Skip to content

Online downside / semicovariance (lower-partial-moment) covariance skater #57

Description

@microprediction

Online downside / semicovariance (lower-partial-moment) covariance skater

Why

The allocation package's Thurstone "ability tilt" becomes tail-risk-consistent when the correlated race is driven by a downside covariance instead of the ordinary (symmetric) covariance. The result (tail-consistency, written up in allocation/papers/thurstone-portfolios): a cluster of assets that crash together (lower-tail dependence) de-duplicates to a single competitor and is de-weighted, with the decorrelated hedge up-weighted — a de-weighting keyed to genuine tail co-movement that a symmetric Σ cannot see.

A simulation sweep (A + 99-name Clayton lower-tail cluster) shows the downside semicovariance recovers almost all of the effect that the full tail copula produces, while full covariance is tail-blind. So the practical lever is an online downside-covariance estimator.

Ask

A skater-style online estimator (same partial_fit / covariance_ shape as the existing precise covariance skaters, so it drops straight into ThurstonePortfolio(covariance=...) in allocation) that maintains a co-lower-partial-moment matrix:

S_ij = E[ min(r_i - τ_i, 0) · min(r_j - τ_j, 0) ]

with:

  • online/EWMA means for the threshold τ (running mean, or a fixed target return),
  • the usual streaming partial_fit(dict-or-row) / covariance_ interface,
  • an option to return the downside correlation (unit diagonal) and a PSD / nearest-correlation repair (the race needs a valid correlation to sample from),
  • ideally a precise-style f-function wrapper for the buffered/streaming convention.

Note

I think precise used to carry a semivariance estimator — if so, this may be a restore/port rather than a fresh build. Worth checking the history.

Acceptance

  • streams row-by-row and over a changing key set (river-style), like the other skaters;
  • matches a batch downside-semicovariance on static data;
  • handles τ = running mean and τ = fixed threshold;
  • returns a valid (PSD, unit-diagonal) downside correlation on request.

Filed from the tail-consistent Thurstone work in microprediction/allocation.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions