Non-linear Aggregations in Large-Scale Multi-Dimensional Cubes Georges Bory, Quartet Financial Systems

1
Non-linear Aggregations in Large-Scale
Multi-Dimensional CubesGeorges Bory, Quartet
Financial Systems

• Distributed and Grid Computing in Computational
  Finance.
• Inria, Sophia Antipolis, October 20th 2008

2
Agenda

• OLAP cubes for finance
• Non linear behaviours
• Time constraint
• ActivePivot solution
• Performance and future work

3
OLAP cubes for Finance

• A lot of data
• And not much time to understand it

4
A lot of data

• Historical Var
• 2 years
• 500 000 deals
• 250 Million values
• Monte-Carlo Var
• 5 000 simulations
• 100 000 deals
• 500 Million values
• Potential exposure amount
• 100 000 deal
• 500 simulations
• 20 future points
• 1 Billion values

5
OLAP cubes for Finance

• Organize data into business hierarchies
• Drill down from top to bottom
• Filter
• Drill thru individual trades, scenario

6
Business Hierarchies

• High Cardinality Levels
• Securities gt10 000
• Counterparty gt 2 000
• Time buckets 80 future strips, gt10 000 days
• Low Cardinality Levels
• Books
• Traders
• Currencies
• Index

7
OLAP cardinality curse
8
Non linear behaviours

• Value at Risk
• Variance, Nth percentile loss
• Potential Exposure Amount
• Max (Expectation, 0)

9
Time constraint

• Any time lost in aggregation is expensive in grid
  hardware costs

10
Agenda

• OLAP cubes for finance
• Non linear behaviours
• Time constraint
• ActivePivot solution
• Performance and future work

11
ActivePivot solution

• Non linear aggregation
• Aggregate objects rather than values
• Apply operators to aggregated objects

12
ActivePivot solution
13
Transactional OLAP engine