When the Bank of Tokyo-Mitsubishi Bank Ltd. took an $83 million derivatives bath in March, it blamed the losses on bad modeling, as though it was just another bad hair day.
But as those traders should have known, how derivatives models predict pricing and the relative movements of the underlying assets can mean millions of dollars in the credit or debit columns.
What makes the difference? How people use the resources at the bank. Making good predictions of price moves means hours in front of the computer pushing algorithmic models to the max by running "what if" scenarios to see what breaks, and thus avoiding mistakes before they hit the trading floor.
While this sort of precaution seems to be mere common sense, it's more rare than it should be, says Tanya Styblo Beder, a principal of New York's Capital Market Risk Advisorsomarking a crucial difference in performance from one trader to another. "What sets apart people who manage model risk well from those who don't is the degree to which people have stress-tested the assumptions, data and methodology that they used to perform their calculations," says Beder. "If using five different models and a whole bunch of assumptions give you numbers right on top of each other, one should have very high confidence in one's modeling." Beder, who tracks publicly-disclosed derivatives losses, says there have been $22 billion of them since 1983 in what she says is a $53 trillion market.
Another good idea is to put the right features into your software, says David Penny, vp of software development for Toronto-based Algorithmics Co. Good pricing software can weed out the bad data points that often slip into the torrent of daily quotes and skew pricing results, he says. The programming can tag these exceptions, run the model with and without them, and discard them if they produce weird results.
Derivatives shops can also take a page out of real rocket science, says Penny, and integrate redundant systemsoroutine in systems like the space shuttleointo the overall pricing program. This approach would compare the pricing results with triangulated relationships between several linked prices with a recognized relationship to the price in question.
Programming and hardware costs wouldn't be cheap, he acknowledges, but would the cost reach $83 million? "Software developed in-house would cost about three months of development," he says. "It'd be peanuts." --reinbach tfn.com