Risk Budgeting Portfolio Concepts

• Risk budgeting portfolios are defined by aligning each asset's risk contribution with a predetermined risk budget vector using Euler’s decomposition.
• They employ various convex optimization methods—such as logarithmic barrier formulations and sequential quadratic programming—to compute weights across different risk measures.
• Numerical methods like cyclical coordinate descent, ADMM, and Mirror Descent enable efficient convergence even under practical constraints and forecast integrations.

A risk-budgeting portfolio is a portfolio in which the ex-ante contribution of each asset to total portfolio risk matches a pre-specified risk budget vector $b$, typically under long-only and full-investment constraints. In its classical form, the approach replaces return forecasting with explicit risk-allocation targets: if $RC_i(w)$ denotes the risk contribution of asset $i$ and $R(w)$ the portfolio risk, the defining condition is $RC_i(w)=b_iR(w)$ for all $i$. The special case $b_i=1/n$ is Equal Risk Contribution, commonly associated with risk parity (Roncalli, 2014, Cetingoz et al., 2022).

1. Mathematical definition

Let $$w=(w_1,\dots,w_d)'\in\Delta_d:=\{w\in\mathbb R_+^d:\sum_i w_i=1\}$$ be the portfolio weights, and let the portfolio-risk map be

$R(w):=\rho(-w'X),$

where $\rho$ is positively homogeneous of degree $RC_i(w)$0, sub-additive, and differentiable on $RC_i(w)$1. By Euler’s theorem on homogeneous functions,

$RC_i(w)$2

The risk contribution of asset $RC_i(w)$3 is therefore defined by

$RC_i(w)$4

and satisfies $RC_i(w)$5. Given a strictly positive budget vector $RC_i(w)$6 with $RC_i(w)$7, a $RC_i(w)$8-risk-budgeting portfolio $RC_i(w)$9 is characterized by

$i$0

These identities provide the general definition independently of the specific choice of risk measure (Cetingoz et al., 2022).

For volatility-based risk budgeting, with covariance matrix $i$1,

$i$2

the marginal contribution to volatility is

$i$3

and the total contribution is

$i$4

The budgeting condition becomes

$i$5

In this standard-deviation case, there is no closed-form $i$6 for $i$7 (Parra-Diaz et al., 28 Apr 2025, Uysal et al., 2021).

2. Optimization formulations and existence theory

Risk-budgeting equations admit several equivalent or closely related optimization formulations. A direct approach is the nonlinear system

$i$8

possibly implemented through a squared-error objective such as

$i$9

A second, widely used formulation introduces a logarithmic barrier: $R(w)$0 followed by normalization $R(w)$1. In the volatility case, standard convex reformulations use auxiliary variables $R(w)$2, for example

$R(w)$3

or, equivalently in another representation,

$R(w)$4

These formulations are central because they transform the allocation problem into a strictly convex program over exposures, after which the final weights are recovered by normalization (Roncalli, 2014, Parra-Diaz et al., 28 Apr 2025, Uysal et al., 2021).

A general existence-and-uniqueness result is available when $R(w)$5 is continuous, convex, $R(w)$6 on $R(w)$7, positive on $R(w)$8, and positively homogeneous of degree $R(w)$9, with $RC_i(w)=b_iR(w)$0 and $RC_i(w)=b_iR(w)$1. Under these hypotheses, there exists a unique solution $RC_i(w)=b_iR(w)$2 to

$RC_i(w)=b_iR(w)$3

One route to the proof introduces the strictly convex function

$RC_i(w)=b_iR(w)$4

where $RC_i(w)=b_iR(w)$5 is convex, increasing, and $RC_i(w)=b_iR(w)$6, and shows that the unique minimizer $RC_i(w)=b_iR(w)$7 yields the risk-budgeting solution after normalization. This framework extends the theory beyond variance to a wide spectrum of homogeneous risk measures (Cetingoz et al., 2022).

3. Numerical computation

Classical numerical methods include Jacobi-type or cyclical coordinate-descent updates. In one standard scheme, starting from $RC_i(w)=b_iR(w)$8, one iterates

$RC_i(w)=b_iR(w)$9

followed by renormalization to $i$0. Newton, quasi-Newton, sequential quadratic programming, and root-finding methods are also used for the barrier or equation-based formulations (Roncalli, 2014, Cetingoz et al., 2022).

For constrained problems, a unified large-scale solver combines cyclical coordinate descent, ADMM, proximal operators, and Dykstra’s algorithm. The constrained objective is written as

$i$1

with ADMM splitting between the risk-budgeting term and the indicator of the feasible set $i$2. The $i$3-update is handled by cyclical coordinate descent, while the $i$4-update is the Euclidean projection onto $i$5, potentially via Dykstra’s algorithm when $i$6 is an intersection of simpler sets. An outer bisection or Newton search on $i$7 enforces the budget normalization condition (Richard et al., 2019).

A different line of work avoids auxiliary optimization problems altogether. For a general risk measure with Euler contributions, one may define

$i$8

and choose $i$9 so that $b_i=1/n$0 with $b_i=1/n$1. Under the stated assumptions, $b_i=1/n$2 is a Cauchy sequence in the simplex, its limit satisfies the risk-budgeting condition, and the induced fixed-point map $b_i=1/n$3 supports existence-and-uniqueness arguments. In the standard-deviation case, each iteration costs $b_i=1/n$4, convergence is geometric with rate at least $b_i=1/n$5, and numerical experiments up to $b_i=1/n$6 report that the fixed-point algorithm with $b_i=1/n$7 typically converges in under $b_i=1/n$8 seconds with accuracy uniformly below $b_i=1/n$9 in $$w=(w_1,\dots,w_d)'\in\Delta_d:=\{w\in\mathbb R_+^d:\sum_i w_i=1\}$$0 (Fassino et al., 16 Mar 2026).

When the risk measure is only available through simulation, cutting-planes and stochastic methods become natural. For arbitrary coherent distortion risk measures, one may iteratively evaluate the sampled risk $$w=(w_1,\dots,w_d)'\in\Delta_d:=\{w\in\mathbb R_+^d:\sum_i w_i=1\}$$1, compute a subgradient, and add supporting hyperplanes to an outer approximation. For Expected Shortfall, specialized cutting-planes based on the Rockafellar–Uryasev representation and projected SGD on the exposure vector $$w=(w_1,\dots,w_d)'\in\Delta_d:=\{w\in\mathbb R_+^d:\sum_i w_i=1\}$$2 and threshold $$w=(w_1,\dots,w_d)'\in\Delta_d:=\{w\in\mathbb R_+^d:\sum_i w_i=1\}$$3 are available (Costa et al., 2023).

Mirror Descent has recently been adapted to the same convex exposure formulation. Using the negative-entropy prox-function

$$w=(w_1,\dots,w_d)'\in\Delta_d:=\{w\in\mathbb R_+^d:\sum_i w_i=1\}$$4

and a tamed gradient $$w=(w_1,\dots,w_d)'\in\Delta_d:=\{w\in\mathbb R_+^d:\sum_i w_i=1\}$$5 with $$w=(w_1,\dots,w_d)'\in\Delta_d:=\{w\in\mathbb R_+^d:\sum_i w_i=1\}$$6, deterministic and stochastic Mirror Descent algorithms converge to the unique minimizer. For the averaged deterministic method, the paper states an $$w=(w_1,\dots,w_d)'\in\Delta_d:=\{w\in\mathbb R_+^d:\sum_i w_i=1\}$$7 rate with $$w=(w_1,\dots,w_d)'\in\Delta_d:=\{w\in\mathbb R_+^d:\sum_i w_i=1\}$$8, and for the stochastic method it establishes an $$w=(w_1,\dots,w_d)'\in\Delta_d:=\{w\in\mathbb R_+^d:\sum_i w_i=1\}$$9-type bound up to log-factors along averaged iterates. Numerical experiments up to dimension $R(w):=\rho(-w'X),$0 report that naive “classical-SGD” without the $R(w):=\rho(-w'X),$1-factor diverged in $R(w):=\rho(-w'X),$2–$R(w):=\rho(-w'X),$3 of trials, while Tamed-SGD and stochastic Mirror Descent never diverged across $R(w):=\rho(-w'X),$4 repetitions (Iglesias et al., 2024).

4. Extensions of the risk measure and time horizon

Risk budgeting is not restricted to variance. The general theory explicitly targets positive homogeneous and sub-additive risk measures and has been developed for Expected Shortfall, spectral or distortion measures, deviation measures, and Variantiles (Cetingoz et al., 2022, Iglesias et al., 2024).

A dynamic extension replaces the static risk functional by a time-consistent dynamic risk measure $R(w):=\rho(-w'X),$5. For portfolio weights $R(w):=\rho(-w'X),$6, one defines the risk-to-go recursively by

$R(w):=\rho(-w'X),$7

and the dynamic risk contribution of asset $R(w):=\rho(-w'X),$8 at time $R(w):=\rho(-w'X),$9 by a one-sided Gâteaux derivative. Under coherent dynamic distortion risk measures, the allocation problem is recast as a sequence of strictly convex optimization problems

$\rho$0

Each problem has a unique minimizer, the first-order conditions yield $\rho$1, and the backward sequence of solutions defines a self-financing dynamic risk-budgeting strategy with initial wealth $\rho$2 (Pesenti et al., 2023).

In continuous time, terminal variance can itself be treated as the risk measure: $\rho$3 The Gâteaux derivative is represented by a signed measure on the predictable $\rho$4-algebra, with density $\rho$5, interpreted as the instantaneous marginal risk contribution process. The asset-level risk contributions are

$\rho$6

and satisfy the aggregation property

$\rho$7

Prescribed predictable budget processes $\rho$8 are enforced by the stochastic convex program

$\rho$9

whose minimizer satisfies $RC_i(w)$00. In this framework, the Moreira–Muir volatility-managed portfolio is recovered as a risk-budgeting solution, while continuous-time mean-variance allocation appears concentrated in terms of risk contribution (Zhao et al., 2020).

Tail-relative versions have also been studied. Under the Normal-Tempered-Stable market model, one can define portfolio CoVaR and CoCVaR relative to a benchmark and compute the marginal contributions

$RC_i(w)$01

Risk budgeting can then be implemented either through the nonlinear system $RC_i(w)$02 or through local LP steps that reduce CoVaR or CoCVaR while respecting an expected-return constraint (Kim, 2023).

5. Constraints, forecasts, factors, and learning-based models

The unconstrained theory extends imperfectly to practical constraints. With additional feasibility restrictions $RC_i(w)$03, a consistent constrained formulation preserves the logarithmic barrier and solves

$RC_i(w)$04

rather than merely minimizing squared deviations of realized from target risk contributions. This accommodates box bounds, sector constraints, turnover limits, and liquidity constraints (Richard et al., 2019).

Return forecasts and transaction costs can be integrated while maintaining convexity. One formulation considers

$RC_i(w)$05

subject to

$RC_i(w)$06

Using a factorization $RC_i(w)$07, each risk-budget constraint is rewritten as the second-order-cone condition

$RC_i(w)$08

This produces a tractable SOCP that scales to asset universes around $RC_i(w)$09–$RC_i(w)$10, with the reported case studies emphasizing reduced drawdowns in equity–bond and NASDAQ 100 applications (Bhardwaj et al., 2022).

At the factor level, risk budgeting can be posed on factor exposures rather than asset weights. With a linear factor model $RC_i(w)$11, factor exposures $RC_i(w)$12, and factor-risk measure

$RC_i(w)$13

a factor-budgeting portfolio matches

$RC_i(w)$14

A combined Asset–Factor Risk Budgeting model solves

$RC_i(w)$15

thereby interpolating between pure asset-level and pure factor-level diversification. A related convex NMF approach imposes positive factor loadings and yields factors with a quasi-diagonal correlation matrix, promoting diversified long-only allocations (Cetingoz et al., 2023, Spilak et al., 2022).

Recent end-to-end neural architectures embed risk budgeting as an implicit optimization layer. In a model-based architecture, a neural network maps input features to a budget vector $RC_i(w)$16 through a softmax layer, after which a differentiable convex program solves for the allocation $RC_i(w)$17. In market data from 2017–2021, the model-based end-to-end approach trained on Sharpe ratio achieved an out-of-sample Sharpe ratio of $RC_i(w)$18, compared with $RC_i(w)$19 for nominal risk parity and $RC_i(w)$20 for equal weight; a gated version with stochastic asset selection reached $RC_i(w)$21 (Uysal et al., 2021). A later deep-declarative framework replaced the standard softmax with a bounded softmax, preventing vanishing budgets and improving conditioning of the implicit layer. On seven broad-market ETFs over 2011–2021, this reduced the dispersion of cumulative returns across $RC_i(w)$22 random seeds from $RC_i(w)$23 to $RC_i(w)$24 in-sample and from $RC_i(w)$25 to $RC_i(w)$26 out-of-sample, while consistently outperforming the risk parity benchmark in the Sharpe-trained configuration (Parra-Diaz et al., 28 Apr 2025).

6. Interpretation, variants, and recurrent issues

Risk budgeting is often presented as a diversification doctrine, but the literature qualifies that interpretation. Asset-level Equal Risk Contribution can still produce concentration in latent factors, which motivates factor budgeting and joint asset–factor formulations. In the MSCI BARRA US application, Equal Factor Risk Contribution improved factor diversification but induced heavier concentration in asset weights, whereas the combined Asset–Factor Risk Budgeting construction was designed as a compromise between the two objectives (Cetingoz et al., 2023).

Homogeneity is a second recurrent issue. The core Euler decomposition depends on the risk measure being positively homogeneous of degree $RC_i(w)$27, and constrained formulations can break the scale invariance present in the unconstrained case. The “scaling incompatibility” emphasized in constrained risk budgeting arises because the logarithmic barrier enforces normalization through the Lagrange multiplier $RC_i(w)$28, while many practical constraints are not themselves homogeneous. As a result, apparently equivalent formulations of a bound can lead to distinct constrained solutions (Richard et al., 2019).

A third issue is sensitivity to the covariance structure and the asset universe. The neural end-to-end literature notes that risk-based portfolios can be sensitive to the underlying asset universe and that low-volatility assets with low returns can hurt the portfolio, motivating embedded gating mechanisms for asset filtering (Uysal et al., 2021). The online-learning literature makes a related point in a different language. Universal Risk Budgeting defines a scheme that weights each risk budget, rather than each capital budget, by its historical performance record, and proves mathematical equivalence to a novel type of universal portfolio using a new family of prior densities. The proposed construction is presented as more flexible because it allows the algorithmic trader to incorporate prior knowledge about the covariance structure of instantaneous asset returns; in particular, if there is dispersion in the volatilities of the available assets, then the uniform or Dirichlet priors standard in universal portfolio theory generate a dangerously lopsided prior distribution over the possible risk budgets (Garivaltis, 2021).

These strands together indicate that “risk budgeting portfolio” is best understood not as a single recipe but as a family of allocation principles built from Euler risk decomposition, convex or fixed-point computation, and preassigned budgets over assets, factors, times, or tail states. The shared invariant across these variants is that portfolio construction is governed by the allocation of risk rather than by the allocation of capital.