Umeå universitets logga

Value at Risk (VaR) is a financial risk measurement that Nasdaq calculates using Historical Simulation VaR (HVaR) in one of their products. In Nasdaq's current solution, the calculations are performed in two steps, Step 1 and Step 2. Step 1 produces a large output matrix that is used in Step 2 to compute the VaR measure. Step 1 is characterized by being memory-intensive but not latency-critical. However, Step 2 is latency-critical because it can be triggered by a user. The calculations are performed on a single computer/process (VaR node), which has become problematic because the memory usage is too high.

In this thesis, a distributed solution for calculating VaR is designed, implemented, evaluated, and compared to Nasdaq's current solution. The proposed design is based on Nasdaq's current solution but uses theory from distributed systems and parallel programming to distribute the resource usage across multiple VaR nodes that communicate over a network, while attempting to keep the latency of Step 2 low. Constraints in the proposed solution design included the requirement that the majority of existing code be reused.

The implementation of the proposed design is found to drastically reduce the memory usage per VaR node, for both Step 1 and Step 2, in some cases by around 78\% compared to Nasdaq's current solution. However, the calculation latency of Step 2 increases with the implementation of the proposed design when running with multiple VaR nodes, in the best case, doubling it.

In addition, some issues with the proposed design related to distributed systems theory are identified, such as availability, and potential solutions and mitigations are discussed.