Log-Optimal Portfolio

• Log-optimal portfolio is a strategy that maximizes expected logarithmic growth of wealth by optimizing log-utility under fully invested, no-shorting constraints.
• It uses constant rebalancing and robust estimation methods, including ambiguity sets and online learning, to handle uncertainty in asset returns.
• Extensions of the model include continuous-time semimartingale formulations, duality via deflators, and cost-sensitive adjustments incorporating transaction costs and risk controls.

A log-optimal portfolio is a portfolio that maximizes the expected logarithmic growth of wealth. In the discrete-time formulation with random return vector $r\in\mathbb R^n$ and portfolio weight vector $w\in\mathbb R^n$, one-period wealth satisfies $V(1)=V(0)(1+w^Tr)$, and over $N$ i.i.d. periods,

$V(N)=V(0)\prod_{k=0}^{N-1}(1+w^Tr(k)).$

The associated expected per-period log-growth rate is

$$g(w)=\frac1N\mathbb E[\ln V(N)/V(0)]=\mathbb E[\ln(1+w^Tr)].$$

Under the fully invested, no-shorting simplex constraints $\Delta=\{w\in\mathbb R^n:\mathbf 1^Tw=1,\;w\ge 0\}$, the classical problem is

$\max_{w\in\Delta}\;\mathbb E[\ln(1+w^Tr)].$

Across the literature this object is also described as the Kelly portfolio, the growth-optimal portfolio, or, in many continuous-time settings, the numéraire portfolio (Hsieh, 2022, Cuchiero et al., 2016).

1. Classical criterion and foundational formulations

The classical log-optimal problem is a concave maximization over portfolio weights. In the constant-rebalancing setting, the objective is the expected logarithmic growth (ELG) of wealth, and the admissibility condition $r_i>-1$ ensures $1+w^Tr>0$ and hence well-defined log-utility (Hsieh, 2022). In the notation of the General Log-Optimal Strategy (GLOS), one chooses $w\in\mathbb R^n$0 to solve

$w\in\mathbb R^n$1

Under mild regularity conditions, the GLOS is stated to have three properties: “Information Benefit,” “Greed,” and “Long-Term Superiority” (Guo et al., 2018).

A central structural fact is that the log-optimal objective is sensitive to compounding rather than to linearized expected return. This distinguishes it from mean-variance or one-step expected-return criteria. In the frequency-based formulation with $w\in\mathbb R^n$2-period compound return $w\in\mathbb R^n$3, the ELG is

$w\in\mathbb R^n$4

and the optimizer depends on the rebalancing interval unless further structure is imposed (Hsieh et al., 2023, Hsieh, 2021).

Several papers isolate special cases in which the solution collapses to a boundary portfolio. If asset $w\in\mathbb R^n$5 is dominant, meaning it is relatively more attractive than every other asset in the sense that

$w\in\mathbb R^n$6

for every $w\in\mathbb R^n$7, then the unique log-optimal weight is $w\in\mathbb R^n$8, and

$w\in\mathbb R^n$9

This boundary result is therefore conditional on the dominant-asset hypothesis rather than a universal property of log-optimality (Hsieh, 2021).

The same basic criterion has also been generalized by replacing the unweighted log-growth with weighted log-growth. In these discrete-time formulations, the weighted logarithmic growth minus a cumulative weighted entropy term is shown to be a supermartingale, and the optimal martingale strategy is proportional betting. The papers further state that the form of the optimal strategy does not depend on the form of the weight function, although the optimal rate does (Suhov et al., 2015, Kelbert et al., 2017).

2. Numéraire interpretation, universality, and model-free asymptotics

In probabilistic portfolio theory, the log-optimal portfolio is closely linked to the numéraire portfolio. In the discrete-time ergodic Markov setting on the simplex $V(1)=V(0)(1+w^Tr)$0, the local growth-rate functional is

$V(1)=V(0)(1+w^Tr)$1

and any measurable selector $V(1)=V(0)(1+w^Tr)$2 achieving the maximum is called the log-optimal or numéraire map. Under ergodicity and integrability,

$V(1)=V(0)(1+w^Tr)$3

The same paper shows that, under appropriate assumptions, the optimal long-run yield coincides for Cover’s universal portfolio, the hindsight-best portfolio in a regular class, and the log-optimal numéraire portfolio, in both discrete and continuous time (Cuchiero et al., 2016).

A distinct model-free route uses calibration and Blackwell approachability. In that construction, one forms an “artificial” conditional law $V(1)=V(0)(1+w^Tr)$4 for returns from well-calibrated forecasts, then solves at each round

$V(1)=V(0)(1+w^Tr)$5

The resulting wealth process asymptotically performs at least as well as any stationary portfolio that redistributes investment at each round using a continuous function of side information, and the method makes no stochastic assumptions about market values (V'yugin, 2014).

Recent work on stationary markets with latent side information pushes this asymptotic viewpoint further. There the dynamic log-optimal strategy, defined via the one-step conditional maximization

$V(1)=V(0)(1+w^Tr)$6

is shown to admit an asymptotically equivalent constant strategy when the joint process is stationary. Specifically, there exists a possibly random constant portfolio $V(1)=V(0)(1+w^Tr)$7 such that

$V(1)=V(0)(1+w^Tr)$8

This suggests that, in a stationary market, the asymptotic growth rate of the optimal dynamic strategy may decay into an equilibrium state matching that of a constant strategy, even when side information is partly unobservable (Lam, 12 Jan 2025).

These universality results place the log-optimal portfolio at the intersection of stochastic optimization, game-theoretic forecasting, and online learning. They also show that long-run growth optimality is not confined to i.i.d. return models.

3. Continuous-time semimartingale theory and duality

In the general semimartingale framework, the log-optimal portfolio is formulated on a filtered probability space with discounted risky-asset price process $V(1)=V(0)(1+w^Tr)$9. For a predictable integrand $N$0, wealth is

$N$1

Admissibility is imposed under “no-unbounded-profit-with-bounded-risk” (NA1), equivalently

$N$2

together with an integrability condition on the negative logarithm of jumps. The optimization problem is

$N$3

A complete characterization is given without assuming NFLVR. For $N$4-special, quasi-left-continuous semimartingales, the following are equivalent: existence of a log-integrable deflator, existence of a unique predictable minimizer $N$5, existence of a unique optimal deflator $N$6 minimizing $N$7, and existence of a unique $N$8 maximizing expected log-utility. Moreover,

$N$9

The optimal portfolio is characterized locally as the minimizer of a convex Hamiltonian-like function $V(N)=V(0)\prod_{k=0}^{N-1}(1+w^Tr(k)).$0, with first-order condition

$V(N)=V(0)\prod_{k=0}^{N-1}(1+w^Tr(k)).$1

(Choulli et al., 2018).

In many enlarged-filtration models the same portfolio coincides with the numéraire portfolio whenever it exists. For a market model stopped at a random time $V(N)=V(0)\prod_{k=0}^{N-1}(1+w^Tr(k)).$2, with progressive enlargement from $V(N)=V(0)\prod_{k=0}^{N-1}(1+w^Tr(k)).$3 to $V(N)=V(0)\prod_{k=0}^{N-1}(1+w^Tr(k)).$4, the stopped price process is $V(N)=V(0)\prod_{k=0}^{N-1}(1+w^Tr(k)).$5. The log-optimal portfolio in $V(N)=V(0)\prod_{k=0}^{N-1}(1+w^Tr(k)).$6 is studied through deflators, the Azéma supermartingale $V(N)=V(0)\prod_{k=0}^{N-1}(1+w^Tr(k)).$7, and $V(N)=V(0)\prod_{k=0}^{N-1}(1+w^Tr(k)).$8-observable characteristics of $V(N)=V(0)\prod_{k=0}^{N-1}(1+w^Tr(k)).$9. The setting covers credit risk and life insurance, where $$g(w)=\frac1N\mathbb E[\ln V(N)/V(0)]=\mathbb E[\ln(1+w^Tr)].$$0 represents default time or death time (Choulli et al., 2018).

A complementary enlargement problem concerns the model after a random time, $$g(w)=\frac1N\mathbb E[\ln V(N)/V(0)]=\mathbb E[\ln(1+w^Tr)].$$1. There, existence of the log-optimal portfolio is reduced to a dual problem over deflators, and the utility increment relative to the initial filtration decomposes into “cost of late investing,” “correlation risk,” and “information premium” (Alharbi et al., 2022).

A recent rough-paths treatment develops a fully deterministic framework for the continuous-time log-optimal portfolio in an Itô diffusion model. For a fixed price trajectory generated by a rough differential equation, one defines

$$g(w)=\frac1N\mathbb E[\ln V(N)/V(0)]=\mathbb E[\ln(1+w^Tr)].$$2

constructs the pathwise gain

$$g(w)=\frac1N\mathbb E[\ln V(N)/V(0)]=\mathbb E[\ln(1+w^Tr)].$$3

and recovers the portfolio and capital process pathwise. The paper establishes pathwise stability estimates with respect to model parameters and pathwise error estimates for time discretization (Allan et al., 24 Jul 2025).

4. Robust, cost-sensitive, and risk-controlled extensions

The classical ELG criterion assumes that the true distribution of returns is known. A major line of work relaxes this by replacing the expected log-growth under a single law with a worst-case objective over an ambiguity set. In the polyhedral setting, if $$g(w)=\frac1N\mathbb E[\ln V(N)/V(0)]=\mathbb E[\ln(1+w^Tr)].$$4 are return scenarios and

$$g(w)=\frac1N\mathbb E[\ln V(N)/V(0)]=\mathbb E[\ln(1+w^Tr)].$$5

the robust ELG is

$$g(w)=\frac1N\mathbb E[\ln V(N)/V(0)]=\mathbb E[\ln(1+w^Tr)].$$6

and the distributionally robust Kelly problem is

$$g(w)=\frac1N\mathbb E[\ln V(N)/V(0)]=\mathbb E[\ln(1+w^Tr)].$$7

Because $$g(w)=\frac1N\mathbb E[\ln V(N)/V(0)]=\mathbb E[\ln(1+w^Tr)].$$8 is concave, it can be under-approximated by supporting hyperplanes. The resulting supporting hyperplane approximation reformulates a class of robust log-optimal problems as a linear program, with auxiliary variables enforcing

$$g(w)=\frac1N\mathbb E[\ln V(N)/V(0)]=\mathbb E[\ln(1+w^Tr)].$$9

The framework allows transaction costs, leverage and shorting, survival trades, and diversification considerations, and includes an algorithm for selecting the number of hyperplanes required for a uniform approximation error $\Delta=\{w\in\mathbb R^n:\mathbf 1^Tw=1,\;w\ge 0\}$0 (Hsieh, 2022).

A Wasserstein-based extension incorporates convex transaction costs directly into the robust objective:

$\Delta=\{w\in\mathbb R^n:\mathbf 1^Tw=1,\;w\ge 0\}$1

where $\Delta=\{w\in\mathbb R^n:\mathbf 1^Tw=1,\;w\ge 0\}$2 is a Wasserstein ball around an empirical distribution. The paper establishes robust survivability conditions under convex transaction costs, derives a finite convex reformulation by duality, and reports that without transaction costs the optimal portfolio converges to an equal-weighted allocation, while with transaction costs the portfolio shifts slightly towards the risk-free asset (Hsieh et al., 2024).

Transaction costs also alter the frequency-dependent problem materially. If rebalancing occurs every $\Delta=\{w\in\mathbb R^n:\mathbf 1^Tw=1,\;w\ge 0\}$3 periods and proportional costs are embedded in $\Delta=\{w\in\mathbb R^n:\mathbf 1^Tw=1,\;w\ge 0\}$4, the objective becomes

$\Delta=\{w\in\mathbb R^n:\mathbf 1^Tw=1,\;w\ge 0\}$5

This is equivalent to a concave program over the simplex. The same work proves a dominance theorem with costs, shows that transaction costs may cause a bankruptcy issue because $\Delta=\{w\in\mathbb R^n:\mathbf 1^Tw=1,\;w\ge 0\}$6 may become negative with positive probability, introduces the quadratic surrogate

$\Delta=\{w\in\mathbb R^n:\mathbf 1^Tw=1,\;w\ge 0\}$7

derives necessary and sufficient KKT-type optimality conditions, and proves a two-fund theorem stating that any convex combination of two optimal weights is still optimal (Hsieh et al., 2023).

Risk control may also be imposed directly on log-returns. In a continuous-time complete market, when risk is measured by weighted Value-at-Risk on the continuously compounded log-return $\Delta=\{w\in\mathbb R^n:\mathbf 1^Tw=1,\;w\ge 0\}$8, the efficient frontier is described as a concave curve that connects the minimum-risk portfolio with the growth optimal portfolio, as opposed to the vertical line when WVaR is applied to terminal wealth (Wei et al., 2021).

5. Data-driven, online, and learning-based implementations

A direct data-driven approach is the sliding-window estimator of the unknown return law. At each date $\Delta=\{w\in\mathbb R^n:\mathbf 1^Tw=1,\;w\ge 0\}$9, using the most recent $\max_{w\in\Delta}\;\mathbb E[\ln(1+w^Tr)].$0 realized returns $\max_{w\in\Delta}\;\mathbb E[\ln(1+w^Tr)].$1, one solves

$\max_{w\in\Delta}\;\mathbb E[\ln(1+w^Tr)].$2

then applies $\max_{w\in\Delta}\;\mathbb E[\ln(1+w^Tr)].$3 for the next period and repeats. The objective is concave and differentiable on $\max_{w\in\Delta}\;\mathbb E[\ln(1+w^Tr)].$4, so the problem has a unique global maximizer. The paper notes that no asymptotic convergence or regret bounds are proved and describes the approach as conceptually akin to model-predictive control. Computationally, each step solves an $\max_{w\in\Delta}\;\mathbb E[\ln(1+w^Tr)].$5-dimensional concave maximization with $\max_{w\in\Delta}\;\mathbb E[\ln(1+w^Tr)].$6 samples, with off-the-shelf methods running in roughly $\max_{w\in\Delta}\;\mathbb E[\ln(1+w^Tr)].$7 per step (Wang et al., 2022).

A different approximation, the Robust Log-Optimal Strategy (RLOS), replaces the exact GLOS objective by a second-order Taylor expansion around the mean return. Writing $\max_{w\in\Delta}\;\mathbb E[\ln(1+w^Tr)].$8 and $\max_{w\in\Delta}\;\mathbb E[\ln(1+w^Tr)].$9, the surrogate allocation utility is

$r_i>-1$0

Because this surrogate depends only on the first two moments, it avoids high-dimensional CDF estimation and is described as resisting the “Butterfly Effect” caused by estimation error. The paper then embeds the moment-based signal into a reinforcement-learning architecture, RLOSRL, where the state is a $r_i>-1$1 tensor of recent open/high/low/volume data, the signal is the RLOS-suggested weight vector, the action is produced by a 1D-CNN with shared filters and softmax output, and the reward is realized log-return (Guo et al., 2018).

Online portfolio-selection theory formulates the problem as logarithmic loss minimization on the simplex. With round-$r_i>-1$2 loss

$r_i>-1$3

regret against the best constant rebalanced portfolio is

$r_i>-1$4

Cover’s Universal Portfolio achieves $r_i>-1$5 regret but requires high-dimensional integration. The VB-FTRL algorithm instead minimizes current logarithmic loss regularized by a hybrid logarithmic-volumetric barrier and achieves

$r_i>-1$6

with amortized per-round runtime $r_i>-1$7 (Jézéquel et al., 2022). The Ada algorithm, based on Online Mirror Descent with a “Newton + barrier” regularizer and an adaptive outer loop on $r_i>-1$8, achieves $r_i>-1$9 regret with reported per-round complexity $1+w^Tr>0$0 (Luo et al., 2018).

These implementations show that the log-optimal portfolio has become a computational object as much as a theoretical one: it can be approximated by empirical likelihood, by moment surrogates, by online convex optimization, or by RL-enhanced predictors, depending on what information and latency regime are assumed.

6. Special structures, empirical evidence, and interpretive issues

Several papers clarify boundary cases and common interpretive errors. The dominant-asset theorem establishes that the log-optimal solution is “all-in” only under a specific relative-attractiveness condition. The same work then shows that if a dominant asset exists and a buy-and-hold portfolio assigns it positive weight, the ELG gap satisfies

$1+w^Tr>0$1

so $1+w^Tr>0$2 at the sublinear rate $1+w^Tr>0$3 (Hsieh, 2021). This suggests that asymptotic log-optimality can coexist with ad hoc initial diversification in the presence of a dominant asset.

Another structural connection appears under approximate log-normality of portfolio gross returns. In that setting, closed-form expressions for logarithmic-utility weights are obtained, the resulting portfolio belongs to the set of mean-variance feasible portfolios, and necessary and sufficient conditions are given for mean-variance efficiency (Bodnar et al., 2018).

Empirical studies in the supplied literature are illustrative rather than uniform, but they reveal recurring patterns.

Study	Setting	Reported finding
(Hsieh, 2022)	Top 15 S&P 500 stocks; in-sample Jan 2–Jun 30 2021; out-of-sample Jul 1–Dec 31 2021	In-sample growth rates nearly identical: robust SHA-LP $1+w^Tr>0$4 vs nominal Kelly $1+w^Tr>0$5; out-of-sample final wealth SHA $1+w^Tr>0$6 vs nominal $1+w^Tr>0$7; drawdown SHA $1+w^Tr>0$8 vs nominal $1+w^Tr>0$9
(Wang et al., 2022)	VT, BND, BNDX daily closes Feb 14 2018–Feb 14 2020	In the “Bull” example, sliding-window with $w\in\mathbb R^n$00 gives cumulative return $w\in\mathbb R^n$01 versus $w\in\mathbb R^n$02 for classical ELG; in the “Bear” example, classical cumulative return is $w\in\mathbb R^n$03 while $w\in\mathbb R^n$04 gives $w\in\mathbb R^n$05
(Guo et al., 2018)	CSI300 back tests	RLOS steadily outperforms Naïve-Average, Follow-the-Winner, and Follow-the-Loser; the dashed red RLOSRL equity line lies strictly above RLOS in all tests

A recurring misconception is that log-optimality is only meaningful when the true law of returns is known in advance. The robust formulation over ambiguity sets, the sliding-window empirical formulation, the Blackwell-approachability construction, and universal/online portfolio algorithms are all explicitly designed for settings in which the law is unknown, only partially known, or not modeled probabilistically at all (Hsieh, 2022, Wang et al., 2022, V'yugin, 2014, Cuchiero et al., 2016). Another misconception is that the criterion necessarily prescribes unrestricted leverage or a perpetual corner solution. In the supplied literature, leverage bounds, shorting decompositions, box constraints, no-ruin constraints, convex transaction costs, WVaR constraints, and filtration-enlargement effects are incorporated directly into the optimization problem (Hsieh, 2022, Hsieh et al., 2023, Hsieh et al., 2024, Wei et al., 2021).

Taken together, these results depict the log-optimal portfolio not as a single formula but as a family of growth-maximizing constructions. The family includes classical ELG maximization, semimartingale numéraire portfolios, model-free calibrated strategies, distributionally robust and cost-sensitive variants, and computational approximations ranging from linear programming to online convex optimization and reinforcement learning.