Why Multicurve Calibration Systems Often Favor Bootstrapping

# Comparison of multicurve calibration methods


# Comparison of multicurve calibration methods












It seems that there are mayor softwares around offering a multicurve framework based on bootstrap. I find this puzzling nowadays, given the distinct advantages of best-fit optimization methods and the hurdles in extending bootstrap techniques to the multicurve setting (e.g. cyclic interdependencies among curves, nontrivial products, usage of nonliquid instruments, overall coherence, dates mismatches, TOY effect, pre-first-tenor forwards, joint curves+term structure dynamics calibration etc). Therefore I must be missing some major drawback of full calibration or overestimating issues for bootstrap. The literature is outdated and mostly partisan, just like those I asked to dismiss either choice altogether. Could someone please shed some more light on these two possibilities and respective drawbacks?

(Moreover, Henrard mantains that for best-fit calibration a Newton-Rhapson optimizer suffices, while in my opinion the landscape is not so well-behaved... any views on this?)

## Answer by Phil H (score 2)

https://quant.stackexchange.com/a/14253

- Predictability - we all know what a bootstrapped curve will do when we shift a value. A minimisation, however, could jump to a new minimum at any moment. They also have unpredictable performance; sometimes a minimisation is fast, sometimes slow.

- Robustness - these codes have been around forever, and they work. New codes, not so much.

- Defendability - why is your X over the 17th so high? Because a cubic spline said so? It's much harder to defend something based on maths than something based on proven usefulness, even if it's more correct, and particularly when the boss has only known that type of construction.

- Speed - Bootstrapped codes, because they are predictable and because they are old, are fast. Optimisation methods can be 10-100x slower, (or more!) if you have a lot of complex criteria, or it happens to be a funny day for Futures vs Swaps.

- That's how other people do it - if you're trying to agree a stub rate with your counterparty, or the moneyness of a position for margining, anything where you have to justify the rates, it is much easier to go with something "normal". I heard it took Goldman (Goldman!) years to convince everyone that OIS was the right rate for margin accrual and discounting.

- For compatability with mid-office systems - if your P&L tomorrow morning is calculated by a mid-office setup which uses bootstrapped curves, you're going to want to know at least what that is likely to say, even if you also use a sophisticated spreadsheet

- Systems lag spreadsheets - following the above, the systems that are available lag market practices, so unless the software house has major customers that require more sophisticated curves, they just won't have caught up yet.

- It's not a significant difference - if you know that the difference is small, or that it is overwhelmed by something else that is far more important but impossible with a newer curve, then for all it's sophistication it is still not a better solution.