Re: [Quantlib-dev] Jamshidian engine with start delay

[Peter C. <pca...@gm...>]

For zero days between option and start date the results are identical up to
1E-14, for example the old engine resp. new engine gives

0.06257488979659526
0.06257488979659394

for a 5y into 5y atm swaption vs. Euribor 6m, 3% forward flat yts, 1% volatility and
reversion, asof 30-04-2013.

The differences in the test suite displayed below are much bigger because there
a start delay of 2 business days applies. We could maybe produce cached
values with the old engine for a test case with expiry = start date
before moving to the new engine.

Thanks for your comments.

regards
 Peter


Ferdinando Ametrano <na...@am...> writes:

> I'm ok with your extension provided that the new values collapse back to the
> old ones (with a reasonable tolerance) in the case of expiry date being equal
> the value date.
> Is the data you've posted related to this case?
>
> It's a while now I do not work for a vol desk, but I would never underestimate
> the huge difference of analytic vs numerical methods when it comes to
> calibration.
>
> This said I would also add that I'm always amazed how poor the production setup
> is, even in very sophisticated banks. Old models stick around for very long
> time, just because of the huge effort required to update them in production
> systems. The multi-curve framework updates I've seen so far rival with Mary
> Shelley's Frankenstein approach
>
>
> On Sat, May 11, 2013 at 1:37 PM, Peter Caspers <pca...@gm...> wrote:
>
>     Hello,
>
>     in the JamshidianSwaptionEngine the option expiry date and the value date
>     of the underlying swap are handled a bit simplified assuming both dates
>     equal (see the warning in the code). Though the impact is usually not very
>     big we might want to improve this detail in the library ? See below for a
>     possible approach. Thank you Sebastian for our discussions on the topic.
>
>     Aside I would be interested whether the Jamshidian method is still in use
>     for model calibration in the world of multi curve enhanced models (where by
>     enhanced I mean something simple like a static spread correction) because I
>     believe the generalization of the method to this setting is not
>     straightforward. Also I feel that numerical integration does nearly a just
>     as efficient and accurate job and it directly allows for multiple curve
>     computations. Or do you ignore multi curve in the calibration phase and
>     only adjust the curves for the actual pricing ?
>
>     Back to Jamshidian and the start delay. Some theoretical background and
>     numerical examples can be found here
>
>     http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2246054
>
>     A possible implementation goes as follows. First we need to provide an
>     extended zerobond option method, which we can add to AffineModel in
>     model.hpp
>
>     https://github.com/pcaspers/quantlib/commit/
>     e16b4ea5ffbfe33bd6acd0ee6cb3ecd8a43f72a4
>
>     The default implementation uses the same simplification as mentioned above
>     ignoring the bond start delay. To improve the pricing in the
>     JamshidianEngine we have to overwrite this method in the model
>     implementations for which we want it. For the Hull White model I did it
>     here
>
>     https://github.com/pcaspers/quantlib/commit/
>     e8b5912cac2e236fe59a885e8cd1e2ed9243cc47
>
>     Finally we have to modify the Jamshidian engine a bit
>
>     https://github.com/pcaspers/quantlib/commit/
>     019f37a498846d9a6e89a897300f126c01d6ef86
>
>     (maybe we should keep some warning in the code because you are not forced
>     to support the start delay in your model implementations)
>
>     Not suprisingly the test suite breaks when comparing computation results to
>     cached values computed with the simplified engine, so the cached values
>     should be updated (given that we believe in the new engine)
>
>     1>  Testing Hull-White calibration against cached values...
>     1>  shortratemodels.cpp(126): error in
>     "QuantLib::detail::quantlib_test_case(&
>     ShortRateModelTest::testCachedHullWhite)": Failed to reproduce cached
>     calibration results:
>     1>  calculated: a = 0.0464041, sigma = 0.00579912, f(a) = 0.1158,
>     1>  expected:   a = 0.0488565, sigma = 0.00593662, f(a) = 0.121599,
>     1>  difference: a = -0.00245242, sigma = -0.000137495, f(a) = -0.00579896,
>     1>  end criteria = StationaryFunctionValue
>
>     regards
>       Peter
>
>    
>     ------------------------------------------------------------------------------
>     Learn Graph Databases - Download FREE O'Reilly Book
>     "Graph Databases" is the definitive new guide to graph databases and
>     their applications. This 200-page book is written by three acclaimed
>     leaders in the field. The early access version is available now.
>     Download your free book today! http://p.sf.net/sfu/neotech_d2d_may
>     _______________________________________________
>     QuantLib-dev mailing list
>     Qua...@li...
>     https://lists.sourceforge.net/lists/listinfo/quantlib-dev