Skip to content

comparison.md

Latest commit

 

History

History
95 lines (78 loc) · 5.39 KB

comparison.md

File metadata and controls

95 lines (78 loc) · 5.39 KB
Raw

cvxcla vs. Bailey–López de Prado (2013)

A comparison of the cvxcla paper (Schmelzer, The Critical Line Algorithm as implemented in cvxcla) with Bailey, D. H. and López de Prado, M. (2013), "An Open-Source Implementation of the Critical-Line Algorithm for Portfolio Optimization", Algorithms 6(1):169–196 (SSRN abstract 2197616).

This is not an arbitrary comparison: Bailey–López de Prado (BLdP) is the bailey2013 reference that cvxcla benchmarks against, and the PyPortfolioOpt CLA used in the experiments is an implementation of it. The comparison therefore doubles as a fact-check of cvxcla's §8 ("Relation to the Bailey–López de Prado algorithm").

Shared lineage

Both papers compute the entire efficient frontier exactly as a finite list of Markowitz turning points, in Python. They share several concrete design choices:

  • First turning point. The same greedy construction: start every weight at its lower bound, fill assets in descending expected-return order until the budget binds; the last (partially filled) asset is the sole free asset (BLdP initAlgo, Snippet 2; cvxcla init_algo, §3).
  • Minimum-variance endpoint. Both append an explicit turning point at λ = 0, and both explicitly note this departs from Niedermayer et al., who keep searching for turning points at negative λ.
  • Maximum-Sharpe portfolio. Both locate it by a one-dimensional search over the convex combination of the two turning points that bracket the discrete Sharpe maximum (BLdP: golden-section search, Snippets 13–14; cvxcla: one-dimensional search, §6).
  • Frontier interpolation. Both reconstruct intermediate frontier portfolios as convex combinations of adjacent turning points.

Head-to-head differences (verified against the BLdP source)

Axis Bailey–López de Prado (2013) cvxcla
Covariance representation Explicit dense matrix, sliced via getMatrices / reduceMatrix (Snippet 5) Operator protocol (matvec, solve_free, cross) + structured backends
Free-block linear algebra Explicit np.linalg.inv(covarF) at each step (Snippets 3, 7–9) Single block-eliminated KKT solve via an m×m Schur complement; no explicit inverse
Asset-entry search Full inverse recomputed inside the candidate loop for i in b (Snippet 3) → O(n) inverses per turning point Vectorised event search; one solve per step
Structured risk models Not supported Diagonal-plus-low-rank factor backend via the Woodbury identity (no n×n matrix ever formed)
Constraints Box bounds + single (budget) equality only (inputs: mean, covar, lB, uB) Arbitrary m linear equality constraints A w = b
Degeneracy handling No general anti-cycling rule; a single hack (mean[-1]+=1e-5) for all-equal means Bland-style lowest-index tie-break + √ε slope floor (§4.3–4.4)

The central claim checks out

cvxcla's most load-bearing §8/§9.3 claim — that PyPortfolioOpt "recomputes that full inverse once for every candidate" in the asset-entry search — is literally BLdP Snippet 3: covarF_inv = np.linalg.inv(covarF) sits inside the for i in b candidate loop. This single structural fact explains the ≈340× runtime gap and is faithfully represented in cvxcla.

One nuance worth noting: BLdP's removal branch (Snippet 7) computes the inverse once outside its loop, so the per-candidate rebuild is specific to the addition branch — which is exactly how cvxcla phrases it. The characterization is accurate and fair.

Where each paper is stronger

Bailey–López de Prado, on its own terms

  • Pedagogy. It is a tutorial: 17 complete code snippets, step by step. It established the open-source CLA reference (it became PyPortfolioOpt and is widely cited).
  • Novelty-at-the-time. In 2013 it filled a genuine gap — the first documented open-source CLA in a scientific language (prior art was the Markowitz–Todd VBA-Excel implementation). That is higher-impact novelty than an interface refinement.
  • Ground-truth validation. It reproduces the canonical Markowitz–Todd 10-asset example to 15 significant figures.

cvxcla

  • Scale demonstrated. BLdP only ever runs its n = 10 toy example. cvxcla shows n = 640 benchmarks, an n = 20,000 factor stress test, and a real n = 494 S&P 500 frontier.
  • Robustness. BLdP never discusses conditioning, cycling, or tolerances and has no general anti-cycling rule. cvxcla hardens the event logic and honestly documents where it still fails (§9.6).
  • Representation-agnosticism. The covariance-operator abstraction (and the resulting sub-quadratic factor backend) is the one element not present in BLdP — or, per cvxcla, in any standard CLA implementation.
  • Independent exactness check. cvxcla validates against a general convex QP solver (CVXPY/Clarabel) at solver precision, rather than against a single known reference answer.

Verdict

BLdP is a pedagogical reference implementation that established the open-source baseline; cvxcla is an engineering / performance paper that re-architects that baseline — operator abstraction, block elimination, structured backends, hardened degeneracy handling — and quantifies the payoff. cvxcla's characterization of BLdP is accurate and fair; if anything it under-credits BLdP's pedagogical completeness, which a single sentence in §8 could acknowledge.