covariance matrix constructed from the feature space data points to detect and reduce multicolinearity in the given cluster of index positions while still estimating the beta of underlying index (shaving off multicolinear and similalry clustered stocks in the index) PCA/K-means in service of live/simulation-live portfolio management.
the unsupervised methods currently in 0.1-0.2v serve as feature-space diagnostics informing model interpretation. Their natural extension — operating on the return covariance structure for index-tracking position reduction and wash-sale replacement selection — is deferred to the portfolio construction layer, which requires at least a simulation environment (likely with RL instead of GBT) as a prerequisite.
Unsupervised layer will work hand in hand with the classification model (whether that be a supervised or RI based) for the lot to make a $\sigma_TE$ optimization decision of (if colinear lots available) under the assumption of an already dim-reduced portfolio of whether to in service of reducing TE to replace a harvested lot with an immediate fully or semi-colinear substitute or the second option of buying back the same asset 30d later (wash clock rule). This is closer to how applied quantitative systems are structured: different ML methods solve different subproblems rather than one giant model doing everything.
covariance matrix constructed from the feature space data points to detect and reduce multicolinearity in the given cluster of index positions while still estimating the beta of underlying index (shaving off multicolinear and similalry clustered stocks in the index) PCA/K-means in service of live/simulation-live portfolio management.
the unsupervised methods currently in 0.1-0.2v serve as feature-space diagnostics informing model interpretation. Their natural extension — operating on the return covariance structure for index-tracking position reduction and wash-sale replacement selection — is deferred to the portfolio construction layer, which requires at least a simulation environment (likely with RL instead of GBT) as a prerequisite.
Unsupervised layer will work hand in hand with the classification model (whether that be a supervised or RI based) for the lot to make a$\sigma_TE$ optimization decision of (if colinear lots available) under the assumption of an already dim-reduced portfolio of whether to in service of reducing TE to replace a harvested lot with an immediate fully or semi-colinear substitute or the second option of buying back the same asset 30d later (wash clock rule). This is closer to how applied quantitative systems are structured: different ML methods solve different subproblems rather than one giant model doing everything.