Merton Portfolio Choice Problem

• Merton portfolio choice problem is a continuous-time stochastic control framework that defines optimal consumption and investment using feedback rules under risk dynamics.
• The framework underpins modern extensions including learning, ambiguity, rough volatility, and robust control, which modify classical HJB and duality methods.
• Research applications span strategic interaction, optimal stopping, and neural policy optimization, driving scalable solutions for high-dimensional portfolio choices.

The Merton portfolio choice problem is the continuous-time stochastic control problem of selecting consumption and portfolio policies so as to maximize expected utility of consumption, terminal wealth, or both, under stochastic asset-price dynamics. In its canonical form, wealth evolves under a risk-free asset and one or more risky assets, and the investor chooses a consumption rate and risky-asset exposure subject to wealth nonnegativity. The problem is foundational because it yields explicit feedback rules in benchmark cases, yet it is also a template for modern extensions involving learning, ambiguity, illiquidity, stopping, strategic interaction, rough volatility, jumps, and neural policy optimization (Moehle et al., 2021, Huh et al., 22 Jan 2025).

1. Canonical formulation and classical policy rules

A standard finite-horizon formulation with $n$ risky assets specifies wealth dynamics

$$dX_t = [r X_t + X_t\boldsymbol{\pi}_t^\top (\boldsymbol{\mu} - r\mathbf{1}) - C_t ] dt + X_t\boldsymbol{\pi}_t^\top \mathbf{V} d\mathbf{W}_t,\qquad X_0 = x_0 > 0,$$

and objective

$$J(\boldsymbol{\pi}_t, C_t) = \mathbb{E}\left[ \int_{0}^{T} e^{-\rho t} U(C_t)\,dt + \kappa e^{-\rho T} U(X_T) \right],$$

with CRRA utility

$U(x) = \frac{x^{1-\gamma}}{1-\gamma}.$

A closely related single-asset formulation writes

$$dX_t = X_t[\pi_t (\mu - r) + r]dt + X_t \pi_t \sigma dW_t - C_t dt,$$

with infinite-horizon objective

$$V(x) = \sup_{(\pi, C)} \mathbb{E} \left[ \int_0^\infty e^{-\delta t} \frac{C_t^{1-R}}{1-R} dt \right].$$

These formulations differ in notation and horizon, but share the same core structure: continuous-time control of risky exposure and consumption under multiplicative wealth dynamics (Huh et al., 22 Jan 2025, Herdegen et al., 2020).

In the unconstrained constant-parameter case, the classical infinite-horizon solution is explicit. For the multi-asset model,

$$\boldsymbol{\pi}_t^* = \frac{1}{\gamma}\Sigma^{-1}(\boldsymbol{\mu} - r\mathbf{1}), \qquad C_t^* = \nu X_t,$$

with

$$\nu = \frac{\rho - (1-\gamma)\left[r + \frac{1}{2}(\boldsymbol{\mu} - r\mathbf{1})^\top \Sigma^{-1} (\boldsymbol{\mu} - r\mathbf{1})\right]}{\gamma}.$$

In the single-asset Black-Scholes-Merton setting, the optimal controls can be written

$$\hat{\pi} = \frac{\lambda}{\sigma R},\qquad \hat{\xi} = \eta,\qquad C_t^*=\hat{\xi}X_t,$$

where

$$\eta := \frac{1}{\delta - (1-R)\left( r + \frac{\lambda^2}{2R} \right)}.$$

These formulas encode the defining Merton property: optimal risky exposure is proportional to the instantaneous risk premium and inversely proportional to risk aversion, while optimal consumption is proportional to wealth (Huh et al., 22 Jan 2025, Herdegen et al., 2020).

2. Homogeneity, HJB structure, and alternative solution theories

The classical solution is usually derived from dynamic programming. In an infinite-horizon CRRA setting, homogeneity implies

$$dX_t = [r X_t + X_t\boldsymbol{\pi}_t^\top (\boldsymbol{\mu} - r\mathbf{1}) - C_t ] dt + X_t\boldsymbol{\pi}_t^\top \mathbf{V} d\mathbf{W}_t,\qquad X_0 = x_0 > 0,$$0

suggesting the power ansatz

$$dX_t = [r X_t + X_t\boldsymbol{\pi}_t^\top (\boldsymbol{\mu} - r\mathbf{1}) - C_t ] dt + X_t\boldsymbol{\pi}_t^\top \mathbf{V} d\mathbf{W}_t,\qquad X_0 = x_0 > 0,$$1

The associated stationary HJB equation becomes

$$dX_t = [r X_t + X_t\boldsymbol{\pi}_t^\top (\boldsymbol{\mu} - r\mathbf{1}) - C_t ] dt + X_t\boldsymbol{\pi}_t^\top \mathbf{V} d\mathbf{W}_t,\qquad X_0 = x_0 > 0,$$2

and direct maximization yields

$$dX_t = [r X_t + X_t\boldsymbol{\pi}_t^\top (\boldsymbol{\mu} - r\mathbf{1}) - C_t ] dt + X_t\boldsymbol{\pi}_t^\top \mathbf{V} d\mathbf{W}_t,\qquad X_0 = x_0 > 0,$$3

For high risk aversion $$dX_t = [r X_t + X_t\boldsymbol{\pi}_t^\top (\boldsymbol{\mu} - r\mathbf{1}) - C_t ] dt + X_t\boldsymbol{\pi}_t^\top \mathbf{V} d\mathbf{W}_t,\qquad X_0 = x_0 > 0,$$4, recent work emphasizes that well-posedness is delicate and is characterized by the finiteness condition

$$dX_t = [r X_t + X_t\boldsymbol{\pi}_t^\top (\boldsymbol{\mu} - r\mathbf{1}) - C_t ] dt + X_t\boldsymbol{\pi}_t^\top \mathbf{V} d\mathbf{W}_t,\qquad X_0 = x_0 > 0,$$5

That analysis derives the explicit solution without requiring transversality conditions, using homogeneity and a “half-verification” argument (Biffis et al., 4 Aug 2025).

The classical HJB route is no longer the only solution theory. A variational-analysis approach formulates expected utility as a functional $$dX_t = [r X_t + X_t\boldsymbol{\pi}_t^\top (\boldsymbol{\mu} - r\mathbf{1}) - C_t ] dt + X_t\boldsymbol{\pi}_t^\top \mathbf{V} d\mathbf{W}_t,\qquad X_0 = x_0 > 0,$$6 on the admissible controls and characterizes the optimal portfolio through the vanishing of the Gâteaux derivative. In a general continuous semimartingale market with stochastic coefficients, the optimal portfolio satisfies the feedback form

$$dX_t = [r X_t + X_t\boldsymbol{\pi}_t^\top (\boldsymbol{\mu} - r\mathbf{1}) - C_t ] dt + X_t\boldsymbol{\pi}_t^\top \mathbf{V} d\mathbf{W}_t,\qquad X_0 = x_0 > 0,$$7

where

$$dX_t = [r X_t + X_t\boldsymbol{\pi}_t^\top (\boldsymbol{\mu} - r\mathbf{1}) - C_t ] dt + X_t\boldsymbol{\pi}_t^\top \mathbf{V} d\mathbf{W}_t,\qquad X_0 = x_0 > 0,$$8

and $$dX_t = [r X_t + X_t\boldsymbol{\pi}_t^\top (\boldsymbol{\mu} - r\mathbf{1}) - C_t ] dt + X_t\boldsymbol{\pi}_t^\top \mathbf{V} d\mathbf{W}_t,\qquad X_0 = x_0 > 0,$$9 is part of a coupled FBSDE. In the deterministic-coefficient log and CRRA cases, this recovers the classical Merton rules (Al-Aradi et al., 2020).

A distinct primal method replaces the utility $$J(\boldsymbol{\pi}_t, C_t) = \mathbb{E}\left[ \int_{0}^{T} e^{-\rho t} U(C_t)\,dt + \kappa e^{-\rho T} U(X_T) \right],$$0 by the perturbed utility $$J(\boldsymbol{\pi}_t, C_t) = \mathbb{E}\left[ \int_{0}^{T} e^{-\rho t} U(C_t)\,dt + \kappa e^{-\rho T} U(X_T) \right],$$1, where $$J(\boldsymbol{\pi}_t, C_t) = \mathbb{E}\left[ \int_{0}^{T} e^{-\rho t} U(C_t)\,dt + \kappa e^{-\rho T} U(X_T) \right],$$2 is the candidate optimal consumption stream. This stochastic perturbation makes $$J(\boldsymbol{\pi}_t, C_t) = \mathbb{E}\left[ \int_{0}^{T} e^{-\rho t} U(C_t)\,dt + \kappa e^{-\rho T} U(X_T) \right],$$3 strictly positive and simplifies the supermartingale and transversality steps, giving

$$J(\boldsymbol{\pi}_t, C_t) = \mathbb{E}\left[ \int_{0}^{T} e^{-\rho t} U(C_t)\,dt + \kappa e^{-\rho T} U(X_T) \right],$$4

and then recovering the original value by letting $$J(\boldsymbol{\pi}_t, C_t) = \mathbb{E}\left[ \int_{0}^{T} e^{-\rho t} U(C_t)\,dt + \kappa e^{-\rho T} U(X_T) \right],$$5. This removes many of the parameter and admissibility restrictions that appear in earlier primal verifications (Herdegen et al., 2020).

Another reformulation is deterministic rather than stochastic. A certainty equivalent problem yields the exact same value function and optimal policy as the base Merton problem: $$J(\boldsymbol{\pi}_t, C_t) = \mathbb{E}\left[ \int_{0}^{T} e^{-\rho t} U(C_t)\,dt + \kappa e^{-\rho T} U(X_T) \right],$$6 After discretization, the certainty equivalent problem becomes a second-order cone program, which the paper proposes as a practical starting point for model predictive control (Moehle et al., 2021).

3. Learning, ambiguity, and robust variants

A major departure from the benchmark model is incomplete information about expected returns. Under drift uncertainty, the investor observes prices but not the true drift vector $$J(\boldsymbol{\pi}_t, C_t) = \mathbb{E}\left[ \int_{0}^{T} e^{-\rho t} U(C_t)\,dt + \kappa e^{-\rho T} U(X_T) \right],$$7, which is treated as an unknown static parameter with prior $$J(\boldsymbol{\pi}_t, C_t) = \mathbb{E}\left[ \int_{0}^{T} e^{-\rho t} U(C_t)\,dt + \kappa e^{-\rho T} U(X_T) \right],$$8. The Bayesian estimator

$$J(\boldsymbol{\pi}_t, C_t) = \mathbb{E}\left[ \int_{0}^{T} e^{-\rho t} U(C_t)\,dt + \kappa e^{-\rho T} U(X_T) \right],$$9

becomes an endogenous state variable, and wealth evolves as

$U(x) = \frac{x^{1-\gamma}}{1-\gamma}.$0

For Gaussian priors,

$U(x) = \frac{x^{1-\gamma}}{1-\gamma}.$1

and the CARA-optimal portfolio is

$U(x) = \frac{x^{1-\gamma}}{1-\gamma}.$2

This differs from the naive plug-in rule by a “learning-anticipation effect,” since the control anticipates future information updates (Bismuth et al., 2016).

Robust Merton formulations replace statistical learning by worst-case control. In the ellipsoidal ambiguity model, drift belongs to

$U(x) = \frac{x^{1-\gamma}}{1-\gamma}.$3

and the HJB becomes a max-min HJB-Isaacs equation. The inner minimization yields the penalty

$U(x) = \frac{x^{1-\gamma}}{1-\gamma}.$4

With CRRA utility, the optimal robust risky allocation is a shrunk Merton rule,

$U(x) = \frac{x^{1-\gamma}}{1-\gamma}.$5

where

$U(x) = \frac{x^{1-\gamma}}{1-\gamma}.$6

If $U(x) = \frac{x^{1-\gamma}}{1-\gamma}.$7, the investor holds only the safe asset. This directly formalizes the idea that ambiguity acts by reducing the effective Sharpe ratio (Biagini et al., 2015).

A more recent distributionally robust Bayesian formulation places a single Wasserstein ambiguity set on the drift prior rather than using time-rectangular adversarial control. The robust value is

$U(x) = \frac{x^{1-\gamma}}{1-\gamma}.$8

where $U(x) = \frac{x^{1-\gamma}}{1-\gamma}.$9 is the Karatzas-Zhao Bayesian Merton functional. A minimax swap reduces the problem to optimizing over priors, and for small radius the worst-case prior is an asymptotically optimal pushforward

$$dX_t = X_t[\pi_t (\mu - r) + r]dt + X_t \pi_t \sigma dW_t - C_t dt,$$0

The paper explicitly contrasts this prior-level ambiguity with time-rectangular DRC, arguing that the latter induces over-pessimism, whereas the former preserves learning and tractability (Blanchet et al., 1 Dec 2025).

4. Preference modifications and equilibrium reinterpretations

The classical model implies deterministic feedback rules conditional on observed states. That implication is relaxed in "Merton’s Problem with Recursive Perturbed Utility" (Dai et al., 14 Feb 2026), where the investor receives utility from randomization itself through a recursive entropy term: $$dX_t = X_t[\pi_t (\mu - r) + r]dt + X_t \pi_t \sigma dW_t - C_t dt,$$1 The optimal risky-exposure policy is Gaussian,

$$dX_t = X_t[\pi_t (\mu - r) + r]dt + X_t \pi_t \sigma dW_t - C_t dt,$$2

with

$$dX_t = X_t[\pi_t (\mu - r) + r]dt + X_t \pi_t \sigma dW_t - C_t dt,$$3

For $$dX_t = X_t[\pi_t (\mu - r) + r]dt + X_t \pi_t \sigma dW_t - C_t dt,$$4, the recursive formulation reduces to additive perturbed utility, and the mean portfolio coincides with the classical Merton rule. In complete markets, the hedging term vanishes (Dai et al., 14 Feb 2026).

Non-exponential discounting leads to time inconsistency rather than full commitment optimality. In a complete-market formulation with general utility and non-constant discounting, the relevant controls are subgame perfect strategies. The consumption-to-wealth and investment-to-wealth ratios are given in feedback form by

$$dX_t = X_t[\pi_t (\mu - r) + r]dt + X_t \pi_t \sigma dW_t - C_t dt,$$5

and the constant discount rate of the classical model is replaced by a utility weighted discount rate $$dX_t = X_t[\pi_t (\mu - r) + r]dt + X_t \pi_t \sigma dW_t - C_t dt,$$6. A fixed point iteration is used to compute $$dX_t = X_t[\pi_t (\mu - r) + r]dt + X_t \pi_t \sigma dW_t - C_t dt,$$7, and under the stated asymptotic assumptions the subgame perfect strategy is the same as the optimal strategy with that replacement (Mbodji, 20 Feb 2026).

Benchmark-relative objectives produce another systematic deformation of the Merton rule. If utility depends on both absolute and relative terminal wealth,

$$dX_t = X_t[\pi_t (\mu - r) + r]dt + X_t \pi_t \sigma dW_t - C_t dt,$$8

then the optimal risky weights become

$$dX_t = X_t[\pi_t (\mu - r) + r]dt + X_t \pi_t \sigma dW_t - C_t dt,$$9

The first term is the classical Merton allocation; the second subtracts benchmark exposures according to the relative-performance exponents. In the CAPM setting with a fixed beta constraint and an investable benchmark, the optimal risky allocation is only in the benchmark and the risk-free asset (Sarantsev, 2021).

5. Illiquidity, stopping, sharing, and strategic interaction

Several extensions transform Merton’s problem from pure stochastic control into mixed control-stopping or game-theoretic systems. In "Merton problem with one additional indivisible asset" (Trybuła, 2014), the investor owns one unit of an indivisible asset with value process

$$V(x) = \sup_{(\pi, C)} \mathbb{E} \left[ \int_0^\infty e^{-\delta t} \frac{C_t^{1-R}}{1-R} dt \right].$$0

independent of liquid-wealth noise. The asset can be sold only once, at a stopping time $$V(x) = \sup_{(\pi, C)} \mathbb{E} \left[ \int_0^\infty e^{-\delta t} \frac{C_t^{1-R}}{1-R} dt \right].$$1, after which the proceeds are folded into standard Merton wealth. By scaling, the value function takes the form

$$V(x) = \sup_{(\pi, C)} \mathbb{E} \left[ \int_0^\infty e^{-\delta t} \frac{C_t^{1-R}}{1-R} dt \right].$$2

and the optimal sale time is

$$V(x) = \sup_{(\pi, C)} \mathbb{E} \left[ \int_0^\infty e^{-\delta t} \frac{C_t^{1-R}}{1-R} dt \right].$$3

The problem is thus a free-boundary HJB-ODE system with smooth pasting.

Healthcare investment yields a related control-stopping structure with endogenous mortality. In "On a Merton Problem with Irreversible Healthcare Investment" (Ferrari et al., 2022), the state variables are wealth $$V(x) = \sup_{(\pi, C)} \mathbb{E} \left[ \int_0^\infty e^{-\delta t} \frac{C_t^{1-R}}{1-R} dt \right].$$4 and health capital $$V(x) = \sup_{(\pi, C)} \mathbb{E} \left[ \int_0^\infty e^{-\delta t} \frac{C_t^{1-R}}{1-R} dt \right].$$5, mortality is

$$V(x) = \sup_{(\pi, C)} \mathbb{E} \left[ \int_0^\infty e^{-\delta t} \frac{C_t^{1-R}}{1-R} dt \right].$$6

and the agent chooses consumption, portfolio, and an irreversible lump-sum healthcare investment time $$V(x) = \sup_{(\pi, C)} \mathbb{E} \left[ \int_0^\infty e^{-\delta t} \frac{C_t^{1-R}}{1-R} dt \right].$$7. Duality transforms the primal problem into a two-dimensional optimal stopping problem. The free boundary is Lipschitz continuous and characterized as the unique solution to a nonlinear integral equation, and in the original coordinates the health investment is undertaken when wealth exceeds an age- and health-dependent transformed boundary.

In a cooperative multi-agent variant, "Optimal Sharing Rule for a Household with a Portfolio Management Problem" (Huu et al., 2014) studies two investors with separate consumption streams and a shared terminal wealth. In a complete market, the value decomposes as

$$V(x) = \sup_{(\pi, C)} \mathbb{E} \left[ \int_0^\infty e^{-\delta t} \frac{C_t^{1-R}}{1-R} dt \right].$$8

so the household problem becomes the choice of an initial sharing rule followed by three independent Merton problems. In the reported numerical example with mean-reverting market price of risk, changing risk aversion can affect the consumption satisfaction proportion by up to 50% or more, while the effect of market price of risk is at most about 8%.

The price-taking assumption can also be dropped entirely. "Portfolio Choice In Dynamic Thin Markets: Merton Meets Cournot" (Gupta et al., 2023) models two large investors whose trading rates exert permanent linear price impact: $$V(x) = \sup_{(\pi, C)} \mathbb{E} \left[ \int_0^\infty e^{-\delta t} \frac{C_t^{1-R}}{1-R} dt \right].$$9 Each investor solves a best-response singular control problem inside a non-zero-sum stochastic differential game. Under constant volatility, the unique Nash equilibrium is deterministic and explicitly given by hyperbolic-function formulas for the equilibrium portfolios. A plausible implication is that the classical Merton framework is not confined to frictionless price taking; it can be embedded in strategically coupled equilibrium systems.

6. Rough volatility, jumps, and scalable computation

The Markovian Brownian setting is no longer standard in recent work. Under the Volterra Heston model,

$$\boldsymbol{\pi}_t^* = \frac{1}{\gamma}\Sigma^{-1}(\boldsymbol{\mu} - r\mathbf{1}), \qquad C_t^* = \nu X_t,$$0

the volatility process is non-Markovian and, for rough kernels, non-semimartingale. Classical HJB methods are unavailable, so the problem is solved by the martingale optimality principle using an auxiliary process and Riccati-Volterra equations. For power utility,

$$\boldsymbol{\pi}_t^* = \frac{1}{\gamma}\Sigma^{-1}(\boldsymbol{\mu} - r\mathbf{1}), \qquad C_t^* = \nu X_t,$$1

while for exponential utility,

$$\boldsymbol{\pi}_t^* = \frac{1}{\gamma}\Sigma^{-1}(\boldsymbol{\mu} - r\mathbf{1}), \qquad C_t^* = \nu X_t,$$2

The numerical experiments report a sign reversal with respect to roughness: as $$\boldsymbol{\pi}_t^* = \frac{1}{\gamma}\Sigma^{-1}(\boldsymbol{\mu} - r\mathbf{1}), \qquad C_t^* = \nu X_t,$$3 decreases, optimal investment decreases for power utility but increases for exponential utility (Han et al., 2019).

A multivariate jump extension replaces the auxiliary Riccati-Volterra system by a Riccati BSDE with jumps. In "Optimal Merton's Problem under Multivariate Affine Volterra Models with Jumps" (Dro et al., 1 May 2026), the market allows multiple risky assets, rough stochastic variance, and volatility jumps driven by an independent Poisson random measure. The optimal strategies remain semi-closed form. For logarithmic utility,

$$\boldsymbol{\pi}_t^* = \frac{1}{\gamma}\Sigma^{-1}(\boldsymbol{\mu} - r\mathbf{1}), \qquad C_t^* = \nu X_t,$$4

while exponential and power utilities involve $$\boldsymbol{\pi}_t^* = \frac{1}{\gamma}\Sigma^{-1}(\boldsymbol{\mu} - r\mathbf{1}), \qquad C_t^* = \nu X_t,$$5-dependent hedging corrections obtained from time-dependent multivariate Riccati-Volterra equations. Numerical experiments on a two-dimensional rough Heston model indicate that both path roughness and jumps materially affect value functions and optimal portfolios.

Jump structure can also be introduced through self-exciting events. In "Merton Investment Problems in Finance and Insurance for the Hawkes-based Models" (Swishchuk, 2021), the risky asset is driven by a general compound Hawkes process and then approximated by diffusion. In the finance application, the optimal log-utility risky fraction is

$$\boldsymbol{\pi}_t^* = \frac{1}{\gamma}\Sigma^{-1}(\boldsymbol{\mu} - r\mathbf{1}), \qquad C_t^* = \nu X_t,$$6

so the Merton rule depends explicitly on Hawkes intensity, excitation, and Markov-chain jump-size parameters.

Computational research increasingly treats the Merton problem as a benchmark for high-dimensional continuous-time learning algorithms. "Pontryagin-Guided Policy Optimization for Merton's Portfolio Problem" (Huh et al., 2024) tracks a policy-fixed adjoint BSDE and enforces Pontryagin first-order conditions during policy optimization; the alignment penalty

$$\boldsymbol{\pi}_t^* = \frac{1}{\gamma}\Sigma^{-1}(\boldsymbol{\mu} - r\mathbf{1}), \qquad C_t^* = \nu X_t,$$7

improves convergence speed and stability. "Pontryagin-Guided Deep Learning for Large-Scale Constrained Dynamic Portfolio Choice" (Huh et al., 22 Jan 2025) extends the same idea to constrained multi-asset problems, reporting tractability for up to 1,000 assets in the abstract and far exceeding the longstanding DP-based limit of around seven assets. These developments do not replace the analytical Merton problem; rather, they use its PMP structure as a supervisory signal for scalable policy learning.

A common misconception is that the Merton problem is exhausted by its closed-form classical solution. The body of work surveyed here shows the opposite: once one alters observability, preferences, market completeness, tradability, volatility regularity, or strategic environment, the problem becomes a flexible research program spanning HJB and HJB-Isaacs PDEs, duality, FBSDEs, Riccati-Volterra systems, free-boundary optimal stopping, stochastic differential games, and neural direct policy optimization.