Maximizing predictive performance at origination and beyond!

Transcription

1 Maximizing predictive performance at origination and beyond! John Krickus, Experian Joel Pruis, Experian Amanda Roth, Experian Experian and the marks used herein are service marks or registered trademarks of Experian Information Solutions, Inc. Other product and company names mentioned herein may be the trademarks of their respective owners. No part of this copyrighted work may be reproduced, modified, or distributed in any form or manner without the prior written permission of Experian Information Solutions, Inc.

2 Agenda Score performance in financial crisis Model performance by model type Impact of data Custom model process Decisioning options 2

3 Model performance in the financial crisis 3

4 Consumer marketplace Financial Institutions experienced higher delinquency across all products Increase actually began in 2004 / 2006 for real estate portfolio Bankcard experienced 33% increase in the 2007 / 2009 period 4

5 Performance of consumer score Although some change was experienced, scores remained relatively stable over the analyzed periods Percentage of change experienced from development time: Overall 9% Auto 4% Bankcard 2% Real estate 30% Real estate experienced the greatest degradation primarily due to: Capacity verification Home value depreciation Job loss increases Validation results using GINI statistic 5

6 Employee size trends Six-month trend shows slight decline for smaller firms Risk score March 2010 Average DBT Percent dollars delinquent Percent 91+ delinquent Risk score Average DBT Six-month change Percent dollars delinquent For small business (0-20), the percent dollar delinquent and the percent 91+ increasing slightly Percent 91+ delinquent National average % 5.2% -0.9% 1% -1% 0% Number of employees Nonemployer % 4.7% -0.7% 0.2% -3.3% -1.5% 1 to % 7.6% -1.4% 1.0% 0.1% 0.9% 5 to % 7.2% -1.3% 0.9% 0.7% 2.4% 10 to % 5.9% -1.1% 0.9% -0.3% 1.5% 20 to % 4.8% -0.9% 0.7% 0.4% 3.0% 50 to % 3.5% -0.7% 0.7% -0.2% 2.0% 100 to % 2.8% -0.7% 0.5% -0.2% -3.3% 250 to % 2.2% -0.8% 0.2% 0.4% 1.3% 500 to % 1.8% -1.1% 0.4% -0.1% -2.1% 1,000 and over % 2.0% 1.3% 1.7% 0.3% 1.4% Source: Experian Business Information Services 6

7 Performance of business score: Intelliscore Plus SM Pre vs. post financial crisis, bottom 10% bad capture Bad rate capture for bottom 10% remained strong across model segments Note: Large company segment and demographic only records less then 5% of total, not shown 7

8 Lending At the peak of the cycle Economic growth What are you doing with those clients that are just above your acceptable score threshold? Expansion Recession Time 8

9 Unsubstantiated response to recession 9

10 Loan production change Pre vs. post Category Pre Post Change Impact Application volume per financial institution 1,185 1, $8,624,860 Average request $112,186 $95,012 - $17,174 - $11,297,401 Approval percentage 62% 58% -4% - $4,032,309 $82,423,054 $58,468,484 - $23,954, % Net Interest Margin = $718,637 10

11 Summary: financial crisis performance Consumer / business / lending Consumer scores experienced slight declines, major exception: real estate Business scores, bottom 10% bad capture held up extremely well Small business payment performance has finally improved Combination of factors drove drop in lending 84% of drop related to number of applications, amount requested, 16% related to approval rate 11

12 Model performance by model type 12

13 Evolution of model performance All industry Industry-specific Custom model Model performance is driven by several factors: Bad definition, model build population, depth/breadth and quality of data Performance gains have been realized in each model type Which is the best fit for your operation? Factors to consider: Volume / account size / score dependency Cost benefit analysis costs vs. operation savings and performance improvement 13

14 Performance of all industry models Bad capture All industry Intelliscore Plus SM (October 2008) vs. previous two models Intelliscore Plus SM Commercial Intelliscore SM 210 Commercial Intelliscore SM 113 Significant performance gains for IP: 10% and 20% bad capture up around 40% vs. 113 model Bad definition: 90 day payment delinquency or bankruptcy Model built off of Experian s all industry database, BizSource SM, bad rate of 14.7% 14

15 Performance of industry-specific vs. all industry Industry-specific Small Business Credit Share SM : Bad capture bottom 20%, private card 80% bad capture for industry model vs. 50% for all-industry model (bottom 20% bads) Bad definition: Two 60-day late pays or one 90-day late card payment Model built off of SBCS financial trade and general business database, bad rate: 7.81% 15

16 Highest evolution of model performance Custom Custom model Across all client types, superior performance (bottom 20% bad capture) 3% lift 46% lift 41% lift 53% lift Custom model provided continuing lift in bad capture, worse 10%, up to 53% improvement Performance gains measured against industry-specific Small Business Credit Share SM models Both Small Business Credit Share SM and BIS business data available; data selected for model is client specific 16

17 Impact of data on model performance 17

18 Predictive information impact on score* All industry Intelliscore Plus SM Model weighting* 5-10% 50-60% 10-15% 5-10% 10-15% Business aggregates Historical payment behavior, delinquency trends Current payment status number / trade balance / percent of accounts delinquent Credit utilization, in commercial data using highest credit as proxy for credit line Company profile years in file, industry risk, employee size, number of inquiries Derogatory items collections, liens, judgments, bankruptcy * Estimate based on inquiry patterns and score model segments utilized 18

19 Blended data improves performance Significant gains using blended vs. commercial only Percentage of bads eliminated Intelliscore Plus SM blended, meaning commercial data and consumer data on owner / guarantors vs. commercial 29% improvement in screening out bad accounts at the 90% cutoff level 71% of bads excluded vs. 55% Almost screen out as many bads at higher approval level, 90%, as were screened out at 80% 71% vs. 74% 19

20 Financial industry specific data provides lift Small Business Credit Share SM data expands total number of trades Commercial credit database ALL001 Total number of trades Small Business Credit Share SM FTC001 Total number of financial trades ALL001 Total number of trades FIL001 Total number of financial installment loans FLS001 Total number of commercial lease trades FCC001 Total number of commercial credit card trades 20

21 Predictive data used in all industry vs. custom models Data selected that has the most impact on your portfolio Data aggregate for Intelliscore SM RTB030 average balance, recently delinquent accounts RTO076 past presence of delinquency TTC series, number of accounts: current / delinquent / derogatory TTC090 percent of balance seriously delinquent TTP series, percent of accounts: active / delinquent / seriously delinquent TTP093 utilization: total balance to total high credit OTB leasing balance Data aggregate for custom model RTB030 average balance, recently delinquent accounts RTO076 past presence of delinquency TTC series, number of delinquent accounts / current /derogatory TTC090 percent of balance seriously delinquent TTP series, percent of accounts: active / delinquent / seriously delinquent TTP093 utilization: total balance to total high credit OTB leasing balance 21

22 Custom model process 22

23 Custom model advantages The ultimate in performance Custom model Custom models utilizes your... Bad definition Portfolio profile Best data fit Each custom model is based on your bad definition, delinquency, nonperforming, etc. Uses your specific account portfolio characteristics and performance Model will utilize the business data elements that best fits your portfolio 23

24 Key advantage of custom model design Using your bad definition Custom model Bad definition is specific to your portfolio, not a generic or even industry model Model performance improves significantly when built directly on a specific bad event Early stage delinquency: 15, 30, 60 etc. Deposit requirement Write-off 24

25 Key benefits of custom model output Custom model Customized odds charts Results driven by portfolio characteristics Odds chart tied directly to your portfolio s performance Options to build Internal analytic resources External consultants Experian analytics, combo of above Improved performance 30%, 45% improved performance Huge ROI on bad debt / slow pay elimination with same or increase account approval 25

26 Custom model lift: Major cable / satellite Major lift in KS and bad capture compared with Intelliscore SM Custom model 45% lift in bad dollar capture 75% lift in KS KS Summary results by account type Bad $ capture worst scoring 20% 26

27 Custom model process three steps Weeks, not months Custom model Initial data files from with performance based on agreed to bad definition Average of three weeks, however can vary widely Decision science Two weeks to develop the model after the receipt of clean data: segmentation / model development / validate against holdout / performance statistics Technology Three weeks for implementation and deployment: coding and audit / validate programming Result: In eight weeks, or less, a custom model is produced! 27

28 Decisioning 28

29 Decision tree Enhancing score prediction Start at top with any standard score, Intelliscore Plus SM / Small Business Credit Share SM Identify, deploy key variables Combined trade count Combined trade balance DBT (days beyond terms) Years in file Recent high credit Derogatory legal count Optimize bad identification with decision paths and resulting decision nodes 29

30 Decision tree example Overall 1.19% write-off rate Start with portfolio and Intelliscore Plus (IP), bad rate of 1.19% Score range of 17-37, resulting bad rate of 1.94% Branch off of number of tradelines, 0-2, 2-7, 7+ IP score of and 0-2 trades, bad rate now 3.32% Further branch down to DBT, days beyond terms, of (below score of 17-27, nbr trades 0-2) DBT of 16-36, bad rate is 4.61% 30

31 Learn more in the Vision Expert Annex and online Meet Network Research Learn Vision Expert Annex Open every morning and afternoon as well as during session breaks Interact with Vision speakers and seek-out information on products and services Schedule one-on-one meetings with Experian experts Network with peers Review and request Experian thought leadership materials and research Visit the Internet café or relax in the Expert Annex lounge Vision Community Site Join the conversation at 31

32 For additional information, please contact: