LQ Risk-Sensitive Portfolio Optimization

• Linear–quadratic risk-sensitive portfolio optimization is a dynamic asset allocation framework that models state-space dynamics and quadratic objectives with diverse risk formulations.
• It showcases multiple risk criteria, including dynamic mean–variance, entropic, and equilibrium formulations, enabling optimal feedback via Riccati equations and Kalman filtering.
• The approach extends to jump-diffusion, reinforcement learning, and constraint settings, offering robust strategies for managing practical asset allocation challenges.

Linear-quadratic risk-sensitive portfolio optimization denotes a family of dynamic asset-allocation problems in which portfolio, wealth, or factor dynamics are modeled in linear state-space form and the objective is chosen so that optimal strategies can be characterized through quadratic value functions, Riccati equations, Kalman filtering, or closely related forward-backward systems. In the literature represented here, “risk-sensitive” has several distinct meanings rather than a single canonical one: a dynamic Markowitz mean–variance criterion, an exponential entropic criterion of the form $\frac{1}{\theta}\log\mathbb{E}[e^{\theta(\cdot)}]$, quadratic risk minimization, and explicit constraints on predictive or ergodic variance (Abeille et al., 2016, Broek et al., 2012, Dong et al., 9 Jul 2025).

1. Formulations of risk sensitivity

A central distinction is between Markowitz-style and exponential risk sensitivity. In the LQG portfolio formulation of "LQG for portfolio optimization" (Abeille et al., 2016), the investor minimizes an infinite-horizon average of

$$\lambda\,\mathbb{V}(PnL_{t,t+1}\mid\mathcal F_t)-\mathbb{E}(PnL_{t,t+1}\mid\mathcal F_t),$$

so risk aversion enters through the scalar $\lambda>0$ as a per-period variance penalty. In that setting, the criterion is risk-sensitive in the mean–variance sense, not in the exponential-of-quadratic sense.

A second formulation uses the entropic or exponential criterion

$$J^\theta_u(t,x)=\frac{1}{\theta}\log \mathbb{E}_{t,x}\!\left[\exp\!\big(\theta C^u(t,X^u)\big)\right],$$

with $\theta\neq 0$ (Broek et al., 2012). Its small-$|\theta|$ expansion yields an expected-cost term plus $\frac{\theta}{2}$ times the variance of the cost, so $\theta>0$ is risk-averse and $\theta<0$ is risk-seeking. The portfolio application in "Relationship between Maximum Principle and Dynamic Programming Principle for Risk-Sensitive Stochastic Optimal Control Problems with Applications" (Dong et al., 9 Jul 2025) adopts this entropic structure in terminal log wealth,

$$-\frac{1}{\theta}\log \mathbb{E}\!\left[\exp\{-\theta\log V_T\}\right],$$

which becomes an LQ risk-sensitive control problem after an Itô expansion and a change of measure.

A third strand is mean–variance with time inconsistency. Hu, Jin, and Zhou formulate a stochastic LQ problem in which the cost contains a quadratic term in the conditional expectation of terminal wealth and a state-dependent linear term, so the Bellman principle fails and an equilibrium, rather than an optimal control in the dynamic-programming sense, is required (Hu et al., 2011). In that literature, the variance penalty is still quadratic and LQ in form, but the solution concept is an open-loop equilibrium characterized by a flow of BSDEs.

These formulations are related but not interchangeable. A common misconception is that “risk-sensitive” automatically means exponential utility. The cited literature shows instead that linear-quadratic portfolio models use at least three non-equivalent notions: dynamic mean–variance (Abeille et al., 2016), entropic risk sensitivity (Broek et al., 2012), and equilibrium mean–variance in time-inconsistent settings (Hu et al., 2011).

2. Linear state-space portfolio models

The linear–quadratic structure is generated by modeling economically meaningful quantities as components of a linear state. In the discrete-time LQG portfolio model of (Abeille et al., 2016), the latent state satisfies

$$\lambda\,\mathbb{V}(PnL_{t,t+1}\mid\mathcal F_t)-\mathbb{E}(PnL_{t,t+1}\mid\mathcal F_t),$$0

where $$\lambda\,\mathbb{V}(PnL_{t,t+1}\mid\mathcal F_t)-\mathbb{E}(PnL_{t,t+1}\mid\mathcal F_t),$$1 is the trade vector and $$\lambda\,\mathbb{V}(PnL_{t,t+1}\mid\mathcal F_t)-\mathbb{E}(PnL_{t,t+1}\mid\mathcal F_t),$$2 can contain inventory $$\lambda\,\mathbb{V}(PnL_{t,t+1}\mid\mathcal F_t)-\mathbb{E}(PnL_{t,t+1}\mid\mathcal F_t),$$3, predictors, impact states, and lagged returns. The PnL is defined from decision prices, execution prices, positions, and trades,

$$\lambda\,\mathbb{V}(PnL_{t,t+1}\mid\mathcal F_t)-\mathbb{E}(PnL_{t,t+1}\mid\mathcal F_t),$$4

so execution costs and price impact are built into the objective rather than appended ex post.

That framework accommodates return predictability, transient and permanent impact, and partial observability in a single state-space representation. The impact state can have exponential decay, and execution returns are modeled so that impact costs are included directly through the execution-price convention (Abeille et al., 2016). This suggests a broad LQ template: augment the state until PnL, inventory dynamics, and predictive signals become linear functions of state and control.

Continuous-time portfolio models achieve the same reduction through factor dynamics and measure changes. In the jump-diffusion factor model of Davis and Lleo, the factor process is linear,

$$\lambda\,\mathbb{V}(PnL_{t,t+1}\mid\mathcal F_t)-\mathbb{E}(PnL_{t,t+1}\mid\mathcal F_t),$$5

while the asset prices are jump diffusions with drifts affine in $$\lambda\,\mathbb{V}(PnL_{t,t+1}\mid\mathcal F_t)-\mathbb{E}(PnL_{t,t+1}\mid\mathcal F_t),$$6 (Davis et al., 2009). After a Kuroda–Nagai change of measure, the optimization reduces to a diffusion control problem in the factor state, with jumps absorbed into the running cost and admissibility set. In the recent factor-based risk-sensitive portfolio application of (Dong et al., 9 Jul 2025), the factor state obeys

$$\lambda\,\mathbb{V}(PnL_{t,t+1}\mid\mathcal F_t)-\mathbb{E}(PnL_{t,t+1}\mid\mathcal F_t),$$7

and the risky asset drift is affine in $$\lambda\,\mathbb{V}(PnL_{t,t+1}\mid\mathcal F_t)-\mathbb{E}(PnL_{t,t+1}\mid\mathcal F_t),$$8, producing a one-factor LQ risk-sensitive portfolio problem.

A different but related construction appears in benchmarked asset allocation. In (Lleo et al., 18 Jun 2026), the benchmarked excess log-return $$\lambda\,\mathbb{V}(PnL_{t,t+1}\mid\mathcal F_t)-\mathbb{E}(PnL_{t,t+1}\mid\mathcal F_t),$$9 contains a controlled Itô integral, so the original problem is not directly Markovian. Free energy–entropy duality and a Girsanov change of measure convert it into a linear-quadratic-Gaussian stochastic differential game in the factor state. The resulting dynamics under the distorted measure are linear in state and adversarial control, and the running objective is quadratic in portfolio and adversarial controls.

3. Riccati equations, filtering, and optimal feedback

Once the state-space model and criterion are fixed, the solution architecture is typically Riccati-based. In the LQG portfolio construction of (Abeille et al., 2016), the dynamic Markowitz objective can be rewritten as

$\lambda>0$0

so the problem becomes a standard LQG control problem. Under stabilizability and detectability assumptions, the separation principle holds: a steady-state Kalman filter produces $\lambda>0$1, the algebraic Riccati equation yields a stabilizing matrix $\lambda>0$2, and the optimal control is

$\lambda>0$3

Risk aversion enters through the cost matrices $\lambda>0$4 because the variance term scales with $\lambda>0$5.

In entropic LQ control, the Riccati equation is modified by the noise covariance. The LEQG synthesis summarized in (Broek et al., 2012) yields, for

$\lambda>0$6

a quadratic value function

$\lambda>0$7

with linear feedback

$\lambda>0$8

and a risk-sensitive Riccati equation of the form

$\lambda>0$9

The extra term $$J^\theta_u(t,x)=\frac{1}{\theta}\log \mathbb{E}_{t,x}\!\left[\exp\!\big(\theta C^u(t,X^u)\big)\right],$$0 is the classical risk-sensitive correction.

The portfolio-specific continuous-time LQ model of (Dong et al., 9 Jul 2025) makes the same structure explicit. After the measure change, the factor process remains linear, the running cost is quadratic in the portfolio fraction $$J^\theta_u(t,x)=\frac{1}{\theta}\log \mathbb{E}_{t,x}\!\left[\exp\!\big(\theta C^u(t,X^u)\big)\right],$$1, and both the maximum principle and dynamic programming yield an affine state-feedback policy

$$J^\theta_u(t,x)=\frac{1}{\theta}\log \mathbb{E}_{t,x}\!\left[\exp\!\big(\theta C^u(t,X^u)\big)\right],$$2

Here $$J^\theta_u(t,x)=\frac{1}{\theta}\log \mathbb{E}_{t,x}\!\left[\exp\!\big(\theta C^u(t,X^u)\big)\right],$$3 solves a Riccati equation and $$J^\theta_u(t,x)=\frac{1}{\theta}\log \mathbb{E}_{t,x}\!\left[\exp\!\big(\theta C^u(t,X^u)\big)\right],$$4 a linear ODE. The paper identifies the adjoint variables from the maximum principle with the gradient and Hessian of the value function from dynamic programming, showing $$J^\theta_u(t,x)=\frac{1}{\theta}\log \mathbb{E}_{t,x}\!\left[\exp\!\big(\theta C^u(t,X^u)\big)\right],$$5 and $$J^\theta_u(t,x)=\frac{1}{\theta}\log \mathbb{E}_{t,x}\!\left[\exp\!\big(\theta C^u(t,X^u)\big)\right],$$6 (Dong et al., 9 Jul 2025).

In the benchmarked LQG game of (Lleo et al., 18 Jun 2026), the value function is quadratic,

$$J^\theta_u(t,x)=\frac{1}{\theta}\log \mathbb{E}_{t,x}\!\left[\exp\!\big(\theta C^u(t,X^u)\big)\right],$$7

and the saddle-point controls are affine:

$$J^\theta_u(t,x)=\frac{1}{\theta}\log \mathbb{E}_{t,x}\!\left[\exp\!\big(\theta C^u(t,X^u)\big)\right],$$8

with an analogous affine formula for the adversarial control $$J^\theta_u(t,x)=\frac{1}{\theta}\log \mathbb{E}_{t,x}\!\left[\exp\!\big(\theta C^u(t,X^u)\big)\right],$$9. A notable economic interpretation follows from the fractional Kelly decomposition: the optimal allocation is a sum of a fractional Kelly portfolio, a benchmark-tracking portfolio, and an intertemporal hedging portfolio (Lleo et al., 18 Jun 2026).

4. Well-posedness, non-arbitrage, jumps, and constraints

The existence of a stabilizing Riccati solution is not automatic in financial applications. In (Abeille et al., 2016), the quadratic cost matrix need not be positive definite because the objective rewards expected PnL while penalizing its variance. The authors therefore use a Popov criterion and prove an equivalence with a dynamical non-arbitrage condition: for deterministic round trips, the total PnL must satisfy

$\theta\neq 0$0

In that model, existence and uniqueness of the stabilizing Riccati solution are equivalent to the absence of profitable round trips generated purely by impact. The well-posedness of the LQG portfolio problem is therefore tied to a no-manipulation condition rather than to a generic coercivity assumption (Abeille et al., 2016).

Jump models preserve part of the LQ structure while adding non-quadratic terms. Davis and Lleo show that in a jump-diffusion factor model, the running cost is quadratic in the control through the Brownian term but contains additional nonlinear integral terms from jumps; nonetheless the HJB equation has a classical solution, and the optimal control is a Markov feedback obtained from the Hamiltonian maximizer (Davis et al., 2009). A striking feature of that formulation is that the transformed factor-control problem has no jumps, so the dynamic programming equation is an HJB PDE rather than a PIDE.

Constraints introduce another layer of structure. Li and Zheng study stochastic LQ portfolio control with random coefficients and convex portfolio constraints, including cone constraints, and derive necessary and sufficient optimality conditions for both primal and dual problems through coupled FBSDEs (Li et al., 2015). In the pure quadratic risk-minimization case, the optimal wealth and portfolio processes can be written explicitly in terms of adjoint processes, and the constrained problem yields extended stochastic Riccati equations. This suggests that constrained LQ risk-sensitive portfolio optimization is naturally expressed in primal–dual FBSDE form when control constraints are intrinsic.

Risk-sensitive jump-diffusion models with regimes push the framework beyond pure LQ but preserve the same transform logic. In (Das et al., 2016), the finite-horizon objective is

$\theta\neq 0$1

with age-dependent semi-Markov regime switching and jump-diffusion asset dynamics. After the multiplicative ansatz

$\theta\neq 0$2

the HJB equation reduces to a nonlocal linear PDE for $\theta\neq 0$3, and existence and uniqueness are established through an equivalent Volterra integral equation. In the pure diffusion, constant-coefficient limit, the local Hamiltonian becomes quadratic in the portfolio weight, which is the standard LQ risk-sensitive form (Das et al., 2016).

5. Time inconsistency, mean-field interactions, and robustness

Mean–variance portfolio selection is LQ in form but can be dynamically inconsistent. Hu, Jin, and Zhou show that when the objective contains both $\theta\neq 0$4 and a state-dependent linear weight on $\theta\neq 0$5, the correct notion is an open-loop equilibrium control, defined through spike variations (Hu et al., 2011). The equilibrium is characterized by a flow of BSDEs, and in the scalar deterministic-coefficient case the equilibrium strategy is linear feedback with coefficients given by coupled Riccati-type ODEs. In the continuous-time mean–variance portfolio application, the wealth equation is linear in the transformed control and the equilibrium strategy remains affine in wealth, but it differs from feedback equilibria derived from generalized HJB equations.

Risk-sensitive LQ portfolio problems also admit multi-agent and robust reinterpretations. The matrix-valued mean-field-type game framework of (Barreiro-Gomez et al., 2019) studies risk-neutral, risk-sensitive, and adversarial linear-quadratic games with conditional mean-field interactions, jump diffusion, and regime switching. In the risk-sensitive case without jumps and with one regime, each player’s value admits a quadratic ansatz in the deviation $\theta\neq 0$6 and the conditional mean $\theta\neq 0$7, and the optimal strategy is a state-and-conditional-mean feedback. The paper also proves that, in the team case, full cooperation enlarges the well-posedness domain under risk-sensitive decision-makers by means of population risk-sharing.

A further point is the formal relation between risk sensitivity and robustness. In (Barreiro-Gomez et al., 2019), the same Riccati equation governs a risk-sensitive control problem and a robust adversarial problem with an auxiliary disturbance control. This parallels the benchmarked allocation model of (Lleo et al., 18 Jun 2026), where free energy–entropy duality rewrites the risk-sensitive benchmarked objective as a min–max LQG stochastic differential game under an equivalent probability measure. Across these works, a plausible implication is that robust-control interpretations are not merely heuristic but structurally embedded in several LQ risk-sensitive portfolio models.

6. Data-driven, reinforcement-learning, and risk-constrained extensions

Recent work extends LQ risk-sensitive portfolio optimization beyond exact model knowledge. In continuous time, (Jia, 2024) studies a risk-sensitive RL problem with the exponential objective and shows that risk sensitivity can be represented through an additional penalty term, the quadratic variation of the value process. For the Merton investment problem, power utility becomes an exponential risk-sensitive objective in terminal log wealth, and the optimal exploratory policy is Gaussian with mean

$\theta\neq 0$8

which coincides with the classical Merton proportion, and variance $\theta\neq 0$9. The same paper gives a q-learning construction for an LQ control example and reports that moderate negative risk-sensitivity parameters can improve finite-sample estimation error.

A directly financial RL formulation appears in (Lleo et al., 18 Jun 2026). There, the explicit LQG-game solution guides a continuous-time q-learning actor–critic method: the critic is quadratic in the factor state, while both the portfolio actor and the adversarial actor are affine. The learned allocation inherits the fractional Kelly decomposition of the analytical saddle-point solution, and the proof-of-concept implementation shows that the portfolio actor learns the optimal policy with high accuracy.

Online learning of risk-sensitive LQ controllers has also been analyzed in episodic settings. The episodic risk-sensitive LQR study (Xu et al., 2024) considers the finite-horizon criterion

$|\theta|$0

with unknown system matrices. A least-squares greedy algorithm achieves $|\theta|$1 regret under an identifiability condition, while adding exploration noise yields $|\theta|$2 regret when identifiability fails. Although that paper is not a portfolio paper, it provides an explicit learning-theoretic template for model-adaptive risk-sensitive linear feedback.

A separate line replaces the entropic criterion by explicit variance constraints. The predictive-variance-constrained LQ framework of (Tsiamis et al., 2021) augments the classical cost with a bound on the total expected predictive variance of the state penalty, producing an affine controller relative to the MMSE estimate and an inflated Riccati equation. The ergodic-risk constrained policy optimization framework of (Talebi et al., 10 Feb 2025) constrains the asymptotic variance of a long-run martingale risk functional and solves the resulting problem by a primal–dual method, again reducing each inner step to an augmented LQR. These constructions are not portfolio models per se, but they are presented as directly interpretable in linear–quadratic portfolio settings and show that “risk-sensitive” control can also mean long-run variance control rather than entropic penalization.

Taken together, the literature presents linear-quadratic risk-sensitive portfolio optimization as a broad research area rather than a single model class. Its common denominator is the reduction of portfolio choice to linear dynamics with quadratic or exponential-quadratic objectives, but the meaning of risk sensitivity varies across dynamic Markowitz criteria, entropic control, equilibrium mean–variance, explicit risk constraints, and robust or adversarial reformulations. The recurring analytical motifs are state augmentation, measure change, Riccati equations, Kalman filtering, and FBSDE systems; the recurring economic motifs are predictability, impact, partial information, benchmarking, constraints, and robustness to diffusive or jump risk (Abeille et al., 2016, Dong et al., 9 Jul 2025).