FNBM Payment Float Strategy Model Development
Transcription
FNBM Payment Float Strategy Model Development
,l I l I i'-, ! ; i i I B \: r i rffiionut BANK OF MAR'N FNBM Payment "Float" Strategy Model Development Amelia Brakefield Robert Gladd Don Blackhurst FNBM CorporateRisk Management May 2002 rffiiorul BANK OF MAR'N FNBM Payment t'Float" Strategy Model Development Amelia Btakefield Robert Gladd Don Blackhurst FNBM Coqpomte Risk Management Mav 2002 Executive Summary This paper documents a study undetaken to derive a datz-diven risk-based strategy for attenuating the aggregate adverseimpact of catdholder payments returned for "non-sufficient funds" (f{SF) by identi$'ing payment and account stanrs criteria predictive of elevated likelihood of NSF teturn. Using logistic regression and Classification and Regtession Ttee (CART) methods, such elements were found to include delinquency stage at time of payment, payment amount and type, Payment magnitude as a percentage of credit line, recent priot history of NSF returned payments, and number of close proximity (same cycle) payment attempts. The study yielded a set of payment "floaf'l criteria that appear to be significantly more effective than the current strategy deployed in the FDR Payment Defender system. Relative "effectiveness" is evaluated in terms of the proportions of "good" payments floated f no floated versus "bad" payments floated /not floated. A floated "good" payment (one honored by the issuing source) is a "false positive," whereas a payment not floated yet returned as NSF is a "false negative." False positives rePfesent ttansient lost utilization (and thereby income) oppotunity, and can result in catdholdet dissatisfaction with the unavailability of good funds. False negatives represent elevated risk in terms of additional unwarranted cedit availability. Th. goal is to minimize both types of classification errors to the extent pmcticable.2 NSFrecenthistoricalsummary and culminatingin April, 2002,monthly NSFs have Dnring the period beginningwith January,2001, of the number of paymentstemitted and 5.5o/oof paymentdollars compdsedapproximately3.8o/o (seeFigure2 andTable 2hercinafter).Fot the fust fout months of 2002,NSFs have averaged approximately$1.47million per month (equivalently,-$17.6 million annualized). 1 Fnnds unavailable to the cardholdet for a specified hold time' 2 Statisticians refer to the "sensitivity" versus "specificity" of predictive measures. We could, of course, float all payments, a tactic that would hold all bad payments (perfect sensitiviry), but at the unacceptable cost of also holding all good ones (zero specificity)-the vast maiority of cardholder payments. Given a low prevalence circumstance-e.9., our io"grtly 3-4%monthly incident NSF rate-at once achieving high specificity and sensitivity becomes problematic. Consequently, sensitivity/specificity rnisclassification trade-offs must be evaluated in terms of estimated quantitative risk-cost/benefit analysis, and we did so in concluding this study. 'v li: FNBMNSFFloat Model Development NSF|'Jrn 2n/0l-AprilAna $1,800,000 $1,600,000 $l,4m,ooo $1,2oo,ooo f1,0m,000 $800,000 $600,000 $,loo,ooo $200,000 $o CCo.p g*-r-.tg ,.,f-f -f CgrT,'""..$".tr.o,S ""p Figure1: NSFdollarsby month,January2001- April2002 NSF incidcnt rrtc w. S ntc 8/e 7o/o 6% 5o/o 4% 3Vc T/o lo/o V/o ,.,f -,Sa,,f ,*A+of.,.'$r"" -^f, CC "S oo$g" oo*+pe,.*P of numberof monthlypayments(lower,orangeline) Flgure2: NSFsas a poroantagp andbbl amountof rnonhlypayments(uppergreā¬nline).Dataftttedwithpolynomialtrendlines -2- . FNBMNSFFloatModelDevelopment Jan-O1 Feb-01 Mar-01 Apr-01 May-O1 Jun-O1 Jul-01 Aug-01 Sep-01 Oct-01 Nov-O1 Dec-01 Jan-02 Feb-02 Mar-02 Apr-02 496,955 570,467 578,347 525,1,02 534,182 548,710 565,163 588,818 603,130 619,959 622,597 630,958 631,,209 679,037 61,4,782 288,968 272,527 300,450 283,684 284,270 282,728 290,809 291,,134 285,313 324,242 305,144 302,296 378,749 306,623 322,907 310,550 58.1.5o/o 53.39o/o 57.960/o 54.02% 53.22% 51,.530/o 51.460/o 49.44o/o 47.31.o/o 52.31,o/o 49.07o/o 47.91.o/o 50.50% 49.53o/o 52.520/o 50.81,o/o 1,.1,9 1,.1,3 1.18 1..27 1.15 1,.1,4 1.15 1,.1,4 1,.1,2 1.18 1..1,4 1.15 1,.16 1.13 1,.1,7 1,.1,6 343,891, fi20,41,5,1,97.37$59.37 308,920 ff21,,502,955.68 $69.61 355,925 $23,524,1,21,.98$66.09 361,,646 $23,386,096.26 fi64.67 325,783 $21,055,600.19 $64.63 323,230 $21,,275,288.40$65.82 334,1,82 fi21,,571.,231.67$64.55 331,,545 fi22,352,085.21,fi67.42 31,9,304 $21,,31,2,592.11 $66.75 1,,1,70.25 $66.57 384,1,25 fi25,57 80.64 $68.40 348,871, $23,863,7 348,603 fi22,755,357.77 $65.28 370,21,0 $ 2 5 , 0 5 4 , 1 5 7 . 8 3$ 6 7 . 6 8 347,934 $27,899,892.67 $80,19 377,827 $28,061 ,893.94 ff71.27 361,.076 $26,045,210.90 ff12.13 5,543,072 $375,646,632.87 fi67.77 datia,January2001- April2002.Source:ME Histtables. Table1: Payments Jan-01 Feb-O1 Mar-01 Apr-O1 May-01 Jun-O1 Jul-O1 Aug-01 Sep-O1 Oct-01 Nov-01 Dec-01 Jan-02 Feb-02 Ma*02 Apr-02 72,953 10,226 11,538 11,,269 17,407 12,454 12,674 14,559 12,573 15,657 14,047 '1,3,768 16,184 12,326 12,226 212,535 fi7,066,235.75 $82.32 $917,658.32 $89.74 $1,040,597.70 $ 9 0 . 1 9 fi961,999.43 $8s.37 $1,490,007.93 $85.02 $1,158,977.85 $93.06 fi1,,21.6,077.A3$9s.9s $7,335,421.67 $91.72 $1,,177,641.09 $93.66 $1,496,765.08 $9s.60 fiL,579,844.26 $1,1,2.40 $1,,486,932.77$108.00 fi1,597,31,0.47 $98.70 fft,474,074.13 $ 1 1 9 . 5 9 $1,479,998.48 $120.96 fi1,,322,987.76 $104.39 790,429.06 {97.82 Table2: NSFdata,January2001- April2002.Source:ME Histtables -3 - 3.770/o 5.220/o 3.31,oh 4.27Yo 3.24Yo 4.42oio 3.1.2o/o 4.11.o/o 5.34%o 7.030h 3.85o/o 5.45o/o 3.79o/o 5.640/o 4.39Yo 5.97%o $19,348,961.62 $20,585,297.36 $22,483,521,28 $22,424,096.83 $19,575,592.26 $20,116,310.55 $20,355,154.64 $21,016,663.54 $20,134,951.02 3,940/o 5.53o/o 4.080/o 5.85o/o 4.03o/o 6.62Yo 3.95%o 6.53o/o 4.37%o 6.38o/o 3.54o/o 5.28Yo 3.24o/o 5.27%o 3.5"1,o/o 5.08% fi24,074,405.17 fi22,284,936.38 ff21,,268,425.06 $23,456,847.36 fi26,425,818.54 $26,582,995.46 $24,722,223.74 3.830/o 5.53o/o $354,856,203.81 FNBMNSFFloatModelDevelopment Paymentand NSFtypes \Ufeassembled a six-month payment and NSF history dataset from the FDR TranHist files spanning October 2001through Febnrary 2002. After retaining only exact dollar NSF matches to priot payments, we wer'e able to tabulate the following estimated NSF distributions by incident and dollar rates, along with statistical days-to-post intervals.3 Direct Check Check Certified Funds MoneyGram Western Union $ 4,930,485.69 63.88o/o 68.72o/o $ 110.78 21,970 fi 2,017,453.833',t.450/o 28.12o/o $ 92.08 4.509 2,370 $ 126,886.42 3.40o/o 486 $ 44,300.84 0.70o/o 357 52,751.75 0.510h $ 475.00 0.030/o $ tl 351.00 0.02o/o AutoPav $ 0.07o/o 1,578.15 9 AsencY Payment $ Xffib 69,6tt $ 7,17438268 Tell-a-Friend t9 7.77o/o $ s3.s4 0.620/o $ 91.1s 0.74o/o $ 1,47.76 0.01,o/o $ 2s.00 0.00o/o $ 31.91 0.02o/o $ 1,75.35 $ rc2.97 Tabfe3a: NSFdataby remittancetype,October2001- February2002. Direct Check 1 4 5 6 6 8 Check 1 8 10 1.2 1,4 18 Certified Funds 1 3 5 8 14 3 4 7 1,1 MoneyGram 1 2 2 Western Union 1 1 1 2 4 15 Tell-a-Friend 3 10 1,7 29 37 AutoPay 3 5 18 18 Agency Payrnent 10 35 64 720 37 26 t20 26 t20 T*lN 2002. 2001- February type,October daysto postbyremittance Table3b:NSFdatia, Direct Checks ('DCs'), private checks, and Certified Funds payment methods comprise nearly 99o/o of the NSF experience, with DC payments alone accounting for approximately t'wo thuds of NSFs (-640/0 of incident NSF tate, -690/o of NSF dollars). DCs account fot approximately 25o/oof overall remittances. Consequently, the DC payment is significantly more likely to be returned NSF than othet payment methods 3 Current float strategy holds all designated payments for 10 days irrespective of remittance tyPe. -4- FNBMNSFFloatModelDevelopment Dataanalysisand modeling At the outset of this study, March 2002 payment and NSF data were the most recent available,and so were chosen to comprise the development study set. Data were assembled from the FDR "short month-end ME Hist" table HIST0302.dbf, a dataset containing number and dollar amounr of payments and NSFs per account, as well as balance and credit line amounts. Only accounts showing one or more remittances in March were retained for the study. To these data we added, from the Febnrary 2002 FDR full month-end file, account attributes such as ACS Behaviot Score, FICO NextGen Scote, delinquency bucket, months on books, and a tally of any NSFs (incident) for the most recent 4 months,'1,2 months, and lifetime. The NSF history was obtained from the Risk Group SAS DelHist0302 portfolio history table. Utilization (balance/line), payment-percent-of-balance,"dollars-at-risk" (balance-savings)and payment-percent-of-line were also calculated.After the metging of the several source files, 295,932 account records were retained. 9,024 accountshaving at least one NSF in March were taggedas "bad" (-3.05o/o).Good/Bad statistics are sutnmanzed in Table 4 below: 0.609 0.1,97 # of NSFs last 4 rnonths 0.459 0.087 # of NSFs last 12 months 0.834 0.209 # of NSFs accountlifetime # of prior NSFs this cycle 1.248 0.400 0.091 0.000 # of payment attempts this cycle 0.015 0.001 Months on books 1,9.7 24.9 FICO NextGen Score 516 550 FDR Behavior Score 560 620 fi104.44 fi70.77 $5s2.33 $4e0 fi463.92 $365.45 0.81,2 $286.61 0.755 0.231 0 . 15 3 0.153 0 . 18 4 # of cvcles delinquen Pavment amount Balance prior to payment Credit line Dollars at risk (balance-savinEs) Utiliz2lien Pavment o/oof ltne Pavment o/oof balance $508 Table4 Univariate statistics for the Good (no NSF) group were uniformly more favorable than were those pertainingto the "Bad" (9,024 NSF) Soup.However, while many of these variables were found to be statistically"significant'according to a subsequentSAS logrstic regression trial, the resulting regression model (using these variables to predict the likelihood of making a good payment) was found to have a weak "adiusted R-Square" value of 0.1114, owing to pervasive drstnbuuonal nonnormalities (i.e. ske*) and ovedap between Good/Bad predictor values. In short, a logistic regressionmodel employing these predictors would yield a hrgh misclassification r^te. -5- FNBMNSFFloatModelDevelopment CARTmodelingtrials The Classification and RegtessionTree (CART) methodology is frequendy employed for explo rz,tory data analysis,and was used in this snrdy. A sgruficant advantage of CART is that it is considered a "non-parametdc" ptocedute not bound by the critical data normality assumptions of Least-Squares analytical and modeling methods such as regression.Moreovet, CART can operate of a mix of categorical, ordinal (tank-ordered) and intenralf rao.olevel data. CART is a "ttee-based" method also known as "binaq/ recursive partitioning" in which the dependent variable must be expressed as a dichotomous outcome (hence "binary")-in this case Good=1 (no NSD versus Good=0 G.{SD.CARTiteratively attempts to find the optimal separation of Good /Bd outcomes as a function of the ptedictor variables, splitting the data into hierarchical "tree nodes" (hence "tecursive partitiooitg') using the strongest predictor each time until pre-set node size limits are reached. The splits tesulting from the final iteration are called terminal nodes. One terminal node will conain the highest proportion of Good casesfound (no NSF), and another will contain the lowest propottion of Goods. An example of our S-Plus CART output follows:o ( 1 ) r o o t 2 9 5 9 3 2 8 7 . 1 9 . 0 00 . 9 6 9 5 ( 2 ) S U i l . N S F 4 < o . s2 7 3 2 7 2 6 2 4 8 . 0 0 0 . 9 7 6 6 ( 4 ) D C < 0 . 52 1 0 0 8 1 2 6 0 7 . 0 0 0 . 9 8 7 4 ( 8 ) D E L . C Y C0<. 5 1 8 8 4 5 0 1 5 9 6 . 0 0 0 . 9 9 1 5 (16) 830202<630.5 59222 825.20 0.9859 * (17) 8S0202>636.5129228 768.40 0.9940 . ( 9 ) D E L . C Y C > o .251 6 3 1 9 8 1 . 0 0 0 . 9 5 2 4 ( 1 8 ) A V G . P i l T < 2 s . 94 21 8 1 6 9 . 7 9 0 . 9 8 3 0 * ( 1 9 ) A V G . P I T > 2 5 . 9 21 7 4 5 O 9 0 6 . 3 0 0 . 9 4 5 0 (38) 850202<223.51688 156.10 0.8969 * ( 3 9 ) B S O 2 0 2 > 2 2 3 . 51 5 7 6 2 7 4 s . 9 0 0 . 9 5 0 2 * ( 5 ) D C > 0 . 56 3 1 9 1 3 5 3 5 . 0 0 0 . 9 4 0 5 ( 1 0 ) P P 7 < 0 . 56 2 9 3 8 3 4 4 2 . 0 0 0 . 9 4 1 9 (20) ltoB<lo.65 26183 1781.00 0.9266 ( 4 0 ) U T I L < I . 0 5 3 0 52 1 9 0 9 1 3 5 4 . 0 0 0 . 9 3 3 8 ( 8 0 ) A V G . P t t T < s 8 . 4 3 51 3 7 6 1 7 2 5 . 1 O O . 9 4 4 2 ' ( 8 1 ) A V G . P M T > 5 8 . 4 83 15 4 8 6 2 4 . 9 0 0 . 9 1 6 3 * ( 4 1 ) U T I L > I . 0 5 3 0 54 2 7 4 4 1 9 . 9 0 0 . 8 8 9 6 ' ( 2 1 ) i l O B > 1 0 . 6 53 6 7 5 5 1 6 5 0 . 0 0 0 . 9 5 2 9 (42) SUil.NSF<1.5 30843 1209.00 0.9591 ( 8 4 ) U T I L < 1 . 0 8 3 0 52 5 6 9 1 9 2 6 . 0 0 0 . 9 6 2 6 ( 1 6 8 ) A V G . P i l T < 6.'6r 4 3 4 1 7 7 2 9 5 7 1 . 3 0 0 . 9 6 6 7 ( 1 6 9 ) A V G . P M T > 6. 614 3 4 7 9 6 2 3 5 3 . 7 0 0 . 9 5 3 4 * (85) UTIL>1.08305 5152 280.80 0.9422 ' ( 4 3 ) S t i l . N S F > l . 55 9 1 2 4 3 4 . 3 0 0 . 9 2 0 2 * ( 1 1 ) P P 7 > 0 . 52 5 3 61.25 0.5889 * (3) SUil.NSF4>0.5 22660 2322.OO 0.8841 (6) 8S0202<585.55381 855.40 0.8017 * ( 7 ) 8 S 0 2 0 2 > 5 8 5 . 51 7 2 7 9 1 4 1 8 . 0 0 0 . 9 0 9 8 ( 14) 0C<0.5 9778 s64.90 0.9384 ' ( 1 5 ) D C > 0 . 57 5 0 1 8 3 4 . 9 0 0 . 8 7 2 4 . aThe parentheticalnumbers on the left indicate the splits.Asteriskson the right denote terminal nodes.This trial produced 16 terminal nodes.The highestproportion of Goods is found in node 17. The lowest proportion is found in mode 6. Interpretation: no NSFs in pdor four months (node 2), not paying by DC (node 4), account current (node 8), and a Behavior Scoreof 637 or better (node 17) yielded 99.4ohGoods. Conversel/, one or more NSFs in the prior 4 Goods. months (node 3) followed by a Behavior Scoreof 586 or lessresultedin a terminus containing only 80.1,7o/o -6- FNBMNSFFloatModelDevelopment The following is a gtaphical depiction of the CART decision tree paths to the best and worst 'al terminal nodes. This t resulted in 16 terminal nodes. The proportions of Good accounts in the remaining 14 not displayed here vry within the worst-to-best (80.17o/oto 99.4o/o)percentages illustrated below ( Nurter of NSFs in prior four npnths ) (Rerittance type) Cycles delinquenl Behavior Score Figure3:CARTtrialtreepathsummary The CART method is hierarchical in that it chooses the strongest predictor of Good/Bad separation on each iteration for the data remaining in the path aftet a prior split (starting, of course with the entfueset. In this instance, NSFs in the prior four months was strongest on the Fust pass). The number of splits and terminal nodes in a CART ta:ralarc govemed by researcher-supplied parameter conftols such as [1] minimum number of observations prior to a node split, [2] minimum number of observations in any tree node, and [3] a factor called "minimum node deviance." Tradeoff judgments must be made tegatding Good/Bad separation and node size. In other words, an unconstrained CART trial would result in perhaps hundreds or thousands of terminal nodes, a relative few of which would contain either all Goods ot all Bads, but all of which would have very -7- FNBMNSFFloatModelDevelopment few observations. Whete the objective entails yes/no decision criteria such as that required for our float/do not float sttategfr an overly latge and too-detailed tree would serve no useful purpose. Gandidatefloat models The initial statistical andyses heretofote described provided ditection for candidate float sffategy modeling. After dozens of logistic tegtession and CART sessions,a number of predictors showed promise, including o o o o o o o o Time on books; Payment amount; Recent prior NSFs; 'lization; U DC vetsus non DC remittance; Multiple payments this cycle; Payment o/oof bdance or limit; Behaviot Scote Given that our month-end daa do not provide us with direct account-level payment floatf no-float identifiers, we 6rct had to dedve an estimate of the existing FDR Payment DefenderrM float strategy now in ptoduction as they would apply to the Match 2002 study set. The existing float criteria are summarized as follows: IF paynent amount is >=$100Al,lD If External Status is A or I f In te rn a l S ta tu s n o t ' Bl a n k' or If # days since last paymentis less than 8 on If Balance before payment / Credit Line is >=125t on If Account has had more than 3 NSFsin the last 4 months or If Account was more than one cycle delinquent befone payment OR Account was 1 cycle delinquent before paymentand Account has already had an NSFthis cycle OR Time on Books is <=6 months and Account is less than 2 cycles delinquent and Account has had an NSFthis cycle OR Account is less than 2 cycles delinquent and Account has not had any NSFsthis cycle and Paymentis >=50rbof the credit limit OR Time on Books is >6 and <=12 and Paymentis >=50t of the credit limit OR Time on Books is >12 and Paynent is >=100t of the cnedit limit FLOATPAYMENT -8- FNBMNSFFloatModelDevelopment Subsequently,we estimated the follo*ing floatf no float segmentation. Bad/no float and Good/f7oat quadrants constitute misclassification effors. 295.932 Pct of all floated 1,0.91,o/o Pct of Bads floated 22.33o/o Pct of Goods floated 1,0.55o/o Table5 Thus fat the discussion has focused on the incidence of NSFs. A strategy focused solely on improving the "catch r^te" of bad payments while mirumrzrng the floating of good payments, however, might well fall short in terms of aggregateNSF dollars held relauve to good payments floated-and consequently net value to the bank. For example, while the current strategy is estimated to catch only 22.33o/oof the bad payment incident rate, it floats better than 57o/oof the NSF dollats. See table 6 below. NSFs not floated NSFs floated Goodsnot floated Goods floated 7,009 2,075 256,646 30,262 480,574.83$ 68.57 O+8,459.10$ 321.82 15,821,893.74$ 61.65 8,723,367.73fi 268.43 $ $ $ $ 5l.43Yo 33.92o/o Table6 Dozens of predictor combinations were tried (seelist on page 8), gurded by CART analyttcaloutput, with the obiective of at once maximizing NSF dollars floated and minimiznggood payments floated while also minimizing the incident rate of good paying accounts floated. A model emerged that corectly identified better than 56Yoof the incident NSFs and approximately 78o/oof the NSF dollars while floating slightly fewer good payment dollars. Model NSFs not floated Model NSFs floated Model goodsnot floated Model soods floated 3,962 5,062 237,116 49,792 $ $ $ $ 249,475.24 $ 62.95 879,678.69 $ 173.77 76,434,264.39$ 69.31 7,510,990.48 $ 150.85 77.970/o 31,.37o/o Table7 The proposed alternative model cdteria are sftaightforwatd: float where [1] an account is 30+ days delinquent, or plhad more one or more prior payment attempts this cycle, or [3] a DC payment of rnore than $75, or [4] there was at least one NSF during the prior 4 months, or [5] a DC payment and account is 5+ days delinquent, or [6] at least one prior NSF this cycle, or [7] payment is great than 150% of line, or [8] payment is mote than $1,000.5 5 A cost of this strategy is the total number of accounts floated: 32,277 under the current procedure (10.9%) versus 54,854 (18.570) under the new model. -9 - FNBMNSFFloatModelDevelopment Validationsamples We appliedthe float model criteria [anuary / Februaryattributes) to February 2002 payment data, with the following outcome (comparedto the current procedure outcome estimates): Model badsnot floated Model badsfloated Model goodsnot floated Model soods floated Default badsnot floated Default badsfloated Default goodsnot floated Default qoodsfloated 2,622 5,464 189,462 55,524 5,818 2,268 210,148 34.838 $ $ $ $ $ $ $ $ 276,325.96 $ 777,683.63 fi 12,613,947.57 $ 8,600,318.14 $ 345,066.65 $ 648,942.94 fi 71,433,485.75 $ 9,780,773.90 $ 82.50 142.33 66.58 154.89 59.31 286.13 54.41 280.75 Table8, February 2002payment data Finally, we merged three months of payment data covering February, March, and April 2002, keeping only obsenrations with at least one payment attempt during the period (n=325,483). We then applied the float model criteria using the combination of January and February attributes to determine how well the model would predict at least one NSF dunng the February-Apnl period, The findings are set forth below in table 9. Model bads, not 0oated Model bads floated Model goods not floated Model goods floated 17,240I 10,632$ 247,723$ 61,888$ 325,#3 1,760,029.06 1,660,386.75 47,853,432.40 15,402,136.25 41.130h 58.87% 73.10% 26.900/0 57.39o/o 48.610/o 79.620/o 20.38% Table9 The NSF prevalence in this 3-month period is 6.720/oQ1,,872/325,483).The model identifiec approximately 49o/oof the incident NSF and 59o/oof the NSF dollars while floating approximately 20.4o/oof the good payment dollars.6We compare these findings to the estimate yielded by the cunent float strategy over the same three-month period (table 10). Current bads not floated Current bads floated Current goods not floated Curent goods floated 17,980$ 3,892 $ 270,996 $ 32,615 $ 7,743,235.08 1,077,180.13 42,468,223.59 74,787,345.06 61.81% 38.790/o 74.770/o 25.83%o 82.27o/o 1,7.79o/o 89.260/o 1,0.74"h 325,483 Table10 In these data, the new model floats appreciably more NSF dollars while holding significandy fewer good payments relative to the opetative strategy. 6 Noteworthy here is that, while the recent average monthly NSF incident rate is slightly less than 4o/o(see table 2), the likelihood of an NSF across this 3-month period is closer to 7o/o _1,0_ FNBMNSFFloatModelDevelopment The proposed float strategy provides approximately a 48o/olift ovet the current procedure in the monthly development and validation samples and neady a 38o/olift with respect to the 3-month sample (seetable 11). 'Ufhile the proportion of actual NSF accounts ("bads") floated to total accounts floated might seernlow+.g.,9.23o/o in the March sample model-, it is important to keep in mind that the incident prevalence of NSFs in the Match data, for example, is only 3.05o/o.ln other words, the model NSF float captute is three times the rncidence of NSFs in the data. Floctcd Good Bad Developrncot sample 30,262 2.015 34,838 2,268 7o bads floated 6.24% 6.77o/o Vdidation sample 3-month sarnple 32,61.5 3,892 '1,0.66o/o Moddsd Stratesy Floated Good Bad 7o bads floated Development sample Validation sample 49,792 55,524 5.062 9.23% 5,464 8.96Yo 3-month sample 61,888 10,632 1,4.660/o Table11 analysis GosUBenefit These data indicate that the float model discussedin this paper ptoffers the potential of approximately $1 million annualized savings relative to the current strategy. The assumptions and estimate used to derive this finding are tabulated in table 1,2on the follo*ing page. A few caveatsare in order. Presently we are not provided with misclassification incident rates and sulns occurring under the operative float strategy deployed thtough Payment DefenderrM, rendering problematic our ability to make direct, accuratebenchmark champion/challenger comparisons with our model. Moreover, our modeling is based on certain variables that are proxies derived from month-end data, as we do not have accessto many FDR Master File data available to the Adapuve Control System (ACS) that are the actual drivers of any production float strategy. Finally, modeling is consffained to the use of predictors availablein ACS production. Other factors such as, for example, number of outbound and/or inbound collections contacts, PTPs (promises-to-pay), or RTPs (refusals)per cycle m^y have predictive utility with respect to NSFs, but are necessarily excluded. -11 - FNBMNSFFloatModelDevelopment otal NSF $ otal $ Lost if no stra $ 7,729,033.93 648,459.10 $ 480,574.83 $ 442,987.75 $ 254,429.47 $ fi $ fi fi et $ Lost umber of Accts floated 'o Accts of that Don't pay after NSF and have open to buy: Cosu Good $ Floated Current Stratew: 7,570,990.48 $ 50.0%o 85.0% $ 20.0% $ 9 1,9.8o/o 12,467.83 8,123,367.73 6 of Float that could be re-used 49.7o/o 85.0o/o 20.0% 9 19.8% t $ C/O Rate Revenue $ of Accts that call re: Float otal Cust Svc Cost 231,,759.59 (231,1,59.59) 78,114.31 90,697.78 (90,697.78) 22,577 46.1.60/o 85.00% that would have used up the open to buy: 'otal 7,729,033.93 $ 879,678.69 $ 249,475.24 $ 521,1,62.06 $ 345,727.19 97,860.56 54,954 $ I $ fi $ $ fi 73,233.53 5.0% 1.00 1,613.85 254,429.47 8,000.00 13,233.53 1,613.85 22,847.38 237,582.04 -12- (612,371.25) (765.6e) 5.0%o 1.00 2,742.70 $ $ $ $ $ 345,727.79 8,000.00 72,467.83 2,742.70 23,270.53 32't,91.6.66 1,128.85 $ $ $ ]h V $ 90,697.78 (765.6e) 1,128.85 363.t6