Post on 20-Dec-2015
Goal
• Increase Profit
• Reduce Cost of Settlements
• Increase Customer Satisfaction
• Reduce Bank Risk
• Reduce Capital Requirements
Domain
• FX Trading System Relational Database
– 6000 Customers
– 400,000 FX Transactions
– Demographic Information
– Credit Information
• FX Marketing Desk Customer Info Database
– Marketer
– Relationship Manager
– Pricing Information
Foreign Exchange Primer
• Spots and Forwards
• Swaps
• Window Options and Draw Downs
• Multi-currency Accounts
• Settlements
• Customer Credit
• Bank Risk
Methodology
• Action Rules are discovered to meet our Goals.
For Example:
Geography( Canada ) AND CreditLine( NO -> YES)
=> customerRating( Average -> Good )
• Confidence = 100%
• Support = 52 Customers
Methodology
• Data Extraction– SQL
– Statistical Attributes
• Data Nominalization– SQL
– Range Mapping based on Domain
Knowledge and Visualization
• Data Reduction– SQL
– 6,000 Customers to 2,500
Trad ingS ystem
C ustom er In foS ystem
C onso lida teedD ata
N om ina lizedD ata
Methodology
N om ina lizedD ata
R osetta
S upportingA ssocia tion
R ules
• Rosetta
– Reducts
– Association Rules
– Filtering
Methodology
• Custom Application
– Flexible versus Static Attributes
– Association Rule combination
– Filtering
S upportingA ssocia tion
R ules
S upportingA ction R u les
Action.java
Results
• Spot-rating is Strongly correlated to the decision
Attribute.
– Spot-rating as flexible attribute ( 1058 Action Rules )
– Spot-rating as static attribute ( 99 Action Rules )
• Improving Spot-rating improves Customer-rating
Results
• Some Customers would be more profitable by
doing business with a CRM Interface Partner
– 120 Supporting Customers
– Static• Spot-rating = GOOD
• Swap-volume = NONE
– Flexible• primaryDealsrc( Direct -> (9 other partners)
– Decision• BAD -> AVERAGE
Results
• Some Customers would be more profitable by
recovering settlement cost.
– 118 Supporting Customers
– Static• Spot-rating = GOOD
• Swap-volume = NONE
• Geography = US
• Customer Type = Corporate
– Flexible
• Settlement-volume( Medium -> low or high )
– Decision
• BAD -> AVERAGE
Results
• Marketer EBF Could do Better
– 68 Supporting Customers
– Static• Spot-rating = GOOD
• Swap-volume = NONE
• Geography = US
– Flexible
• marketer( EBF -> {13 other} )
– Decision
• BAD -> AVERAGE
Results
• Marketer BKG Could do Better
– 49 Supporting Customers
– Static• Spot-rating = EXCELLENT
• Swap-volume = NONE
• Geography = US
– Flexible
• marketer( EBF -> {5 other} )
– Decision
• AVERAGE -> GOOD
Next Steps
• More holistic view of Profit & Loss of the
products
• More attributes--less derived attributes
• Filter change to find rules with the most financial
impact support, not number of customers
supporting
• Use methodology for continuous attributes to
yield a more precise actions to take. E.g, increase
spread from 3.2% to 3.4% to increase profitability
by 5%