Abstract: Value at Risk(VaR) is a fundamental tool for risk management,providing a quantitative measure of the downside risks to existing positions.Historical VaR is one of the most popular methods of VaR calculation,which is widely used in many financial institutions in the world.Real time or quasi-real-time VaR calculation for financial pro-ducts is quite important to avoid financial risks in time.Due to the increasing complexity and increasing number of financial products,the computational capability of existing CPU platforms has not been able to meet the requirements of risk management for computing speed.To solve this problem,the historical VaR method was implemented and optimized on GPU by using CUDA.By improving the sorting algorithm,overlapping the communication time using Multi-stream,decoupling the data dependency and implementing the fine-grained parallel optimization method,42.6x speedup is achieved when it is compared with the performance of the single core CPU,which provides solution for fast VaR calculation of large amount of bonds.The optimization methods of this paper can also be referenced by other financial algorithms on GPU.

Key words: Value at risk,Historical VaR,CUDA,Risk management