Summary

• Add EfficientMVSK optimizer backed by yand-mvsk, which implements Yau's Affine-Normal Descent algorithm (arXiv:2604.25378)
• Extends BaseOptimizer (like CLA — no cvxpy dependency); yand-mvsk is an optional dependency in all_extras
• 18 new tests, all passing; zero regressions on existing test suite

Motivation

Classical mean-variance optimization ignores skewness (crash risk) and kurtosis (tail risk). MVSK optimization jointly considers all four moments, producing portfolios that are more robust to tail events.

Previous MVSK implementations were impractical because they required materializing O(n3) coskewness and O(n4) cokurtosis tensors. The YAND algorithm computes all moment operations via O(Tn) matrix-vector products, solving 800-asset problems in <100ms:

Usage

from pypfopt import EfficientMVSK
mv = EfficientMVSK(returns, gamma=6.0) # or EfficientMVSK.from_prices(prices)
weights = mv.min_mvsk()
mu, vol, sharpe, skew, kurt = mv.portfolio_performance()

Changes

• pypfopt/efficient_frontier/efficient_mvsk.py — new EfficientMVSK class
• pypfopt/efficient_frontier/__init__.py — register export
• pypfopt/__init__.py — top-level export
• pyproject.toml — add yand-mvsk>=0.2.0 to all_extras
• tests/test_efficient_mvsk.py — 18 tests covering convergence, correctness (beats equal-weight), bounds, performance metrics, serialization

Test plan

• All 18 new MVSK tests pass
• All 87 existing tests pass (EfficientFrontier, CLA, CVaR)
• yand-mvsk gracefully skipped when not installed (soft dependency)