Remove min cov det

[robertmartin8]

Describe the bug
The minimum covariance determinant method gives nonsensical covariance matrices.

df = get_data()
S = risk_models.min_cov_determinant(df, random_state=8)
np.diag(S)

array([2.41028948e-03, 2.03474589e-01, 0.00000000e+00, 0.00000000e+00,
       1.57413320e-01, 5.30639582e-02, 2.86881436e-01, 7.30126563e-02,
       7.33460893e-02, 0.00000000e+00, 6.48495304e-02, 6.14756936e-04,
       2.01800632e-02, 4.18976215e-02, 1.60823736e-01, 2.47871090e-01,
       5.52724202e-05, 7.78379082e-02, 9.82877705e-02, 1.31182474e-01])

Compare with sample covariance:

array([0.09321114, 0.20753691, 0.13720115, 0.10095819, 0.37539405,
       0.08310901, 0.39058571, 0.06909179, 0.18019232, 0.08011557,
       0.06555993, 0.24552417, 0.32116207, 0.05440377, 0.27222973,
       0.26954874, 0.12030304, 0.08550166, 0.14689262, 0.15258853])

min_cov_det is giving values several orders of magnitude below the other methods. There is an error in the code where we set NaNs to 0, but even after fixing this to dropna, the result is still far too low:

array([0.01921634, 0.02855301, 0.02299209, 0.05725663, 0.02720913,
       0.01819129, 0.16490903, 0.01775783, 0.03262536, 0.02917394,
       0.01434993, 0.06429909, 0.32185691, 0.01727326, 0.13953319,
       0.04986281, 0.01974825, 0.01409136, 0.01995168, 0.0199959 ])

I think the implementation of MCD is misguided – I only included it (many years ago) because I wanted as many risk models as possible, without really thinking too much about their applicability.

I will be deprecating it for v1.4.1, and removing in v1.5.

[phschiele]

@robertmartin8 I tested min_cov_det on a few different samples, sometimes the covariances were in a similar range, sometimes a lot smaller. I also compared it against skleanrn's MinCovDet(random_state=8).fit(returns_df).covariance_ * 252, which again gave quite different results.

I am not very familiar with this method but at first glance, there seem to be some issues. I would also be in favor of deprecating it 👍