 that
`are, in turn, further subdivided. Cluster 52 is divided into
`subclusters 56 and 58, while cluster 54 is subdivided into
`subclusters 60, 62 and 64. Cluster and subclusters are shown
`both in a “tree” chart 66 and as boxes in valuation block 68.
`
`These individual asset values are then regrouped into
`tranches 70, 72 and 74 for bid purposes. Any number of
`tranches could be assembled in any arrangement set by the
`seller.
`
`Individual asset data (not shown) for each asset in port-
`folio 12 is entered into a database 76 from which selected
`
`data 78 is retrieved based on a given criteria 80 for the
`
`10
`
`15
`
`20
`
`25
`
`30
`
`35
`
`40
`
`45
`
`50
`
`55
`
`60
`
`65
`
`4
`
`iterative and adaptive process 32. When criteria 80 is
`established for valuation of any asset, that established cri-
`teria 80 is stored in database 76 for use in valuating other
`asset data in database 76 which shares such an established
`
`criteria. Iterative and adaptive valuation process 32 thus
`develops 82 valuations (described below) and groups 84
`them for use in bidding.
`FIGS. 3 and 4 together form a flowchart 85 illustrating a
`functional overview of one embodiment of system 28
`(shown in FIG. 2) for evaluation of a large asset portfolio 12.
`Valuation procedures 14, 34 and 40 (see also FIG. 2) are
`simultaneously and sequentially used in system 28 in a
`manner described below. As described above, full under-
`writing 14 is a first type of valuation procedure. Grouping
`and sampling underwriting 34 with fall underwriting of the
`samples is a second type of valuation procedure. Statistical
`inference 40 is a third type of valuation procedure, which is
`an automated grouping and automated valuation. Procedures
`14, 34 and 40 are based on objective criteria established as
`described below.
`
`“Underwriting” as used herein means a process in which
`a person (“underwriter”) reviews an asset in accordance with
`established principles and determines a current purchase
`price for buying the asset. During underwriting, the under-
`writer uses pre-existing or established criteria 80 for the
`valuations. “Criteria” means rules relevant to asset value and
`
`a rating based on such categories. For example, as a criteria,
`an underwriter might determine three years of cash flow
`history of the borrower to be a category of information
`relevant to asset valuation and might give a certain rating to
`various levels of cash flow.
`
`Full underwriting 14 is done in two ways, a full cash basis
`manner 86 and a partial cash basis manner 88. Both full cash
`basis manner 86 and partial cash basis manner 88 start with
`sets 90 and 92 of assets that are fully individually reviewed
`14 (see FIG. 2). Such full review 14 is usually due to the
`large dollar, or other appropriate currency, amounts of the
`assets being reviewed relative to other assets in the portfolio
`or due to the borrower being so well known or so reliable
`that the assets can be quickly and reliably fully underwritten
`or the assets are marked to market such that there is very
`little variance associated with the value of said assets. Asset
`
`set 90 is evaluated by underwriters 94 and each asset in set
`90 receives a valuation with very little variation such as an
`asset backed with cash or a tradable commodity with full
`cash value and is placed in a full value table 96. Selected
`individual values for assets in table 96 are stored as a fully
`underwritten group value 98.
`Set 92 is evaluated by a team of underwriters 100, which
`could be the same as team 94, but each asset receives a
`discounted or partial value and is placed in a partial value
`table 102. Selected individual values for assets in a tranche
`
`in table 102 are stored as a partial value fully underwritten
`group value 104. Criteria 80 (shown in FIG. 2) for full cash
`basis manner 86 and partial cash basis manner 88 are stored
`in database 76 (shown in FIG. 2) in a digital storage memory
`(not shown) of computer 38 (shown in FIG. 2) for use in
`supervised learning 206 and unsupervised learning 208 of
`automated valuation 40.
`
`Sampling underwriting 34 is accomplished using two
`procedures, a full sampling 106 procedure and a partial
`sampling 108 procedure. Full sampling 106 is utilized for
`categories of large assets and includes a one hundred percent
`sampling 110 of the sample groups in the categories of assets
`being sampled. The assets in full sampling 106 are not
`individually underwritten but rather are underwritten in full
`sampling groups 112 based on a determined commonality. A
`
`

`

`US 7,120,599 B2
`
`5
`resulting full sampling group valuation (not shown) is
`created and then desegregated based on a rule 114 to
`generate an individual full sample asset value table 116.
`Individual full sample asset values in table 116 are then
`uploaded electronically into any full sampling group valu-
`ation 118 required for bidding as suggested by the grouping
`of assets in a tranche. The number of assets in an under-
`
`writing sample grouping can be as little as one to any
`number of assets. Partial sampling 108 is for medium
`categories of assets and includes forming a cluster sample
`group 120 by one hundred percent sampling of a represen-
`tative group from within a cluster of the groups being
`sampled and random sampling of the other groups in the
`cluster. In partial sampling 108, all groups are sampled, but
`some are partly valued by extrapolation from cluster sample
`group 120. Partial sampling 108 includes an asset level
`re-underwrite 122 with manual data entry 125 to produce an
`alpha credit analyst table 126 which is given an asset class
`adjustment 128 to produce an adjusted credit analyst table
`130. As described above, individual assets are selected from
`adjusted credit analyst table 130 according to tranche group-
`ing to produce a partial sampling credit value 132 for use in
`bidding on tranche 70 (shown in FIG. 2).
`Automatic valuation procedure 40 utilizes supervised
`learning process 206, an unsupervised learning process 208
`and an upload from a statistical inferencing algorithm 134 to
`generate an underwriting clusters table 136 which is stored
`in a digital storage device. In supervised learning process
`206, an experienced underwriter who knows what questions
`to ask to establish value, assists the computer in determining
`whether or not an asset is a good investment and how to
`value the asset. In unsupervised learning process 208, the
`computer segments and classifies assets and objectively
`self-evaluates the assets based on feedback from the data. An
`
`underwriter periodically reviews the unsupervised learning
`process 208 to determine whether the computer is making
`sensible underwriting conclusions. The computer uses sta-
`tistical algorithms 134 to make its inferences. For example,
`but not by way of limitation, one embodiment uses the
`Design For Six Sigma (“DFSS”) quality paradigm devel-
`oped and used by General Electric Company and applied in
`a Due Diligence (“DD”) asset valuation process using a
`multi-generational product development (“MGPD”) mode
`to value the asset data with increasing accuracy. Learning
`processes 206 and 208 incorporate the accumulated knowl-
`edge as the valuation progresses into cash flow recovery and
`probability of recovery calculations on an ongoing, real time
`basis. Supervised learning process 206 uses business rules to
`identify clusters of assets having common aspects for valu-
`ation purposes. Unsupervised learning process 208 uses
`feedback from prior data valuations performed by procedure
`40 to determine if progress is being made with respect to
`increasing valuation confidence. Identification of all avail-
`able raw data and discovery of interrelationships of clusters
`of these available raw data is possible due to the use of
`high-speed computers, as is described below.
`In one exemplary embodiment, a fuzzy clustering means
`(“FCM”) process of unsupervised organization of raw data
`using a HELTR scoring technique is employed to infer
`valuations of credit scores onto assets in portfolios, as
`described below. Such clustering techniques have been
`developed in response to more sophisticated classification
`segments to describe assets and high asset counts in port-
`folios that must be assessed in time periods that do not allow
`manual processing.
`One exemplary method first organizes valuation scores
`(static and/or probabilistic recoveries) in a computerized
`
`6
`system. Adjustments are then made to the valuation scores
`for special factors and business decisions. Then a reconcili-
`ation of multiple valuation scores describing the same asset
`and an overall adjustment to interview/override the inferred
`valuation is performed.
`Organizing valuation scores is performed by collating, in
`electronic form, a cluster number, a cluster name, descrip-
`tive attributes of the cluster(s), probabilistic recovery values
`(an illustrative example is a HELTR score) and the under-
`writer’s confidence in each cluster’s valuation based upon
`the strengths of each cluster’s descriptive attributes. The
`cluster number is a unique identifier of a specific set of
`descriptive attributes that are facts about an asset which a
`person skilled in evaluations uses to assess value of an asset.
`Examples of descriptive attributes include, but are not
`limited to, payment status, asset type, borrower’s credit
`worthiness expressed as a score, location and seniority of a
`claim. The cluster name is, in one embodiment, an alpha-
`numeric name that describes the cluster’s descriptive
`attributes or sources. One example of descriptive attributes
`is found in FIG. 12, described below.
`Descriptive attributes are the facts or dimensions or
`vectors that were used to develop the asset’s value. Com-
`puter logic is used to check for replicated clusters, if any, and
`alert the analysts or underwriters.
`Because each asset can be described by many combina-
`tions of descriptive attributes, various levels of value for the
`same asset may occur. Probabilistic recovery values or credit
`score or any numerical indication of the asset’s worth are
`indicators of worth designated at the discrete asset level. All
`of the information from the various descriptive attributes is
`synthesized such that a purchase or sale price can be
`ascertained as a fixed value or a probabilistic one. An
`illustrative embodiment used herein is the HELTR score.
`
`Each cluster has a unique set of descriptive attributes and
`designated HELTR score.
`Every cluster’s unique attributes contribute to a valuation
`of cluster value. Different combinations of attributes provide
`a higher confidence or confidence interval of a particular
`cluster’s score. For example, if any asset was described as a
`green piece of paper with height equal to 2.5" and width
`equal to 5"7one might ascribe a value of 0 to 1000 dollars
`and place very little confidence in this assessment. If this
`same asset was described with one more fact or attribute or
`
`vector as being a real $20 US bill, one would place a very
`high confidence factor on this cluster value of $20 US
`dollars.
`A cluster’s valuation and confidence is determined at a
`
`point in time and recorded. Sometimes new information
`becomes available and the analyst would like to alter the
`value(s). The value is altered manually or automatically with
`a data field and decision rules, in the automated fashion via
`computer code. The prior values are manipulated to reflect
`new information. As an illustrative example, assume the
`prior cluster confidence was recorded at 0.1 and it is learned
`that a different asset with exact descriptive attributes as in
`this cluster just sold for over the predicted “most probable”
`value. Rules were in effect such that if this event occurred,
`cluster confidence is multiplied by 10. 0.l><10:l which is
`the revised cluster confidence.
`
`10
`
`15
`
`20
`
`25
`
`30
`
`35
`
`40
`
`45
`
`50
`
`55
`
`60
`
`The purpose of such a process is to reconcile multiple
`scores for the same asset, controlling for the confidence
`associated with each source of valuation of each dimension
`
`65
`
`of valuation. Using the HELTR as an illustrative example
`with sample data points on a particular asset:
`
`

`

`7
`
`US 7,120,599 B2
`
`High
`85
`
`Exp
`62
`
`Valuative
`Low Time Confidence High
`.15
`3
`3
`
`(.3/1.65)(.85)
`
`Exp
`
`Low
`
`Time
`
`(.3/165)(.62)
`
`(3/1.65)(15)
`
`(.3/1.65)(3)
`
`45
`
`.4
`
`31
`
`.9
`
`.5
`
`.2
`
`3
`
`2
`
`7
`
`65
`
`(.7/1 65)( 45)
`
`(7/1.65)(.4)
`
`(7/165)(31)
`
`(7/165)(3)
`
`( 65/1.65)( 9)
`
`(.65/1.65)(.5)
`
`(.65/1 54)( 2)
`
`( 65/1.65)(2)
`
`1 65
`
`6999
`
`.4792
`
`.2374
`
`2.6059
`
`2
`
`Cluster Cluster
`Number Name
`1
`Lien
`positions -
`recourse
`Asset
`classification -
`industry -
`age
`Coordinates -
`use -
`borrower
`x
`
`3
`
`n
`
`The cluster consensus valuation is a high value of 0.6999,
`most likely 0.4792, low 0.2374 with a timing of 2.6059.
`Different logic can be applied to manipulate any of the
`weights.
`The consensus scores are developed in the context of
`global assumptions. Should a global assumption change
`occur, process steps 128, 138 are included in the method-
`ology to weight the consensus score. Illustrative examples
`are fraud discovery in certain valuation factors, macroeco-
`nomic changes, fangible market value established for an
`asset class, and loss of or increase of inferenced asset
`valuation methodologies relative to other methodologies
`being employed.
`In another embodiment, a cross correlation tool is used to
`quickly understand and describe the composition of a port-
`folio. Typically, the tool is used to correlate a response of a
`user selected variable versus other variables in an asset
`
`portfolio. The tool quickly identifies unexpectedly high or
`low correlation between two attribute variables and the
`
`response variable. Attribute variables are of two types,
`continuous and categorical. The cross correlations are com-
`puted by the correlation tool between all variables of interest
`and their bin or level and presented, in one embodiment, in
`a two dimensional matrix for easy identification of trends
`amongst the assets in the portfolios.
`First, the cross-correlation tool identifies attribute vari-
`ables in the portfolio of assets as one of continuous or
`categorical. For each variable aggregation levels are com-
`puted by bins for continuous variables and by value for
`categorical variables.
`A user looking to identify correlations with the tool will
`select a response variable, Y,, for example, an expected
`recovery or count. For all combinations of pairs of attribute
`variables (x1 and x2) and their levels (a and b), compute the
`average value of the response variable, Y,, according to:
`Yr:sum(Y(x1:a and x2:b))/count(x1:a and x2:b).
`
`An expected value, Y
`calculated according to:
`
`expect:
`
`of the response variable is
`
`Yexpect:(sum(Y(x1:a))*count(x1:a)+sum(Y(x2 :b))*
`count(x2:b)))/count(x1:a)*count(x2 :b)).
`
`A deviation, Yemr, of the chosen response variable, Y,,
`from the expected value, Yexpect, using weighted values of
`occurrence of x1:a and x2:b separately, is calculated by:
`Ymay: Yr- Yexpect'
`
`In one embodiment, expected values and deviations are
`displayed in multi-dimensional displays to make variations
`from expected values easy to identify.
`
`20
`
`25
`
`30
`
`35
`
`40
`
`45
`
`50
`
`55
`
`60
`
`65
`
`In another ex