By Edward Lim, Managing Consultant, CIZA Concept
Established in 2006 as a research institute at the National University of Singapore (NUS), the NUS Risk Management Institute (RMI) is dedicated to financial risk management. Its establishment was supported by the Monetary Authority of Singapore (MAS) under its program on Risk Management and Financial Innovation.
In 2009, RMI embarked on a non-profit Credit Research Initiative (CRI) in response to the financial crisis, with the intent to spur research and development in the critical area of credit rating. Besides being just a typical research project, it wanted to demonstrate the operational feasibility of its research and become a trusted source of credit information.
CRI currently covers more than 35,000 companies in 106 economies in Asia-Pacific, North America, Europe, Latin America, Africa, and the Middle East.
Change of methodology demands more processing power
RMI uses the forward intensity model (analogy to forward rate interest rate models), which relies on input variables for each firm (leverage, liquidity, profitability, stock index return, etc). This dependence is statistically estimated using a quasi-maximum likelihood estimation (QMLE), which is conducted monthly to update estimates of the parameters using the newest data.
With a recent change in the methodology to do the QMLE, RMI needed to depend even more than before on GPU computing.
Under the old methodology, the forecast horizon was up to two years and each month had a set of 13 parameters. By using GPUs instead of CPUs, processing time was cut from five hours to just 30 minutes.
The new methodology extends the forecast to five years and the coefficients for each variable need to be parameterised to regularise the estimates for the long term. Its credit analytics will estimate the probability of default for corporations around the globe
Impossible without GPUs
“Previously, we used GPUs to speed up processes. This made life easier but was not necessarily essential. However, this new methodology cannot be adopted without using GPUs,” said Dr Oliver Chen, Deputy Director of RMI.
To make the switch of methodologies, RMI decided to upgrade from the NVIDIA Tesla C2050 GPU to the NVIDIA Tesla K20 GPU accelerators, which can speed up applications by up to 10 times.
“We are still working on making the GPU code more efficient. Currently, the speed up from CPU to GPU is from one day to only five hours for a one-month update,” said Chen.