Searcharxiv⌕ Search

arXiv · cond-mat/0111503

Noisy Covariance Matrices and Portfolio Optimization

Abstract

According to recent findings [1,2], empirical covariance matrices deduced from financial return series contain such a high amount of noise that, apart from a few large eigenvalues and the corresponding eigenvectors, their structure can essentially be regarded as random. In [1], e.g., it is reported that about 94% of the spectrum of these matrices can be fitted by that of a random matrix drawn from an appropriately chosen ensemble. In view of the fundamental role of covariance matrices in the theory of portfolio optimization as well as in industry-wide risk management practices, we analyze the possible implications of this effect. Simulation experiments with matrices having a structure such as described in [1,2] lead us to the conclusion that in the context of the classical portfolio problem (minimizing the portfolio variance under linear constraints) noise has relatively little effect. To leading order the solutions are determined by the stable, large eigenvalues, and the displacement of the solution (measured in variance) due to noise is rather small: depending on the size of the portfolio and on the length of the time series, it is of the order of 5 to 15%. The picture is completely different, however, if we attempt to minimize the variance under non-linear constraints, like those that arise e.g. in the problem of margin accounts or in international capital adequacy regulation. In these problems the presence of noise leads to a serious instability and a high degree of degeneracy of the solutions.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Szilard Pafka, Imre Kondor. 2001-11-27. Noisy Covariance Matrices and Portfolio Optimization. https://doi.org/10.1140/epjb%2Fe20020153. https://searcharxiv.com/abs/cond-mat/0111503

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Ergotropy of macroscopic classical systems

We address the fundamental question of what is the maximal amount of energy (a.k.a. ergotropy) that can be extracted from a macroscopic classical system by means of a cyclical variation of an external parameter. We show that if the system is prepared in a statistical state of the form $ρ= F(H)$, with $H$ the Hamiltonian and $F$ a generic limited function, then the ergotropy is vanishing in the {large $N$} limit, provided the thermodynamic entropy obeys the scaling $\mathcal S(E) \simeq N^αs ({E}/{N^β})$ with $α,β>0$ and $s$ some strictly increasing function with certain reasonable properties. We stress that such ubiquitous condition also ensures the equivalence of the so-called "surface" and "volume" entropies. We illustrate this by introducing the notion of ``ergotropic rearrangement'' of a phase space density which we then apply to find the exact ergotropy expression of systems with constant (positive or negative) heat capacity being evenly distributed over energy shells in phase space. We discuss cases where the above mentioned equivalence does (does not) break and the ergotropy does (does not) vanish accordingly, in the large $N$ limit. The presented results shed new light on the statistical-mechanical foundations of the second law of thermodynamics.

cond-mat.stat-mech↗

Spectral Softening and the Structural Breakdown of Thermodynamic Equilibrium

Reversible quasistatic thermodynamics requires normalizable instantaneous equilibrium states and dynamics allowing their attainment or tracking. We examine these requirements in a driven two-dimensional quadratic Hamiltonian with an angular-momentum coupling and a spatially uniform momentum shift $p_c(t)$, representable by a time-dependent vector potential. Its undriven part has normal-mode frequencies $ω_\pm=ω_1\pmω_2$ and the Fock--Darwin structure of a charged oscillator in a perpendicular magnetic field. At the exact soft-mode endpoint $ω_-=0$, confinement is lost in one normal-mode sector. For every strictly positive $ω_-$, the canonical Gibbs state remains normalizable; only at the endpoint does the canonical trace cease to be finite. In the quasi-soft regime, where $ω_-\to0^+$, exact adjacent-level couplings yield a dimensionless adiabatic parameter proportional to $|\dot p_c|/ω_-^2$. The standard adiabatic criterion requires this parameter to remain much smaller than unity. At a fixed occupation and a fixed nonzero driving rate, however small, this parameter grows without bound as $ω_-\to0^+$, eventually violating this criterion. In an open system, thermal tracking additionally depends on relaxation relative to the driving timescale. An explicit linear system--bath coupling illustrates why the relaxation rate cannot be inferred from the soft-mode spacing alone. Finally, the thermal Wigner function and its classical limit show that phase-space confinement is lost in both descriptions at the soft-mode endpoint. Taken together, our results show that canonical equilibrium is lost at $ω_-=0$. For $ω_->0$, the Gibbs state exists, while the ability to track it dynamically depends on the driving and dissipation.

cond-mat.stat-mech↗

Path-Extrema Upper Bounds on Mean Entropy Production

Fluctuation relations imply $\langleΣ_T\rangle\ge0$. What can path extrema tell us about an upper bound? A ceiling on the positive entropy-production maximum alone gives no improvement over the trivial endpoint bound: rare negative excursions can supply the exponential weight required by the fluctuation relation. Given the joint law of the running extrema $L_T$ and $H_T$ of $M_t=e^{-Σ_t}$, we derive an upper bound $U_{\rm ext}$ by allocating a normalization budget across realized intervals, favoring those that yield the most entropy per unit cost. The exact identity $\langleΣ_T\rangle=U_{\rm ext} - A - C$ separates an allocation gap across extrema classes from a curvature gap within each class. The bound and decomposition require the terminal relation $\langle M_T\rangle=1$ and the pathwise extrema constraint, not martingality; $U_{\rm ext}$ is sharp in this minimal class. Martingality imposes further constraints, and attainment of the same optimum under them remains open in general. For a three-state Markov-jump cycle, we obtain the finite-time joint law of $(Σ_T,L_T,H_T)$ by reduction to a biased random walk. A companion Monte Carlo study illustrates how undersampling rare negative excursions prevents the fluctuation relation from being resolved at long times. The result is a structural upper constraint, not a more efficient estimator of the mean from complete trajectories.

cond-mat.stat-mech↗