Big Data Developments in Transaction Analytics
27
0
0
Full text
(2) Agenda » Transaction Analytics Streaming Transaction Profiles. » Deeper Understanding Behavior Sorted Lists. » Beyond a Single View Connecting the Holistic Customer. » Self-Calibrating Analytics Essential In-Stream Self-learning. 2. © 2013 Fair Isaac Corporation..
(3) Transaction Analytics Streaming Transaction Profiles. 3. © 2013 Fair Isaac Corporation..
(4) Transaction Fraud Detection Long Term Streaming Analytic Problem 1993 Falcon® introduced » » » » ». Real-time Predictive Analytics Decisions Transaction profiles for recursive variable updates Neural Networks for complex rapid score computation Fraud scores operationalized in authorization systems <10s of millisecond response times. 2013 » 65% of Payment cards transactions processed by Falcon® » More than a petabyte of transactions processed in real-time annually 4. © 2013 Fair Isaac Corporation..
(5) Boiling Down Transaction Histories Cardholder Transaction history June 3, $28.00, Amazon, … June 4, $33.51, Grocery, … June 4, $100.00, ATM, … June 5, $1.00, Preauth, …. Cardholder Transaction Profile. June 5, $53.88, Gas, … June 8, $28.00, Amazon, …. Channel Variables. June 9, $60.00, ATM, … June 9, $9.50, Fast food, … … … … … July 30, $51.18, Grocery, … July 31, $55.11, Gas, … Aug 1, $28.16, Amazon, … Aug 2, $80.00, ATM, … Aug 3, $8.75, Fast food, … … Now. » In many applications it is impractical to retrieve the entire transaction history for making predictions » Transaction Profiles contain recursive variables to summarize the relevant predictors. 5. © 2013 Fair Isaac Corporation..
(6) Capturing Patterns Using Analytics »Summarized data. »Cross Border. »Usage. »Usage. »In Country. »Transaction data. »Month-End. »Time. Two Very Different Behaviors Look The Same With Summarized Data 6. © 2013 Fair Isaac Corporation..
(7) Transaction Profiling Technology » Transaction Profiles Feature Detectors. » Adjust based on a customer’s real-time activity » Individualized to each customer » Optimized recursive mathematical variables. Profile variables include features such as: » Patterns of behavior » Recent trends and deviation from norm » Significant events. Input TRX Data. P R O F I L E. Profiles updated by every transaction » Relevant and most recent information used to make predictions and decisions. » Earlier identification of changes in customer behavior and allows timely scoring decisions. Custom designed for each problem. 7. © 2013 Fair Isaac Corporation.. Profiles give the model the power to compare recent patterns with historical behavior, or undesirable behavior with normal, desirable behavior.
(8) Deeper Understanding Behavior Sorted Lists. 8. © 2013 Fair Isaac Corporation..
(9) Why A Deeper Understanding of the Customer?. Reduce False Positives » Regularly do target behavior » Propensity to do target behavior. Eliminate Embarrassment » » » ». 9. “I use that gas station every week” “I pay that person every month” “I travel to that country every year” “Based on my spending pattern you should know my likely transactions”. © 2013 Fair Isaac Corporation..
(10) Individuals are Creatures of Habit Cardholders use » Favorite ATMs close to work or home » Favorite gas stations along the way of daily commute » Preferred grocery stores just for knowing where to find things » Preferred online stores for internet shopping » Other preferred merchants/localities (postal codes). 10. © 2013 Fair Isaac Corporation..
(11) Frequent Behavior Tracking Algorithm Index. Entity Value. Frequency. Index. Rank. Frequent ATM. 1. ATM_77. F1=3.2. 1. 2. Rank. BLIST table. 2. ATM_318. F2=9.2. 2. 1. Table. 3. ATM_291. F3=0.3. 3. 4. 4. ATM_54. F4=2.7. 4. 3. A New ATM withdrawal at ATM_54. Index. Entity Value. Frequency. Index. Rank. 1. ATM_77. F1*wt. 1. 3. 2. ATM_318. F2*wt. 2. 1. 3. ATM_291. F3*wt. 3. 4. 4. ATM_54. F4*wt+Fo. 4. 2. ATM_318 and ATM_54 are Favorites after update 11. © 2013 Fair Isaac Corporation.. U.S. Patent 8,090,648.
(12) BLIST Differentiating Fraud Risks » In-Pattern means low risk while Out-Pattern means high risk » BLIST helps to reduce Falcon scores for normal cardholder In-Pattern spending even in risky categories. Legacy FFM V-Liquor. Low Risk Frequency of hitting favorites. 12. © 2013 Fair Isaac Corporation.. 820 S-Casino. Rank. V-Liquor. 1. S-Casino. 2. In Pattern. Risk level. High Risk. T&E Merchant. FFM with BLIST. Reduce scores. S-Casino. V-Liquor 670. 930. 710. V-Liquor. V-Liquor. 870. 630. U.S. Patent 8,090,648.
(13) Beyond a Single View Connecting the Holistic Customer. 13. © 2013 Fair Isaac Corporation..
(14) Why Connect Decisions? Better Customer Experience » Eliminate conflicting decisions » Reduce customer impact due to multiple inquiries. Better Predictions » Leverage additional information across more channels » Customer activity is correlated. Better Business Decisions » Know Your Customer (KYC). 14. © 2013 Fair Isaac Corporation..
(15) Customer within a Silo – limited view / understanding » Often effective scores exist in silos. Device?. » The complete customer view requires connecting silos/decisions. Login?. » Must link multiple accounts, devices, logins to the customer. Falcon Score. Device?. Customer?. Login?. Falcon Score 231. 15. © 2013 Fair Isaac Corporation.. 920. Device?. Customer? Account 120912901. Customer?. Login?. Account 109801980. Falcon Score 989. Account 329109823 U.S. Patent 8,296,205.
(16) Connecting Profiles: A Correct Data Model. 16. Field Name recordType. Type CHAR. recordCreationDate. YYYYMMDD. recordCreationtime customerIdFromHeader. HHMMSS CHAR. customerAcctNumber. CHAR. Transaction Record Format. Per Specification Field Level Definitions. © 2013 Fair Isaac Corporation.. Definition This Record's Type. "ACH" “Check" “Credit Card “ “Customer Information“ “Account Information“ “Application“ Etc. Date of the record creation Time of the record creation The Primary Customer Number. Financial institution's unique identifier for the customer. Customer Account Number. Identifier for the account used in the transaction,. U.S. Patent 8,296,205.
(17) Connected Decisions Login. Device (Computer1). Customer. Device (Computer2). Account Account. Account Account. Account. Account. (DDA) (DDA). (Savings) (Savings). (Credit Card). (Loan). Device (Mobile Phone). PAN PID. PAN PID. PID. Customer = The unique identifier for a given individual across all channels. Account = An identifier for an account which may be one of many held by the associated customer. PAN = The primary account number account number encoded or embossed on the payment instrument Payment Instrument Id (PID) = Identifier to distinguish between payment instruments with the same PAN. Device = Identifier for the device the customer uses access his accounts at the financial institution. Login = The login user identifier the customer uses to access his 17. © 2013 Fair Isaac Corporation.. U.S. Patent 8,296,205.
(18) Connecting Profiles: Multiple Profiles Customer Variables. Field Name. Cross-Acccount Variables. Account Summary Profiles. recordType. Customer Profile Database. DDA. Savings. Credit Card. Loan. recordCreationDate recordCreationtime. Account Variables. Access Channel Profiles. Cross-Channel Variables. customerIdFromHeader. customerAcctNumber. Transaction Record Format. Account Profile Database. Card. ACH. Wire. Check. Channel Variables. PAN paymentInstrumentId. 18. © 2013 Fair Isaac Corporation.. Channel Profile Database. U.S. Patent 8,296,205.
(19) Customers have Different Mixes of Accounts. Credit 1. Credit 2. Debit 0. DDA 0. Credit 1. Debit 0. DDA 0. ONLINE 0. Debit 1. ONLINE 0. Unsecured 0. DDA 2. ONLINE 1. Unsecured 0. Checking 0. Checking 0. Unsecured 0. Checking 1. How to train models with this variability / complexity? Will models be relevant?! 19. © 2013 Fair Isaac Corporation.. U.S. Patent 8,296,205.
(20) Self-Calibrating Analytics Essential In-Steam Self-Learning. 20. © 2013 Fair Isaac Corporation..
(21) Why Self Calibrating Models? Data availability issue » Data export restrictions » Gathering of data impractical » Unreliable targets/no targets. Diverse Customers » Peer grouping addresses diversity » NOT 100+ poorly trained models. Dynamic Environment » Real-time adjustment to changes. Emerging Markets Require Flexible Streaming Analytics. 21. © 2013 Fair Isaac Corporation..
(22) Self-Calibrating Analytics. % Population % Population. LowLow RiskRisk. High Risk. High Risk. Rate of High-Dollar Cross Border TrxTrx Rate of High-Dollar Cross Border. Scaling function. q ( xi | s ) where. 22. © 2013 Fair Isaac Corporation.. s p sL sR. xi s p (sL sR ) / c. [0, C ]. are quantiles estimated real-time. U.S. Patent 8,027,439, 8,041,597, 13/367344.
(23) Self-Calibrating Analytics. Score wi q ( xi | t , s ) (t1 , , t m ) T. : segment vector (m segments). ( x1 , , x p ) X. :profile variables (p variables). ( s1 , , sl ) S. : quantile estimates. wi. Peer Group Distributions. : weights. Weights can be assigned » Uniformly » Using Expert Knowledge » Based on limited data. 23. © 2013 Fair Isaac Corporation.. U.S. Patent 8,027,439, 8,041,597, 13/367344.
(24) Multi-layered Self-Calibrating Models US 13/367,344 Pending Input Layer 1. Hidden 1 Layer. 2. 3. 5. 5 2 4 6 3. 6. 6. 3 4. Output Layer. Multi-Layer Self Calibrating Score Weights Tuning. Inspired by Neural Networks » Each hidden node is a separate self-calibrating model » Variables in the hidden node are selected based on Factor group analysis » Ability to do simple tuning of output weighting of hidden nodes to improve performance or study effectiveness of different hidden nodes » Can experiment with new hidden nodes/variables in production, dial up promotion of nodes 24. © 2013 Fair Isaac Corporation.. U.S. Patent 8,027,439, 8,041,597, 13/367344.
(25) Adaptive MLSC Model Architecture. Transaction. Score Peer Grouping Feature Detectors. Segment Scaling Data Extraction. Offline Learning. Profiles Peer Group i Distributions. Regular Updating of Weights. Intelligent Transform + Labeling. Patent Pending 25. © 2013 Fair Isaac Corporation.. U.S. Patent 8,027,439, 8,041,597, 13/367344.
(26) Performance Results US Data. Transaction Detection Rate (%). » Competitive on US Data » Great For International clients » Smaller amounts of data » Rapidly changing » Regional variety. » Adaptive tuning sustains robustness. Unsupervised MLSC obtains. 75% US FFM Credit v17 MLSC model. Non-Fraud Transaction Review Rate (%) 26. © 2013 Fair Isaac Corporation.. of multi-segment neural network trained with frauds on stable US Data.
(27) THANK YOU. Scott M Zoldi, PhD 858-369-8891 ScottZoldi@fico.com Credit Scoring and Credit Control XIII August 28-30, 2013 This presentation is provided for the recipient only and cannot be reproduced or shared without Fair Isaac Corporation's express consent.. 27. © 2013 Fair Isaac Corporation..
(28)
Related documents
Following applications were used to demonstrate the usefulness of integration of Big Data in power system applications: • Improved traveling wave fault location technique
I We also consider a noisy variant with results concerning the asymptotic behaviour of the MLE. Ajay Jasra Estimation of
Executive Ordinance of the Minister of National Education of September 30, 1998 concerning the detailed principles, mode and criteria of granting, repaying and redeeming student
joint level mechanical work, and cumulative muscle activation while cycling at a constant 97.. submaximal power output; (2) to determine the effect of cadence on
As the highest applicable percentage ratio of the Current Transaction is less than 5%, while the highest applicable percentage ratio upon aggregation with the Previous Transaction
Very few small utilities have check meters to monitor the accuracy of these gate stations to ensure the accuracy of the measurement. It is important to know what size cubic foot
These classes are tests based on F statistics (supF , aveF , expF tests), on OLS residuals (OLS-based CUSUM and MOSUM tests) and on maximum likelihood scores (including
gas cleaning engine aftertreatment Limit – Injector technology Automotive (Helmond) Automotive (Helmond) TTAI TÜV Rheinland TNO Automotive International (Helmond) TTAI TÜV Rheinland
