
End of the bumping grind?
In quest for processing power, banks are choosing 'unnatural' calculus over fancy chips

If you were trying to dramatically speed up risk and sensitivity calculations – a process that traditionally involves thousands of repetitive sums, with the inputs being changed marginally, or ‘bumped' each time – the output of a single calculation may not seem the most promising place to start. In the slightly odd world of adjoint algorithmic differentiation (AAD), though, it is the key to running calculations tens, hundreds or even thousands of times faster than the most popular rival methods.