Generic Value at Risk Range Constraints

PortfolioOptimisers.set_risk_constraints! — Method
set_risk_constraints!(
    model::Model,
    i,
    r::GenericValueatRiskRange{<:Any, <:ValueatRisk{<:Any, <:Any, <:Any, <:DistributionValueatRisk}, <:ValueatRisk{<:Any, <:Any, <:Any, <:DistributionValueatRisk}},
    opt::Union{AbstractProgrammeAllocationSet, RiskJuMPOptimisationEstimator},
    pr::AbstractPriorResult,
    args...;
    prefix,
    kwargs...
) -> Any

Add GenericValueatRiskRange constraints to model when both tails are a ValueatRisk with DistributionValueatRisk.

When the two tails read the same factor of the covariance matrix, they read one standard deviation variable and one cone, as ValueatRiskRange does. Otherwise each tail builds a cone of its own, as in the method of set_risk_constraints! for a RangeRiskMeasure. The cone bounds the standard deviation $\lVert \mathbf{G}\boldsymbol{w} \rVert_2$, which does not read the expected returns, so tails with different expected returns share it too. Each tail reads its own expected returns in its expression.

The cone bounds $g$ from below only, so a model is bounded when the coefficient of $g$ is not negative. With a cone for each tail, that is $z_{\mathrm{loss}} \geq 0$ and $z_{\mathrm{gain}} \geq 0$. A Normal with a mean $m$ far from zero makes one of the two negative, and the minimisation then has no finite solution. With one cone, the condition is $z_{\mathrm{loss}} + z_{\mathrm{gain}} \geq 0$, the condition of ValueatRiskRange, which does not depend on $m$.

The tails keep their expressions, so the ub of each tail bounds its own value. A bound on a tail whose coefficient is negative lets $g$ rise above the standard deviation, so the model holds a relaxation of that bound.

Mathematical definition

\[\begin{align} \mathrm{GenVaRRange}(\boldsymbol{w}) &= \left(-\boldsymbol{\mu}_{\mathrm{loss}}^\intercal \boldsymbol{w} + z_{\mathrm{loss}}\, g\right) + \left(\boldsymbol{\mu}_{\mathrm{gain}}^\intercal \boldsymbol{w} + z_{\mathrm{gain}}\, g\right)\,, \quad g \geq \lVert \mathbf{G}\boldsymbol{w} \rVert_2\,. \end{align}\]

Where:

  • $\mathrm{GenVaRRange}(\boldsymbol{w})$: Generic Value-at-Risk range.
  • $\boldsymbol{\mu}_{\mathrm{loss}}$, $\boldsymbol{\mu}_{\mathrm{gain}}$: The expected returns vectors of the loss tail and of the gain tail.
  • $\boldsymbol{w}$: Portfolio weights vector $N \times 1$.
  • $z_{\mathrm{loss}}$, $z_{\mathrm{gain}}$: The z-scores of the loss tail and of the gain tail.
  • $g$: Standard deviation variable.
  • $\mathbf{G}$: Cholesky factor of the covariance matrix.

Algorithm

  1. Read the tails loss and gain with range_tails, and the key name with range_risk_key. Read the Cholesky factor of each tail with chol_sigma_selector. When the two factors are equal, build the cone of the shared standard deviation g_var with set_distribution_var_cone!. Otherwise g_var is nothing, and each tail builds a cone of its own.
  2. Build the two tails, register their sum and bound it, as steps 2 to 5 of the method of set_risk_constraints! for a RangeRiskMeasure do. Each tail reads g_var. This copy of the steps agrees with that method.

Arguments

  • model::JuMP.Model: The JuMP optimisation model.
  • i: Constraint index for unique variable and constraint naming.
  • r::GenericValueatRiskRange: The generic Value-at-Risk range risk measure, with two parametric tails.
  • opt::RiskConstraintOwner: The owner of the risk constraint, a risk-based JuMP optimisation estimator or a programme Allocation Set.
  • pr::AbstractPriorResult: Prior result containing the returns matrix X.

Keyword arguments

  • prefix::Symbol: Model State namespace (default: empty, i.e. the bare key).

Returns

  • range_risk: The combined loss + gain risk expression added to the model, registered under the key genvar_range_risk_.

Related

source