CHAPTER 13Check-Pointing and Calibration

In the previous chapter, we learned to differentiate simulations against model parameters in constant time, and produced microbuckets (sensitivities to local volatilities) in Dupire's model with remarkable speed. In this chapter, we present the key check-pointing algorithm, apply it to differentiation through calibrations to obtain market risks out of sensitivities to model parameters, and implement superbucket risk reports (sensitivities to implied volatilities) in Dupire's model.


Reducing RAM consumption with check-pointing

We pointed out in Chapter 11 that, due to RAM consumption, AAD cannot efficiently differentiate calculations taking more than 0.01 seconds on a core. Longer calculations, which include almost all cases of practical relevance, must be divided into pieces shorter than 0.01 seconds.core, and differentiated separately over each piece, wiping RAM in between and aggregating sensitivities in the end.

How exactly this is achieved depends on the instrumented algorithm. This is particularly simple in the case of path-wise simulations, where sensitivities are computed path by path and averaged in the end. But this solution is specific to simulations. This chapter discusses a more general solution called check-pointing. Check-pointing applies to many problems of practical relevance, in finance and elsewhere. Huge successfully applied it to the differentiation of multidimensional FDM in [90]. Huge and Savine ...

Get Modern Computational Finance now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.