Mean-Variance Portfolio Selection

• Mean-variance portfolio selection is a framework for choosing portfolio weights that balance expected return and risk, tracing the efficient frontier.
• The approach extends from classical static optimization to dynamic continuous-time and multi-period models, addressing challenges like time inconsistency and state-dependent risk aversion.
• Advanced methods incorporate robust optimization, Bayesian filtering, and reinforcement learning to tackle estimation errors and market uncertainties.

Searching arXiv for recent and foundational papers on mean-variance portfolio selection to ground the article in current literature. Mean-variance portfolio selection is the class of portfolio choice problems in which allocation decisions are organized around the trade-off between expected return or terminal wealth and variance as a risk proxy. In the classical single-period formulation, portfolio weights $x$ over risky assets with mean vector $\mu$ and covariance matrix $\Sigma$ are chosen so that portfolio mean is $\mu^\top x$ and portfolio variance is $x^\top \Sigma x$; varying a target return parameter traces the mean-variance efficient frontier (Cesarone et al., 2021). In multi-period and continuous-time settings, the same trade-off is imposed on terminal wealth generated by dynamic trading, but the variance term makes the problem nonseparable and typically time-inconsistent, which has led to pre-committed formulations, equilibrium formulations, mean-field reformulations, robust variants, and data-driven methods that learn policies directly from observations rather than from a fully specified market model (Cui et al., 2013, Huang et al., 2024).

1. Classical formulation and efficient frontier

The classical Markowitz problem in the long-only static setting takes $n$ risky assets with random returns $R_1,\dots,R_n$, portfolio weights $$x\in\Delta:=\{x\in\mathbb{R}^n:\sum_{k=1}^n x_k=1,\ x_k\ge 0\}$$, expected return vector $\mu$, and covariance matrix $\Sigma$. Portfolio return is $\mu$0, expected return is $\mu$1, and variance is $\mu$2. The standard target-return problem minimizes $\mu$3 subject to $\mu$4, $\mu$5, and $\mu$6; varying $\mu$7 traces the mean-variance efficient frontier (Cesarone et al., 2021).

Equivalent scalarized forms maximize expected return minus a variance penalty. In the notation of one modern treatment, the mean-variance utility under distribution $\mu$8 is

$\mu$9

over no-short portfolios $\Sigma$0 (Lam, 2024). This representation makes explicit that mean-variance selection can be viewed either as constrained variance minimization or as direct utility maximization.

Discrete-time multi-period extensions preserve the same conceptual structure but apply it to terminal wealth generated by sequential trading. In one standard model, wealth follows

$\Sigma$1

where $\Sigma$2 is the gross return of the riskless asset, $\Sigma$3 is the vector of risky holdings, and $\Sigma$4 is the excess-return vector of risky assets (Cui et al., 2013). This is the setting in which dynamic mean-variance problems begin to differ sharply from the one-period case, because the objective depends on the joint law of future wealth rather than on a single portfolio return.

A common misconception is that the efficient frontier is exclusively a static geometric object in the $\Sigma$5-plane. The dynamic literature shows that the frontier persists in more elaborate forms: as a discrete-time frontier for multi-period wealth, as a continuous-time frontier for terminal wealth under stochastic dynamics, and in some extensions as a higher-dimensional efficient surface when additional risk criteria such as Value-at-Risk are added (Cesarone et al., 2021).

2. Dynamic continuous-time formulations and time inconsistency

Continuous-time mean-variance portfolio selection is typically formulated in a diffusion market with one risk-free asset and one or more risky assets. In a Black-Scholes setting with one risky asset, one representative wealth equation is

$\Sigma$6

where $\Sigma$7 is income, $\Sigma$8 consumption, and $\Sigma$9 the portfolio proportion in the stock (Yang et al., 2020). In a multi-asset diffusion setting with monetary risky positions $\mu^\top x$0, another standard wealth equation is

$\mu^\top x$1

with excess return $\mu^\top x$2 and covariance $\mu^\top x$3 (Shen et al., 2022).

The defining technical difficulty is time inconsistency. Because the variance term is nonlinear in expectations, Bellman’s principle fails: a strategy optimized at time $\mu^\top x$4 is generally not re-optimized by the same criterion at later times. The literature therefore distinguishes at least two solution concepts. The first is the pre-committed solution, obtained by solving the initial problem over the whole horizon and then following the resulting policy even though later re-optimization may disagree with it (Cui et al., 2013, Li et al., 2024). The second is the equilibrium or game-theoretic solution, in which no infinitesimal deviation on a future interval can improve the objective, a continuous-time analogue of subgame-perfect Nash equilibrium (Yang et al., 2020).

Several mathematical devices have been developed to handle the nonseparability. Zhou and Li’s embedding technique converts the mean-variance objective into a family of quadratic control problems, after which dynamic programming applies (Li et al., 2024). A distinct route is the mean-field reformulation of discrete-time multi-period problems, where the state is augmented by its expectation and fluctuation $\mu^\top x$5, converting the original nonseparable problem into a separable mean-field linear-quadratic control problem (Cui et al., 2013). In equilibrium formulations, extended HJB systems augment the state by conditional moments such as the conditional mean of terminal wealth, its second moment, and the accumulated utility of consumption (Yang et al., 2020).

These formulations also clarify a second misconception: “dynamic mean-variance” is not a single model. It includes pre-committed terminal-wealth problems, intertemporal mean-variance problems, Lagrangian relaxations of bankruptcy constraints, and equilibrium control problems, all sharing the same mean-versus-variance trade-off but differing in admissible controls, state variables, and solution concept.

3. State-dependent risk aversion, human capital, and consumption

A prominent extension introduces time- and state-dependent risk aversion into the mean-variance framework by allowing the penalty on terminal-wealth variance to depend on current wealth and on the present value of future net income. In that model, the key auxiliary quantity is

$\mu^\top x$6

and the risk-aversion coefficient is specified as

$\mu^\top x$7

The state variable $\mu^\top x$8 is interpreted as the total financial position, combining current financial wealth with the discounted value of future net income (Yang et al., 2020).

This structure implies that risk aversion decreases in overall financial wealth and varies over time through the evolution of $\mu^\top x$9. When future income net of consumption is large, $x^\top \Sigma x$0 falls, leading to more risk taking; as time passes and less labor income remains, $x^\top \Sigma x$1 shrinks, which tends to increase effective risk aversion if financial wealth does not rise sufficiently (Yang et al., 2020).

Under an analytically tractable ansatz in which equilibrium consumption is independent of wealth, the model yields explicit equilibrium rules. The risky dollar investment takes the form

$x^\top \Sigma x$2

so the risky position scales with total financial wealth rather than current wealth alone. Consumption is deterministic,

$x^\top \Sigma x$3

and the coefficient functions $x^\top \Sigma x$4 and $x^\top \Sigma x$5 solve coupled ODEs with terminal conditions $x^\top \Sigma x$6 and $x^\top \Sigma x$7 (Yang et al., 2020).

This formulation resolves a limitation explicitly noted relative to an earlier constant-risk-aversion mean-variance-utility model: in the new specification, both wealth and future income-consumption balance matter for risky investment. A plausible implication is that the model is closer to life-cycle investing intuition, because human capital enters directly into the risk-taking rule. The same paper also emphasizes that the consumption path is continuous over time and preference-driven, rather than exhibiting extreme switching behavior (Yang et al., 2020).

4. Constraints, monotonicity, and alternative risk criteria

A large branch of the literature studies mean-variance selection under market and portfolio constraints. One continuous-time formulation imposes both bankruptcy prohibition $x^\top \Sigma x$8 and convex cone portfolio constraints $x^\top \Sigma x$9. A central result is that the constrained problem can be transformed into an equivalent mean-variance problem with bankruptcy prohibition only, but in a modified market whose effective excess return $n$0 is obtained by projecting the original market price of risk onto the cone of feasible trading directions (Li et al., 2015). The optimal terminal wealth then takes the truncated form

$n$1

and the optimal portfolio remains a semi-analytical feedback rule in wealth (Li et al., 2015).

Another strand concerns monotone mean-variance preferences. Classical mean-variance is not monotone: one can have $n$2 almost surely yet prefer $n$3 because variance penalizes favorable upside deviations as well as downside risk. Under diffusion models with conic convex constraints, however, the precommitted optimal strategies for monotone mean-variance and classical mean-variance coincide. The key structural reason is an orthogonality property for the projection $n$4 of $n$5 onto the cone image $n$6,

$n$7

which makes the MMV and MV thresholds identical (Shen et al., 2022). The paper also notes that this equivalence need not survive for general convex non-conic constraints.

Mean-variance can also be augmented rather than constrained. One example is the tri-objective Mean-Variance-VaR model, which minimizes variance subject to a target mean and a Value-at-Risk bound: $n$8 Because the VaR constraint depends on scenario order statistics, the exact scenario-based formulation is a Mixed-Integer Quadratic Programming problem with binary variables $n$9 and a big-$R_1,\dots,R_n$0 device (Cesarone et al., 2021). The same study reports that the out-of-sample performance of Mean-Variance-VaR portfolios seems to be generally better than that of optimal Mean-Variance and Mean-VaR portfolios, while also documenting the resulting computational hardness and the debate with CVaR: CVaR is coherent and often easier to optimize, but VaR is retained there for regulatory realism (Cesarone et al., 2021).

A further augmentation introduces a running tracking-error penalty toward a reference portfolio $R_1,\dots,R_n$1. In continuous time, the cost becomes

$R_1,\dots,R_n$2

and the problem is formulated as a McKean-Vlasov control problem with explicit Riccati ODEs and an optimal feedback control that interpolates between classical mean-variance and pure benchmark tracking (Lefebvre et al., 2020). This suggests a regularization viewpoint: benchmark tracking can be used both to control estimation error and to align the portfolio with institutional benchmarks.

5. Estimation, filtering, and partial information

Mean-variance strategies are highly sensitive to the estimation of expected returns and volatility, so a substantial literature addresses partial information and online parameter updating. In one continuous-time model with a hidden constant drift $R_1,\dots,R_n$3, the investor observes only stock prices. The posterior probability $R_1,\dots,R_n$4 follows

$R_1,\dots,R_n$5

where $R_1,\dots,R_n$6 is the innovation Brownian motion, and the observable-filtration wealth dynamics become

$R_1,\dots,R_n$7

The market is shown to be complete under the observable filtration, and the optimal strategy is characterized through a BSDE as

$R_1,\dots,R_n$8

with $R_1,\dots,R_n$9 represented by a Clark-Ocone formula and approximated numerically via Malliavin calculus and particle methods (Xiong et al., 2019).

In discrete time, Bayesian filtering has been combined with multi-period mean-variance optimization by modeling asset returns with AR or VAR dynamics, updating unknown parameters sequentially through dynamic linear models, and feeding the resulting predictive moments into the dynamic-programming solution of the pre-committed mean-variance problem with uncertain exit time. For a single risky asset, the optimal control under random horizon $$x\in\Delta:=\{x\in\mathbb{R}^n:\sum_{k=1}^n x_k=1,\ x_k\ge 0\}$$0 is

$$x\in\Delta:=\{x\in\mathbb{R}^n:\sum_{k=1}^n x_k=1,\ x_k\ge 0\}$$1

and a corresponding closed-form efficient frontier is obtained for terminal wealth under the random exit-time distribution $$x\in\Delta:=\{x\in\mathbb{R}^n:\sum_{k=1}^n x_k=1,\ x_k\ge 0\}$$2 (Sikaria et al., 2019). The empirical study on S&P 500 data reports that Bayesian updating is strongly favored by the data and that the method is practically implementable (Sikaria et al., 2019).

A different estimation program reformulates the classical pre-commitment solution in terms of current profitability $$x\in\Delta:=\{x\in\mathbb{R}^n:\sum_{k=1}^n x_k=1,\ x_k\ge 0\}$$3 and average profitability

$$x\in\Delta:=\{x\in\mathbb{R}^n:\sum_{k=1}^n x_k=1,\ x_k\ge 0\}$$4

Using an auxiliary wealth process $$x\in\Delta:=\{x\in\mathbb{R}^n:\sum_{k=1}^n x_k=1,\ x_k\ge 0\}$$5 with strategy $$x\in\Delta:=\{x\in\mathbb{R}^n:\sum_{k=1}^n x_k=1,\ x_k\ge 0\}$$6, the paper proves

$$x\in\Delta:=\{x\in\mathbb{R}^n:\sum_{k=1}^n x_k=1,\ x_k\ge 0\}$$7

so AP can be estimated from quadratic variation rather than from direct drift estimation (Li et al., 2024). The same work argues, both theoretically and numerically, that these AP/CP-based estimators are more accurate than MLE-based squared-Sharpe estimates under its stated conditions (Li et al., 2024).

6. Robust, online, and reinforcement-learning approaches

A robust line of work replaces the unknown return distribution by an ambiguity set around the empirical measure. In the Wasserstein formulation, the distributionally robust target-return problem over portfolio weights $$x\in\Delta:=\{x\in\mathbb{R}^n:\sum_{k=1}^n x_k=1,\ x_k\ge 0\}$$8 is equivalent to

$$x\in\Delta:=\{x\in\mathbb{R}^n:\sum_{k=1}^n x_k=1,\ x_k\ge 0\}$$9

where $\mu$0 is the robust feasible set and $\mu$1 is the Wasserstein radius (Blanchet et al., 2018). This identifies distributional robustness with an explicit norm regularization of the empirical variance problem, and the same paper develops a data-driven methodology for choosing both $\mu$2 and the associated robust target return $\mu$3 (Blanchet et al., 2018).

A nonparametric online-learning perspective abandons the train-test separation entirely. With return vectors $\mu$4 arriving sequentially, the proposed strategy recomputes at each time the Markowitz portfolio for the empirical distribution $\mu$5. Under weak convergence $\mu$6 and positive-definite limiting covariance, the resulting dynamic mean-variance strategy asymptotically matches the true constant mean-variance portfolio derived with perfect knowledge of $\mu$7 in empirical utility, Sharpe ratio, and growth rate (Lam, 2024). A second algorithm updates the risk-aversion coefficient over time by selecting, from a candidate set $\mu$8, the sub-strategy with the best empirical Sharpe ratio or growth rate, thereby addressing ambiguity in the appropriate level of risk aversion (Lam, 2024).

At the continuous-time end, reinforcement learning has been used to learn the pre-committed strategy directly in diffusion markets with unknown coefficients. The RL formulation replaces deterministic controls by stochastic policies $\mu$9, adds an entropy penalty, and exploits martingale characterizations of the value process and policy-gradient identities to update actor and critic parameters without estimating market coefficients (Huang et al., 2024). In multi-stock Black-Scholes markets without factors, the paper develops a baseline algorithm with a sublinear regret bound in terms of Sharpe ratio, and its empirical study on S&P 500 constituents reports that the continuous-time RL strategies are consistently among the best especially in a volatile bear market and decisively outperform the model-based continuous-time counterparts by significant margins (Huang et al., 2024).

These developments reflect a broader shift in the field. A plausible implication is that mean-variance selection is no longer only a closed-form model-based problem; it is also a sequential decision problem under model misspecification, ambiguity, and nonstationarity, with solutions ranging from robust optimization to online learning and direct policy learning.

7. Nonlinear wealth dynamics and rough-volatility generalizations

Mean-variance selection has also been extended beyond linear diffusion wealth equations. One line studies nonlinear wealth dynamics of the form

$\Sigma$0

where long and short positions face different mean excess returns, allowing interpretations through long/short asymmetry, taxes, or price impact (Ji et al., 2017). The corresponding mean-variance problem is solved by first tackling an auxiliary quadratic-loss problem via two generalized stochastic Riccati equations, then verifying optimality through convex duality. The optimal wealth process never crosses the vertex of the parabola associated with the quadratic loss, the variance-optimal martingale measure is characterized explicitly, and the efficient frontier is obtained in closed form (Ji et al., 2017). The same paper concludes that people are more likely to invest their money in the riskless asset compared with the classical linear market (Ji et al., 2017).

Another line replaces Markovian stochastic volatility with the Volterra Heston model, where variance satisfies a stochastic Volterra integral equation and may be rough in the fractional-kernel case. The mean-variance problem is formulated as minimizing $\Sigma$1 subject to $\Sigma$2, and the optimal strategy is derived through an auxiliary process $\Sigma$3 built from forward variance and a Riccati-Volterra equation

$\Sigma$4

The optimal feedback control has the form

$\Sigma$5

and the efficient frontier remains a quadratic curve, even though volatility is non-Markovian and typically non-semimartingale (Han et al., 2019). Numerical analysis in that setting indicates that both roughness and volatility of volatility materially affect the optimal strategy (Han et al., 2019).

These generalizations make clear that the mean-variance principle is structurally resilient. Even when the market becomes non-Markovian, state-dependent, or nonlinear, the core object remains the same trade-off between expected terminal wealth and its variance; what changes are the state representation, the control concept, and the analytic machinery required to recover optimal or equilibrium policies.