---
title: Interval Cross-Efficiency in Portfolio Optimization
url: https://www.emergentmind.com/topics/interval-cross-efficiency-ice
type: topic
---

# Interval Cross-Efficiency in Portfolio Optimization

Interval Cross-Efficiency (ICE) is a methodology originally motivated by extensions of Data Envelopment Analysis (DEA) to settings with parameter uncertainty, most notably in uncertain or interval-valued risk-free rates for financial portfolios. ICE systematically evaluates and ranks decision-making units (DMUs) or candidate solutions by aggregating cross-efficiency scores across all admissible parameter values in a specified interval, rather than relying on an arbitrary fixed value. Recent advances have demonstrated the robustness and closed-form computability of ICE in portfolio optimization, where it yields portfolios less sensitive to specification errors or uncertainty in risk-free rates [1610.00937].

## 1. Conceptual Foundations of Cross-Efficiency and ICE

Classical cross-efficiency, developed in the context of DEA, evaluates each DMU’s efficiency not only according to its self-selected optimal weights but also by the optimal weights of all other DMUs. The interval cross-efficiency (ICE) approach generalizes this concept to account for uncertainty in evaluation parameters. In the context of portfolio selection, each DMU corresponds to a tangent portfolio parameterized by a candidate risk-free rate, $r_f \in [r_{\min}, r_{\max}]$.

For each DMU (portfolio corresponding to a specific $r_f$), a self-efficiency score is 1, while cross-efficiencies measure comparative performance under alternative parameterizations. ICE aggregates these cross-efficiencies, integrating over all possible $r_f$ values, thereby yielding an average cross-efficiency for each candidate solution.

## 2. ICE in Portfolio Optimization with Risk-Free Rate Uncertainty

The principal application of ICE in modern research is in robust Sharpe ratio portfolio construction when the risk-free rate is only known to lie within an interval $[r_{\min}, r_{\max}]$, as formalized in "Sharpe portfolio using a cross-efficiency evaluation" [1610.00937]. The classical Maximum-Sharpe-Ratio (MSR) portfolio, for fixed $r_f$, solves:
\[
\max_{w^T1=1} \frac{w^T(\mu - r_f1)}{\sqrt{w^T \Sigma w}}
\]
where $\mu$ is the vector of expected returns, $\Sigma$ is the covariance matrix, and $1$ is the unit vector.

When $r_f$ is interval-valued, one could pose a minimax problem, but ICE considers the full continuum of tangent portfolios $\{w^*(r_f)\}$ and evaluates each via cross-efficiency relative to all others across the interval.

## 3. Mathematical Structure of Interval Cross-Efficiency

Given the family of tangent portfolios and their risk-return pairs,
\[
(\sigma^*(r_f),\,r^*(r_f)) = (\sqrt{w^*(r_f)^T\Sigma w^*(r_f)},\,w^*(r_f)^T \mu),
\]
the cross-efficiency of DMU $i$ evaluated under the optimal weights of DMU $j$ is:
\[
\mathrm{Ef}_i(r_f^j) = \frac{ \frac{ r^*(r_f^i) - r_f^j }{ \sigma^*(r_f^i) } }{ \frac{ r^*(r_f^j) - r_f^j }{ \sigma^*(r_f^j) } } = \frac{ \mathrm{SR}_i(r_f^j) }{ \mathrm{SR}_j(r_f^j) },
\]
where $\mathrm{SR}_i(r_f^j)$ denotes the Sharpe ratio of portfolio $i$ using risk-free rate $r_f^j$. Averaging over the interval, the cross-efficiency score is:
\[
CE_i = \frac{1}{r_{\max}-r_{\min}} \int_{r_{\min}}^{r_{\max}} \frac{(r^*(r_f^i) - r_f) / \sigma^*(r_f^i)}{(r^*(r_f) - r_f) / \sigma^*(r_f)} dr_f.
\]
The ICE portfolio is the candidate $w^*(r_i^*)$ that maximizes $CE_i$.

## 4. Closed-Form ICE Portfolio Solution and Hyperbolic Geometry

When short selling is allowed ($\Sigma$ invertible), each MSR portfolio admits the closed-form expression:
\[
w^*(r_f) = \frac{ \Sigma^{-1} (\mu - r_f 1)}{ 1^T \Sigma^{-1} (\mu - r_f 1) }.
\]
Writing $a = \mu^T \Sigma^{-1} \mu$, $b = 1^T \Sigma^{-1} 1$, $c = 1^T \Sigma^{-1} \mu$, the global minimum variance (GMV) portfolio has $r_{GMV}^* = c/b$, $\sigma_{GMV}^* = 1/\sqrt{b}$. Each tangent portfolio satisfies:
\[
r^*(r_f) - r_f = \frac{ a - c r_f }{ 1^T \Sigma^{-1} (\mu - r_f 1) }, \quad \sigma^*(r_f) = \frac{ \sqrt{ a - 2c r_f + b r_f^2 } }{ 1^T \Sigma^{-1} (\mu - r_f 1 ) }.
\]
The ICE-optimal $r_i^*$ is determined by maximizing $CE_i$ and is given by:
\[
r_i^* = r_{GMV}^* + \sigma_{GMV}^* \frac{ I_2 }{ I_1 },
\]
where $I_1 = \int_{r_{\min}}^{r_{\max}} \frac{ dr }{ \sqrt{ a - 2c r + b r^2 } }$, $I_2 = \int_{r_{\min}}^{r_{\max}} \frac{ r dr }{ \sqrt{ a - 2c r + b r^2 } }$. The final weight vector is
\[
w^*_{ICE} = w^*(r_i^*) = \frac{ \Sigma^{-1} ( \mu - r_i^* 1 ) }{ 1^T \Sigma^{-1} ( \mu - r_i^* 1 ) }.
\]
Closed-form evaluations for $I_1$ and $I_2$ involving logarithmic functions of the hyperbola’s asymptotes are presented in [1610.00937].

## 5. Algorithmic Implementation

In the absence of short-selling, ICE computation involves iterative quadratic programming. In the short-selling-allowed regime, the closed-form solution described above is employed.

The computational workflow consists of:
- Generating a grid of $N+1$ risk-free rates $r_f^j$ over $[r_{\min}, r_{\max}]$,
- Solving for the tangency portfolio $w^{(j)}$ at each $r_f^j$,
- Computing each portfolio’s average cross-efficiency $CE_i$ across the interval,
- Selecting the portfolio achieving maximal $CE_i$.

The explicit pseudocode is as follows [1610.00937]:
```
Input: μ, Σ, [r_min, r_max], grid size N
1.  Grid r_f^j = r_min + (j−1)*(r_max−r_min)/N, j=1…N+1.
2.  For each j solve
      min_x x^T Σ x
      s.t. x^T(μ − r_f^j 1) = 1,  x ≥ 0.
    Set w^(j)= x/ (1^T x). Compute:
      r_j = w^(j)^T μ,   σ_j = √[w^(j)^T Σ w^(j)].
3.  For each i=1…N+1 compute
      CE_i = (1/(N+1)) ∑_{j=1}^{N+1} [(r_i − r_f^j)/σ_i] / [(r_j − r_f^j)/σ_j].
4.  Pick i* = argmax_i CE_i; output w^(i*).
```

With short-sales allowed, use the closed-form $w^*(r_f^j)$.

## 6. Robustness and Interpretation

Portfolios constructed using ICE are robust to misspecification of $r_f$. Rather than choosing a portfolio optimized for an arbitrary or potentially misestimated value of $r_f$, the ICE approach performs a peer-evaluation across all admissible portfolios, relative to all values in $[r_{\min}, r_{\max}]$. This model-averaging mitigates estimation risk by favoring portfolios that maintain strong relative Sharpe performance uniformly across the interval, rather than specializing for a single risk-free rate. This methodology generalizes beyond financial applications wherever DEA-based efficiency evaluations under parameter uncertainty are required [1610.00937].

Source: https://www.emergentmind.com/topics/interval-cross-efficiency-ice