Fix Merton kurtosis formula, GARCH standard errors, LSM order parameter, and SciPy compatibility; add regression tests

[SatishRockzz]

Summary

This PR fixes several verified bugs found during a systematic audit of the package, and adds a small regression test suite that locks the fixes in. Each fix was confirmed numerically (hand-derivation + Monte Carlo cross-check) before being applied.

Math / inference fixes

• Merton kurtosis (Processes.py): the excess-kurtosis formula used 3*sigJ**3 where the 4th cumulant of compound-Poisson-normal jumps requires 3*sigJ**4 (λ·E[J⁴] = λ·(μJ⁴ + 6μJ²σJ² + 3σJ⁴)). With (λ=0.8, μJ=0.1, σJ=0.5, σ=0.2) the old code returned 5.074; the true value is 2.635 (MC-confirmed 2.639). The sibling VG/NIG formulas were verified correct — this was a lone typo.
• GARCH standard errors (Processes.py): the objective minimized in fit_from_data was 2× the true Gaussian NLL (missing ½ factor), so the Hessian-based standard errors were understated by √2 (~29%), narrowing CIs and overstating significance. The optimum (MLE) is unchanged; only inference scales. This also makes the objective consistent with the class's own log_likelihood method.
• LSM order parameter (BS_pricer.py): the Longstaff–Schwartz order argument was silently ignored — np.polyfit was hardcoded to degree 2, so basis-order sensitivity studies returned identical results for every order. Now passed through. The error message for call options was also corrected (calls are unsupported, not "invalid type").

Compatibility / correctness fixes

• scp.mean(...) → np.mean(...) in all five MC pricers (scipy.mean was removed from SciPy years ago; the MC methods crash on any current SciPy). from scipy.linalg.misc import LinAlgError → from scipy.linalg import LinAlgError.
• mesh_plt in the Merton/VG/NIG pricers called self.PDE_price(...), which doesn't exist on those classes → self.PIDE_price(...).
• elif "MM": in VG fit_from_data is an always-true string literal, silently routing unknown method names into method-of-moments → proper comparison plus an explicit ValueError for unknown methods.
• probabilities.py: Q1/Q2 docstrings said P(X<k) while the integral computes P(X>k); corrected. Added a comment explaining the cf(-1.0000000000001j) normalizer (numerics untouched).

New tests (tests/test_regressions.py)

Codifies the cross-method checks the notebooks already perform interactively: kurtosis vs the cumulant formula, put–call parity to 1e-8, MC within 3 standard errors of the closed form (seeded), LSM order sensitivity (order=2 vs 4 now differ under identical seeds), Merton closed-form vs Fourier inversion. 5/5 passing on NumPy 2.5 / SciPy 1.18.

No behavior changes outside the fixes; all public APIs unchanged.

🤖 Generated with Claude Code