[tylrcc]

Closes #19.

Python port of the AdjointHestonModel C++ example, following the pattern of swap-adjoint.py:

• Same setup as the C++ original: DAX data from QuantLib's Heston test (spot 4468.17, 13 strikes x 8 expiries = 104 prices, fit on prices via HestonModelHelper with PriceError), COSHestonEngine(model, 12, 75), Levenberg-Marquardt at tolerance 1e-12 starting from (0.1, 1.0, 0.5, -0.5, 0.1), then a forward-start call (strike set in 6M, expiry 18M) priced with AnalyticHestonForwardEuropeanEngine.
• Sensitivities to all 114 market inputs via the implicit function theorem: Hessian blocks of the calibration objective from central differences of its adjoint gradient (11 recordings), one more recording for dV/dp and the direct dV/dm, and a 5x5 solve (numpy, already in the examples' requirements).
• The eight largest sensitivities are cross-checked inside the example by bump-and-recalibrate.

Testing, with two caveats I want to be explicit about:

1. The current QuantLibAAD wheel is not reachable from my sandbox, so I ran the identical example code against QuantLib-Risks 1.33.3 (the direct predecessor of QuantLibAAD; the tape-construction lines differ slightly between the two APIs). The full flow runs end to end and the cross-check agrees: e.g. zero rate 524d IFT 2887.910476 vs bump-and-recalibrate 2888.227150 (1.1e-4 relative); zero rate 165d -1040.308095 vs -1040.250569; worst of the eight is 4.3e-4 relative.
2. In that older 1.33.3 build, COSHestonEngine segfaults when pricing a helper, so for the local run I substituted AnalyticHestonEngine for the calibration engine only (the submitted file uses COS, as in the C++ example and as the issue describes; plain QuantLib 1.44 prices the same helper with COS without issue, giving the same model value as the analytic engine to 6 significant digits). With the analytic engine the local calibration lands on theta 0.0722, kappa 3.6200, sigma 1.2293, rho -0.5790, v0 0.1289 and a forward-start value of 480.64; the COS calibration in the current build may differ slightly in the last digits. If the COS engine misbehaves in the current build as well, the analytic engine is a drop-in fallback for this example.

Note: this repo requires a signed CLA before merge (per CONTRIBUTING) — I understand it will be handled through the CLA flow.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}