Financial SolutionsRisk ManagementFund ManagementCapital Markets

Prism

Overview

PRISM, designed for the National Stock Exchange, is a risk management software which has the capability to calculate real-time Value-at-Risk of the clearing members at the Exchange at client levels, at a speed of over 100 trades per second.

PRISM was developed by Infofin in alliance with Indira Gandhi Institute of Development Research (IGIDR) and StarCom Software. IGIDR provided the algorithm for calculation of VaRs.

Features

PRISM provides the following features

  • VaR calculations at client level.

  • Management of positions for each client. The trade feed from the Exchange is fed directly to PRISM, without the use of external databases to provide input data.

  • Full Monte Carlo calculation of VaR for portfolios containing Futures and Options on the Index; without any compromises on mathematical correctness.

  • Volatility forecasting using time-series econometrics.

  • VaR implementation focusing on close-to-close risk (i.e. not intra-day price movements).

  • Collateral requirements calculated as the larger of VaR and the regulator's requirements (whatever they may be); the system always computes both and picks the larger value.

  • Intra-day marking to market.

  • Information structures based on a tree of clearing member, trading member, client. VaR is summed up from clients to generate TM collateral requirement and collateral is summed up from TMs to generate the CM collateral requirements.

  • Broad design goals of: 100 clearing members, 1000 trading members, 1,000,000 clients, 10 trades/second (i.e 6,000 trades/minute). Processing 100 trades/second implies doing 200 VaR calculations per second.

  • Implementation of high throughput for VaR calculations using parallel programming.

Technology
  • Parallel computation using off-the-shelf commodity equipment and MPI, a public domain industry standard in the supercomputing industry.

  • Fault tolerance at the level of "child" processors. PRISM is designed to cope with breakdowns of any of the Unix workstations charged with VaR calculations breaks down.

  • Full support for symmetric multi-processing; either mother or child computers could be multiprocessing system.

  • GUI supporting real-time display of information.

Architecture

System Architecture

PRISM uses the model of a ``mother'' machine and a cluster of ``child'' CPUs where the VaR computation is done. The ``mother'' accepts the Exchange trade feed and passes on calculations intelligently to an idle ``child'' for processing.

If any of the ``children'' fail to return a response within a pre-defined time, the ``mother'' passes on the calculation to another idle ``child''. This ensures that the system is able to cope with situations like hardware failure and always sends back a correct response.

The core of the system is written in C using MPI as the messaging interface between the clustered machines. A Java front end monitors the health of the system and gives real time updates on the VaR per trade.

PRISM has been installed and tested on Linux and different variants of Unix. Benefits

  • Real-time computation and monitoring of Value-at-Risk.

  • Significant savings in hardware and operating software costs.

  • Virtually unlimited scalability by adding commodity CPUs.

  • Hardware independence.

Benchmark

PRISM has been stress-tested at CDAC on Param 10000 (India's Premier Super Computer). As VaR calculations take the major system load, the benchmark of PRISM is calculated in terms of nos of VaRs calculated per second. The number of VaRs/sec is a function of number of children added. The scalability of children is shown below in the table.

Sr. Children Trades/sec VaRs/sec
1 1 100 68
2 2 100 136
3 4 150 296
4 6 200 440
5 8 250 587
6 10 400 727
7 12 450 788
8 14 500 861
9 16 600 1146
10 24 900 1492

Here the children denotes the number of CPUs the children are running on. The trades/sec is the number of trades fired. For every trade, PRISM calculates 2 VaRs. The VaRs/sec is the number of VaRs calculated by PRISM per second.

Thus the scalability of PRISM can be seen from the above table and the limit is only the hardware used. Design, development and deployment

From the storyboard to implementation, PRISM took a total of 32 man months. The project commenced in August 1999, went through its advanced second phase in April 2000 and was finally commissioned on June 12th, 2000.

Client

National Stock Exchange of India

Top