Skip to content

Downside (lower-partial-moment) semicovariance skater #58

Description

@microprediction

precise is the natural home for online covariance estimators/skaters, so flagging one that currently lives elsewhere and should migrate here eventually.

What

A downside semicovariance online estimator: an EWMA of the lower-partial-moment cross-products

S_t = (1-α) S_{t-1} + α · min(x_t - τ, 0) min(x_t - τ, 0)ᵀ

(threshold τ, e.g. 0 or a rolling mean), kept PSD with a floored diagonal. It captures how assets co-move on the downside — the lower-tail dependence a plain (full) covariance averages away — and is a drop-in covariance forecaster: same partial_fit / covariance_ interface as the other estimators.

Why it belongs in precise

  • It's a general online covariance-forecasting primitive, not specific to any one allocator — exactly the kind of thing precise collects (cf. the EWMA / shrinkage / factor skaters).
  • Downstream consumers can then pull it via a covariance hook rather than re-implementing it. In particular microprediction/allocation uses it to drive a tail-consistent allocation (the race responds to downside co-movement); it would consume a precise skater instead of shipping its own.

Current location / reference

Implemented today as DownsideSemicovariance in microprediction/allocation (allocation/moments.py), mirroring the EwmaCovariance interface. See allocation PR #11. Happy to port it here as a skater (or hand it off) when there's a good slot.

No urgency — parking it so it isn't lost.

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