Marginal Risk Contribution (MRC)

MRC is formally defined as the partial derivative of portfolio volatility (\sigma_p) with respect to asset weight (w_i):

[MRC_i = \frac{\partial \sigma_p}{\partial w_i} = \frac{(\Sigma w)_i}{\sigma_p}]

where (\Sigma) is the covariance matrix and ((\Sigma w)_i) is the i-th row of the matrix-vector product. This represents the marginal impact of an infinitesimal weight increase on total risk.

The total risk contribution (TRC) of an asset is the product of its weight and its MRC:

[TRC_i = w_i \times MRC_i = w_i \cdot \frac{\partial \sigma_p}{\partial w_i}]

By Euler's homogeneous function theorem, the sum of all TRCs exactly equals total portfolio volatility:

[\sum_{i=1}^{n} TRC_i = \sigma_p]

This additive property is what makes MRC the foundational building block for risk budgeting and equal risk contribution optimization.

MRC is highly sensitive to cross-asset correlations, unlike naive inverse-volatility weighting:

• A highly correlated asset with the rest of the portfolio will have an elevated MRC because its marginal addition compounds existing risk exposures.
• A negatively correlated asset can have a negative MRC, meaning adding more of it actually reduces total portfolio volatility.
• This property makes MRC essential for identifying true diversifiers versus assets that merely appear low-risk in isolation.

MRC is not constant—it changes as portfolio weights change:

• As an asset's weight increases, its MRC typically rises non-linearly due to its growing influence on the portfolio's covariance structure.
• This creates a feedback loop in risk parity optimization: adjusting weights to equalize risk contributions alters the MRCs themselves, requiring iterative convex optimization.
• The non-linearity is why closed-form solutions for risk parity only exist in trivial cases; real implementations rely on sequential quadratic programming or alternating direction method of multipliers.

MRC inherits and amplifies errors from the estimated covariance matrix:

• Sampling error in historical correlations directly distorts MRC values, potentially misidentifying which assets truly drive portfolio risk.
• Estimation error is most severe in high-dimensional portfolios where the number of assets approaches the number of historical observations.
• Mitigation strategies include covariance shrinkage (Ledoit-Wolf), random matrix theory filtering, and hierarchical clustering to denoise the correlation matrix before MRC computation.
• In practice, we validate MRC stability through sensitivity analysis—perturbing the covariance lookback window and measuring weight drift.

MRC can be decomposed beyond individual assets into underlying risk factors:

• Using a linear factor model (r = Bf + \epsilon), the MRC of each factor is computed as the derivative of portfolio volatility with respect to factor exposure rather than asset weight.
• This enables risk factor parity, where risk contributions are balanced across macro factors (inflation, growth, credit) rather than asset classes.
• Factor-level MRC reveals hidden concentration risk—a portfolio with equal asset weights may still have 80% of its risk driven by a single equity beta factor.

The mathematical theorem that makes MRC actionable. For a homogeneous risk function of degree one (like volatility), Euler's theorem states that total portfolio risk equals the sum of each asset's weight multiplied by its Marginal Risk Contribution.

• Perfectly decomposes total volatility into additive parts
• Ensures the sum of component risks equals 100% of portfolio risk
• Underpins all risk budgeting and parity optimization solvers

The specific optimization where MRC is the objective function. ERC algorithms iteratively adjust weights until every asset's Marginal Risk Contribution multiplied by its weight is identical across the portfolio.

• Objective: equalize the product of weight and MRC for all constituents
• Solved via convex optimization or sequential quadratic programming
• Produces weights that are a compromise between inverse-volatility and minimum-variance portfolios, fully respecting the correlation structure

The estimation technique that makes MRC usable out-of-sample. Raw sample covariance matrices produce unstable MRC values due to estimation error; shrinkage blends the sample matrix with a structured target like constant correlation.

• Ledoit-Wolf shrinkage dramatically improves the condition number of the covariance matrix
• Stabilizes MRC gradients so optimization solvers converge reliably
• Reduces turnover in risk parity portfolios by preventing MRC estimates from flipping sign between rebalancing dates

A tail-risk extension where MRC is computed against Expected Shortfall rather than volatility. This shifts the risk decomposition focus from average fluctuations to the expected loss in the worst-case scenarios beyond the VaR threshold.

• MRC under CVaR identifies which assets drive portfolio losses during crashes
• Requires non-parametric or Monte Carlo estimation of the tail distribution
• Equalizing CVaR contributions produces portfolios that are far more robust to drawdowns than volatility-based risk parity

A diversification diagnostic derived directly from MRC values. ENB is the exponential of the entropy of risk contributions, quantifying how many truly independent sources of risk a portfolio holds.

• An ENB of 1.0 means all risk is concentrated in a single bet
• An ENB equal to the number of assets indicates perfect risk parity
• Portfolio managers use ENB trends to detect when MRC concentrations are silently building before they manifest in drawdowns