Exponentially concave functions and high dimensional stochastic portfolio theory

TL;DR

This paper demonstrates that in high-dimensional stochastic markets with specific tail behaviors and volatility conditions, it is possible to construct portfolios that outperform the market over very short periods with high probability, revealing phase transitions based on tail indices.

Contribution

It introduces a novel approach using exponentiated concave functions and high-dimensional geometry to construct portfolios with arbitrage properties in stochastic portfolio theory.

Findings

High-dimensional portfolios can achieve relative arbitrage over short periods.

Phase transitions occur depending on the tail index of market weights.

Construction relies on properties of regular variation and convex geometry.

Abstract

We consider the following problem in stochastic portfolio theory. Are there portfolios that are relative arbitrages with respect to the market portfolio over very short periods of time under realistic assumptions? We answer a slightly relaxed question affirmative in the following high dimensional sense, where dimension refers to the number of stocks being traded. Very roughly, suppose that for every dimension we have a continuous semimartingale market such that (i) the vector of market weights in decreasing order has a stationary regularly varying tail with an index between $-1$ and $-1/2$ and (ii) zero is not a limit point of the relative volatilities of the stocks. Then, given a probability $\eta < 1$ arbitrarily close to one, two arbitrarily small $\epsilon, \delta >0$, and an arbitrarily high positive amount $M$, for all high enough dimensions, it is possible to construct a functionally generated portfolio such that, with probability at least $\eta$, its relative value with respect to the market at time $\delta$ is at least $M$, and never goes below $(1-\epsilon)$ during $[0, \delta]$. There are two phase transitions; if the index of the tail is less than $-1$ or larger than $-1/2$. The construction uses properties of regular variation, high-dimensional convex geometry and concentration of measure under Dirichlet distributions. We crucially use the notion of $(K,N)$ convex functions introduced by Erbar, Kuwada, Sturm in the context of curvature-dimension conditions and Bochner's inequalities.