REJECT INFERENCE FOR CREDIT ADJUDICATION

Transcription

1 REJECT INFERENCE FOR CREDIT ADJUDICATION May 2014

2 THE SITUATION SOMEONE APPLIES FOR A LOAN AND A DECISION HAS TO BE MADE TO ACCEPT OR REJECT. THIS IS CREDIT ADJUDICATION IF WE ACCEPT WE CAN OBSERVE PERFORMANCE * IF WE REJECT WE DO NOT KNOW IF WE MADE THE RIGHT DECISION ** THIS CREATES CHALLENGES FOR MONITORING AND BUILDING MODELS FOR CREDIT ADJUDICATION * There are applications that are accepted and do not book. Not a topic for today ** Lost opportunity analysis whereby bureau lets us know if they booked similar product elsewhere and performance. Expensive and open to interpretation unless similar product was booked elsewhere almost immediately after our rejection 1

3 THE SITUATION WE CAN ONLY OBSERVE PERFORMANCE ON BOOKED. THIS IS A BIASED POPULATION: WE MIGHT SEE LOW INCOME PERFORM VERY WELL WE MIGHT SEE MAJOR DEROGATORIES PERFORM VERY WELL REJECTS HAVE MANY MORE LOW INCOME/MAJOR DEROGATORIES AND THEIR PERFORMANCE IS UNKNOWN AND MAY NOT MATCH COUNTERPARTS IN BOOKED BASING DECISIONS ON WHAT IS OBSERVED AMONG BOOKED MAY CAUSE US TO CONCLUDE THAT LOW INCOME OR MAJOR DEROGATORIES IS A GOOD THING. HOWEVER THE PERFORMANCE OF THESE BOOKED MAY NOT REPRESENT OVERALL QUALITY, THEY MAY HAVE BEEN CHERRY PICKED POSSIBLY EVEN FOR THINGS THAT LENDER CAN ONLY SEE AT THE SAME TIME WE WANT TO GIVE PREVIOUS REJECTS A CHANCE AS THERE MAY BE INSTITUTIONAL BIAS THAT IS NOT WARRANTED BALANCE BETWEEN POSSIBLE RECKLESS LEARNING FROM BOOKED AND POSSIBLE OVER-CONSERVATISM IN REJECTS 2

4 BAYES RULE EXPRESSION LET A = ACCEPT, R = REJECT, THROUGH THE DOOR (TTD) = A + R LET G = GOOD, B = BAD, A = G + B FOR APPLICANT WITH CHARACTERISTICS X P(B) = [P(B/A) * P(A)] + [P(B/R) * P(R)] WE CAN HISTORICALLY MEASURE EVERYTHING EXCEPT P(B/R) WE ASSUME THAT P(B/R) > P(B/A) HOW MUCH CONFIDENCE DO WE HAVE IN APPLYING THE LEARNING FROM THE ACCEPTS TO THE REJECTS THAT IN SOME WAYS LOOK THE SAME? 3

5 APPLYING LEARNING FROM BOOKED TO REJECTS SEGMENT P(B/A) P(A) P(B/R) P(B) % 99% 0.02% 0.02% % 90% 0.04% 0.03% % 80% 0.09% 0.07% % 70% 0.22% 0.15% % 60% 0.53% 0.36% % 50% 1.30% 0.89% % 40% 3.19% 2.29% % 30% 7.79% 6.02% % 20% 19.02% 15.96% % 10% 46.47% 42.53% P(B/A) = BASED ON MODEL OF BOOKED POPULATION PERFORMANCE P(A) = PROPORTION OF THAT TYPE OF PERSON ACCEPTED HISTORICALLY P(B/R) = HOW BAD WILL WE SAY LIKE BUT UNOBSERVED ARE? ODDS(B/R) = ODDS(B/A) x EXP[F x P(R)] AS P(A) APPROACHES 1 THEN P(R) APPROACHES 0 AND WE GIVE P(B/A) TO THE REJECTS AS P(A) APPROACHES 0 THEN P(R) APPROACHES 1 AND WE ASSUME REJECTS WILL BE MUCH WORSE IN THIS EXAMPLE F = 2 SOME SHOPS JUST BUILD P(B/A) MODEL AND USE BAYES RULE AT SEGMENT LEVEL, MAY EVEN USE A SIMPLER ADJUSTMENT SUCH AS P(B/R) = F X P(B/A), EVERYONE IN SEGMENT GETS SAME P(B/R) AND P(B) WE DO NOT PUNISH ALL ACCOUNTS IN SAME SEGMENT EQUALLY, P(A) IS JUST AN AVERAGE EG: IN SEGMENT 4 WE DO NOT INCREASE P(B/R) ABOVE P(B/A) EQUALLY FOR EVERYONE, IT DEPENDS ON ACCEPT RATE, IF HIGH ACCEPT RATE P(B/R) STAYS CLOSE TO 0.12%, IF LOW ACCEPT RATE P(B/R) INCREASES TO EVEN HIGHER THAN 0.22% ACCEPT REJECT MODEL IS USED TO MEASURE P(A) AT ACCOUNT LEVEL 4

6 COMMENTS ON P(B/R) DERVIATION FROM P(B/A) ODDS(B/R) = ODDS(B/A) x EXP[F x P(R)] IS NOT AN ATTEMPT AT A THEORETICAL EXPRESSION FOR REJECT PERFORMANCE. IT IS AN EXPRESSION OF THE UNCERTAINTY APPLIED TO THE LEARNING ON THE ACCEPTS WHEN USED ON REJECTS WITH SAME CHARACTERISTICS. FOR THIS PRESENTATION F=2. WHY 2? WHY NOT 3? WHY NOT 4? ALL REJECT INFERENCE METHODS HAVE JUDGEMENT INVOLVED. USERS HAVE TO DECIDE THEMSELVES WHICH METHOD AND SETTINGS THEY ARE MOST COMFORTABLE WITH I CAN T CLAIM ANY MAGIC THAT I KNOW HOW A REJECT WOULD HAVE PERFORMED, I JUST WANT TO SET UP A SENSIBLE BALANCE BETWEEN POSSIBLE RECKLESS LEARNING FROM THE ACCEPTS AND UNDOING POSSIBLE PAST OVER-CONSERVATISM ON THE REJECTS 5

7 A QUICK WALK THROUGH OF UNIVARIATE EXPLORATION OF POTENTIAL PREDICTORS 6

8 CREDIT BUREAU SCORE SHOWS A NEGATIVE CORRELATION WITH BAD RATE 100% 90% CREDIT BUREAU SCORE 3.0% THERE ARE A FEW MISSING CREDIT BUREAU SCORES IN THE BOOKED POPULATION THAT PERFORM BETTER THAN THE FIRST DECILE 80% 70% 2.5% 2.0% 60% 50% 1.5% 40% 30% 1.0% 20% 0.5% 10% 0% MISSING DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE 0.0% BOOK POPULATION ACCEPT RATE BAD RATE 7

9 VERY LOW NET WORTH PERFORMS BETTER THAN SOME HIGHER NET WORTH BUT THE ACCEPT RATE IS SO LOW THAT UNCERTAINTY IS RAISED 100% 90% 80% NET WORTH 1.8% 1.6% 1.4% 70% 1.2% 60% 1.0% 50% 0.8% 40% 30% 0.6% 20% 0.4% 10% 0.2% 0% DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE 0.0% BOOK POPULATION ACCEPT RATE BAD RATE 8

10 VERY HIGH # OF MAJOR DEROGATORIES PERFORMS THE BEST BUT SO FEW WERE ACCEPTED THAT THERE IS NOT MUCH CERTAINTY TO TRANSFER THIS KNOWLEDGE TO THE REJECTS 100% 90% 80% 70% MAJOR DEROGATORIES 2.0% 1.8% 1.6% 1.4% 60% 1.2% 50% 1.0% 40% 0.8% 30% 0.6% 20% 0.4% 10% 0.2% 0% a. MAJ=0 b. MAJ=1 c. MAJ=2 d. MAJ=3 e. MAJ>3 0.0% BOOK POPULATION ACCEPT RATE BAD RATE 9

11 REVOLVING UTILIZATION SHOWS NEGATIVE CORRELATION TO BAD RATE, EXCEPT FOR LOWEST DECILE 90% 80% REVOLVING UTILIZATION 2.0% 1.8% THIS HOCKEY STICK EFFECT IS VERY COMMON 70% 1.6% 60% 1.4% 1.2% 50% 1.0% 40% 0.8% 30% 0.6% 20% 0.4% 10% 0.2% 0% DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE 0.0% BOOK POPULATION ACCEPT RATE BAD RATE 10

12 THE AGE OF OLDEST TRADE ON BUREAU SHOWS A NEGATIVE CORRELATION WITH BAD RATE 100% 90% AGE OF OLDEST TRADE 2.5% AS IN ALL GRAPHS, LOOK AT ACCEPT RATE AS A CONFIDENCE MEASURE OF ON THE BAD RATE APPLIED TO REJECTS 80% 70% 2.0% 60% 1.5% 50% 40% 1.0% 30% 20% 0.5% 10% 0% DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE 0.0% BOOK POPULATION ACCEPT RATE BAD RATE 11

13 BUREAU REVOLVNG LIMIT THIS RELATION IS HIGHLY INFLUENCED BY THE FACT THE BETTER PEOPLE HAVE BEEN GIVEN MORE LIMIT AND DOES NOT SUGGEST WE CAN MAKE PEOPLE BETTER BY GIVING THEM MORE LIMT 90% 80% 70% 60% BUREAU REVOLVING LIMIT 1.8% 1.6% 1.4% 1.2% 50% 1.0% 40% 0.8% 30% 0.6% 20% 0.4% 10% 0.2% 0% DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE BOOK POPULATION ACCEPT RATE BAD RATE 0.0% 12

14 OWNER IS THE BEST AND A LOT ARE ACCEPTED 90% OCCUPANCY RESIDENTIAL 2.5% 80% 70% 2.0% 60% 1.5% 50% 40% 1.0% 30% 20% 0.5% 10% 0% a. OCC MISSING b. OCC B c. OCC O d. OCC P e. OCC R f. OCC X g. OCC U 0.0% BOOK POPULATION ACCEPT RATE BAD RATE 13

15 LOWER INCOME IN THE BOOKED POPULATION SUGGESTS NON-INTUITVE RELATIONSHIP HOWEVER, ACCEPT RATES THAT SUGGESTS THERE COULD BE QUITE A BIT OF BIAS 90% 80% 70% 60% INCOME 2.0% 1.8% 1.6% 1.4% 1.2% 50% 1.0% 40% 0.8% 30% 0.6% 20% 0.4% 10% 0.2% 0% DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE 0.0% BOOK POPULATION ACCEPT RATE BAD RATE 14

16 DEBT TO SERVICE RATIO VERY IRREGULAR PATTERN 90% DEBT TO SERVICE RATIO 1.4% 80% 1.2% 70% 60% 1.0% 50% 0.8% 40% 0.6% 30% 20% 0.4% 10% 0.2% 0% DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE BOOK POPULATION ACCEPT RATE BAD RATE 0.0% 15

17 CREDIT BUREAU INQUIRIES 80% CREDIT BUREAU INQUIRIES 3.0% 70% 2.5% 60% 50% 2.0% 40% 1.5% 30% 1.0% 20% 10% 0.5% 0% a. INQ=0 b. INQ=1 c. INQ=2 d. INQ=3 e. INQ>3 0.0% BOOK POPULATION ACCEPT RATE BAD RATE 16

18 TIME ON JOB LOW TIME ON JOB LOOKS PRETTY GOOD BUT ACCEPT RATE SUGGESTS THEY WERE CHERRY PICKED 90% 80% 70% TIME ON JOB 1.8% 1.6% 1.4% 60% 1.2% 50% 1.0% 40% 0.8% 30% 0.6% 20% 0.4% 10% 0.2% 0% DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE BOOK POPULATION ACCEPT RATE BAD RATE 0.0% 17

19 IS THIS AN INDIVIDUAL OR A JOINT APPLICATION? 90% INDIVIDUAL VS JOINT 1.4% JOINT LOOKS BETTER AND THEY ARE ALSO ACCEPTED AT A HIGHER RATE 80% 1.2% 70% 1.0% 60% 50% 0.8% 40% 0.6% 30% 0.4% 20% 10% 0.2% 0% a. IJA=I b. IJA=J 0.0% BOOK POPULATION ACCEPT RATE BAD RATE 18

20 TIME AT RESIDENCE 90% TIME AT RESIDENCE 1.8% 80% 1.6% 70% 1.4% 60% 1.2% 50% 1.0% 40% 0.8% 30% 0.6% 20% 0.4% 10% 0.2% 0% DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE DECILE 0.0% BOOK POPULATION ACCEPT RATE BAD RATE 19

21 GOOD-BAD MODEL = P(B/A) THE MODEL ON THE BOOKED POPULATION USES EMPIRICALLY OBSERVED PERFORMANCE BUT MAY NOT BE REPRESENTATIVE OF ENTIRE TTD 20

22 CREDIT RISK PEOPLE LIKE TO DO WEIGHT OF EVIDENCE (WOE) BINNING THIS IS THE OPTIMAL WOE BINNING AGAINST BAD FOR THE BOOKED POPULATION NO NEED AND DETRIMENTAL TO SMOOTH/COLLAPSE SUSPECTED CHERRY PICKING LIKE FIRST BIN OF INCOME (LOW BAD RATE) WOE IS LOG(ODDS BIN) MINUS LOG (ODDS POPULATION) I.E. STANDARDIZED LOG(ODDS) BINNING CONTROLS EFFECT OF OUTLIERS, CAN CAPTURE NON- LINEARITIES BUT CAN ALSO HARM AN UNDERLYING LINEAR RELATION WOE REDUCES FREEDOM OF BIN COEFFICENTS REDUCING OVERFITTING (CAN BE GOOD OR BAD) AND GIVES EXPLANATION FRAMEWORK CREDIT BUREAU SCORE POP % BAD RATE WOE ACCEPT MISSING 0.57% 1.71% % < % 2.15% % >=696 < % 1.10% % >=767 < % 0.57% % >=778 < % 0.70% % >= % 0.30% % AGE OF OLDEST TRADE (MONTHS) POP % BAD RATE WOE ACCEPT < % 1.91% % >=89 < % 1.29% % >=116 < % 1.00% % >=190 < % 0.66% % >= % 0.29% % NET WORTH POP % BAD RATE WOE ACCEPT <74, % 1.51% % >=74,446 <138, % 1.23% % >=138,000 <209, % 1.02% % >=209,190 <549, % 0.59% % >=549, % 0.18% % REVOLVING UTILIZATION POP % BAD RATE WOE ACCEPT <0.623% 10.58% 1.23% % >=0.623% <4.099% 14.31% 0.40% % >=4.089% <26.196% 36.32% 0.73% % >=26.196% <79.900% 31.40% 1.15% % >=79.900% 7.40% 2.25% % CREDIT BUREAU INQUIRIES POP % BAD RATE WOE ACCEPT % 0.65% % % 0.95% % % 1.20% % >= % 2.09% % TIME ON JOB (MONTHS) POP % BAD RATE WOE ACCEPT < % 1.04% % >=28 < % 1.73% % >=54 < % 0.99% % >=85 < % 0.72% % >= % 0.38% % OCCUPANCY RESIDENTIAL POP % BAD RATE WOE ACCEPT O = OWNER 77.82% 0.78% % P = LIVES WITH PARENTS 6.88% 1.32% % B, R, U, X - BOARD/RENT/UNKNOWN 15.30% 1.83% % INDIVIDUAL VS JOINT POP % BAD RATE WOE ACCEPT J 30.91% 0.50% % I 69.09% 1.19% % INCOME POP % BAD RATE WOE ACCEPT <4, % 0.79% % >=4,506 <5, % 1.61% % >=5,783 <7, % 1.27% % >=7,000 <10, % 0.91% % >=10, % 0.57% % 21

23 THE LOGISTIC REGRESSION RESULT CONCORDANCE = 75.5% AR = 56.3% 22

24 ACCEPT-REJECT MODEL GIVES P(A) AND P(R) THIS MODEL IS BUILT ON THROUGH THE DOOR AND IS USED TO DESCRIBE WHO WAS ACCEPTED HISTORICALLY THIS MODEL REPRESENTS PAST DECISIONING WHICH HAS BEEN AFFECTED BY PREVIOUS SCORECARD, STRATEGY AND LENDER OVERRIDE BEHAVIOR WE DO NOT WANT TO MIMIC POOR PAST DECISIONING PAST DECISIONS CONTROLLED THE GATES OF EVIDENCE, IF THE GOOD-BAD MODEL TELLS US THIS PERSON IS THIS GOOD THE ACCEPT- REJECT MODEL TELLS US HOW MUCH OF THIS PERSON WE HAVE OBSERVED IN THE PAST AND HENCE HOW CONFIDENT WE CAN BE IN APPLYING THE GOOD-BAD RESULT TO THE PREVIOUS REJECTS CONCORDANCE = 91.3% AR = 82.7% ACCEPT REJECT MODEL TEND TO BE VERY STRONG 23

25 THE FINAL P(B) RESULTS P(B) = P(B/A) x P(A) + P(B/R) x P(R) P(B/A) = GOOD-BAD MODEL PREDICTION OF BAD P(A) = ACCEPT-REJECT MODEL PREDICTION OF PAST ACCEPT P(B/R) = P(B/A) ADJUSTED UPWARDS TO REFLECT UNCERTAINTY P(R) = ACCEPT-REJECT MODEL PREDICTION OF PAST REJECT 24

26 FINAL MODEL RESULTS INCOME VARIABLE CHOSEN TO ILLUSTRATE INCOME POP % P(B/A) P(A) P(B/R) P(R) P(B) <4, % 2.08% 51.78% 8.50% 48.22% 7.81% >=4,506 <5, % 2.84% 66.48% 8.40% 33.52% 7.31% >=5,783 <7, % 2.08% 71.83% 5.93% 28.17% 5.10% >=7,000 <10, % 1.32% 77.49% 3.56% 22.51% 3.03% >=10, % 0.70% 80.83% 1.95% 19.17% 1.67% WHEN THE BOOKED MODEL P(B/A) IS APPLIED TO THE THROUGH THE DOOR POPULATION IT SUGGESTS THAT INCOME <4,506 IS BETTER THAN 4,506<=INCOME<5,783 AND EQUIVALENT TO 5,783<=INCOME<7,000 RANGE WHEN PAST OBSERVED UNCERTAINTY IS APPLIED AT ACCOUNT LEVEL USING ACCEPT-REJECT MODEL THE INCOME<4,506 PREDICTED BAD RATE P(B) BECOMES THE WORST NOT ALL INCOME<4,506 HAVE P(B)=7.81%, THIS IS AN AVERAGE WITH A DISTRIBUTION. THIS ACCOUNT LEVEL APPROACH BETTER RECOGNIZES THOSE LOW INCOME ACCEPTED IN THE PAST WHO PERFORMED WELL AND CONTINUES TO GIVE THEM A GOOD SCORE AND LEAVES BEHIND THE TYPE OF LOW INCOME WHO DO NOT APPEAR TO HAVE PROVEN CHARACTERISTICS. OTHER METHODS HAVE MORE OF A TENDENCY TO OPEN THE FLOODGATES OR TO TOTALLY SHUTDOWN THE LOW INCOME 25

27 EVALUATION OF FINAL SCORECARD HOW WELL DOES P(B) PERFORM? NO ONE REALLY KNOWS. CAN ONLY MEASURE ON BOOKED SO EVEN MONITORING IS BIASED P(B/A) IS BEST POSSIBLE MODEL ON BOOKED THIS METHODOLOGY TENDS TO NOT DETERIORATE RANKING ON BOOKED AS MUCH AS MORE TRADITIONAL REJECT INFERENCE AR OF P(B/A) ON BOOKED = 56.3% AR OF P(B) ON BOOKED = 54.5% THE BENEFITS OF THE MODEL ARE BOTH PORTFOLIO QUALITY AND EFFICIENCY REMOVING PAST ACCEPTS THAT DID NOT PERFORM WELL AND REPLACING WITH PREVIOUS REJECTS THAT ARE EXPECTED TO PERFORM WELL UPGRADING PREVIOUS MANUAL ACCEPTS THAT PERFORM WELL TO SUCH HIGH SCORE THAT THEY BECOME AUTOMATIC APPROVALS AND FREEING UP THE LENDER 26

28 EVALUATION OF FINAL SCORECARD ON A POPULATION THAT HAS A SIGNIFICANT NUMBER OF REJECTS MEASURING PERFORMANCE ON BOOKED MAY BE MISLEADING HOWEVER PERFORMANCE ON BOOKED IS ALL WE HAVE AND IT IS WHAT EVERYONE LOOKS AT THERE COULD BE SITUATIONS WHERE A SCORECARD THAT IS THE BEST ON THE ENTIRE TTD MAY NOT PERFORM WELL ON THE BOOKED BUT DID A FANTASTIC JOB PREVENTING THE BAD GUYS FROM BOOKING 27

29 MEASURNG SCORECARD ON BOOKED MAY NOT BE MEANINGFUL THE ENTIRE POPULATION CONTAINS 50,000 APPLICANTS OF WHOM 1,500 ARE BAD TWO SCORECARDS EACH WITH A CUTOFF CALIBRATED TO REJECT 11,000 IF SCORECARD1 IS USED IT REJECTS 11,000 OF WHOM 1,100 ARE BAD IF SCORECARD2 IS USED IT REJECTS 11,000 OF WHOM 400 ARE BAD FOR DEALING WITH THROUGH THE DOOR SCORECARD1 IS MUCH BETTER ASSUME YELLOW IS BOOKED POPULATION: SCORECARD1 FAILS NO BADS SCORECARD2 FAILS ½ THE BADS SCORE SEPARATION BETTER ON SCORECARD2 BECAUSE SCORECARD1 HAS NO BOOKINGS FROM LOW RANGE DOOR SCORECARD1 SCOREARD2 REJECT ACCEPT TOTAL REJECT 2,000 (200) 9,000 (200) 11,000 (400) ACCEPT 9,000 (900) 30,000 (200) 39,000 (1,100) TOTAL 11,000 (1,100) 39,000 (400) 50,000 (1,500) 28

30 HOW TO MEASURE SCORECARD? STRONG PERFORMANCE ON BOOKED IS PLEASANT TO SEE LOW LEVEL OF OVERRIDES IN BOTH THE STRATEGIES AND IN THE FIELD AS THIS SUGGESTS THE BUSINESS STRATEGISTS AND LENDERS BELIEVE IN THE SCORECARD SCORECARDS CAN HAVE EXCELLENT AR OR MAXKS BUT BE DOING NOTHING! NEW SCORECARD IS INTRODUCED THAT SHOWS VERY STRONG AR AS MEASURED ON BOOKED BUSINESS STRATEGISTS AND LENDERS JUST KEEP APPROVING/REJECTING SAME PEOPLE THROUGH OVERRIDES, NO REAL CHANGE TO BOOKED POPULATION BOOKED POPULATION UNAFFECTED BY NEW SCORECARD AND MODEL MONTORING REPORT STILL SHOWS LIFT IN AR, MODELLER IS PROUD BUT HAS HAD NO IMPACT ON BUSINESS THIS NEW METHODLOGY SHOWS MORE BUSINESS STRATEGIST AND LENDER COMFORT 29

31 COMMON CRITICAL FEEDBACK BECAUSE ACCEPT-REJECT MODEL IS USED THE MODEL WILL NO LONGER BE VALID IF BANK CHANGES LENDING PRACTISES NOT TRUE, IF BANK VENTURES INTO NEW AREAS IT CAN NOT BE AS CERTAIN ABOUT PERFORMANCE AND THIS METHODOLOGY REFLECTS THAT UNCERTAINTY. IF MORE OF A DIFFERENT TYPE OF PERSON GETS ACCEPTED THEN IN THE FUTURE WE WILL BE ABLE TO ASSESS, THIS WILL BE A NEW OPPORTUNITY TO LEARN, IN FACT LET S ACCEPT EVERYONE FOR A YEAR AND DO SOME REAL LEARNING! THIS METHODLOGY PROPAGATES PAST LENDING NOT TRUE, THIS METHODOLOGY IDENTIFIES A LOT OF PAST ACCEPTS WHO PERFORMED POORLY AND PROPOSES A LOT OF PREVIOUS REJECTS AS GOOD BETS TO BOOK. USING P(B) ON DOOR AND CHOOSING CUTOFF THAT GIVES SAME NUMBER OF ACCEPTS: 20,551 PREVIOUS ACCEPTS WITH A BAD RATE OF 3.31% ARE REMOVED AND REPLACED BY AN EQUAL NUMBER OF PREVIOUS REJECTS WITH AN EXPECTED BAD RATE OF 1.67% THIS MEANS THAT 24% OF PREVIOUS REJECTS ARE BEING SWAPPED IN! SAME NUMBER OF ACCEPTS WITH A PREDICTED FALL IN BAD RATE FROM 1.05% TO 0.81% OLD REJECT OLD ACCEPT TOTAL NEW REJECT NEW ACCEPT TOTAL # BAD # BAD # BAD 65, % 20, % 86, % 20, % 123, % 144, % 86, % 144, % 230, % SWAPSET SIMILAR IN SIZE TO OTHER METHODOLOGIES AND USUALLY MORE PALATABLE/BELIEVABLE TO BUSINESS/LENDERS NEED TO IMPLEMENT TWO SCORECARDS BUSINESS USERS ACTUALLY LIKE THE SEPARATION OF OBSERVED PERFORMANCE AND MEASURE OF UNCERTAINTY HAVE TRIED TO REPRESENT AS ONE LINEAR SCORECARD MATHEMATICALLY CHALLENGING 30

32 BRIEF OVERVIEW OF OTHER METHODS P(B/A) FROM BOOKED IS APPLIED TO REJECTS. REJECTS ARE INFERRED TO BE GOOD OR BAD BASED ON THEIR P(B/A) SCORE. JUDGEMENT USED TO INCREASE CHANCE WE INFER AS BAD, I.E. IF REJECT HAS P(B/A) = 0.1 THEN ONE COULD MAKE THERE BE A 0.1 X F CHANCE WE INFER AS BAD WHERE F=1, 2, 3, THEN FINAL REGRESSION IS RUN ON ENTIRE POPULATION WITH A MIX OF TRUE BADS AND INFERRED BADS TO GET SINGLE SCORECARD BUSINESS USERS AND LENDERS NOT AS COMFORTABLE WITH NEW ACCEPTS AND TEND TO OVERRIDE USERS ARE NOT COMFORTABLE WITH REGRESSION RESULTS THAT USE A MIX OF OBSERVED AND HYPOTHETICAL VALUES FOR THE TARGET ACCEPT-REJECT MODEL IS DEVELOPED AND BOOKED POPULATION IS WEIGHTED BY P(R). THIS CAUSES APPLICATIONS IN BOOKED POPULATION THAT LOOK LIKE REJECTS TO BE MORE REPRESENTED IN BOOKED POPULATION BUSINESS USERS AND LENDERS NOT AS COMFORTABLE WITH NEW ACCEPTS AND TEND TO OVERRIDE 31