Mean–Variance Optimisation With a Return Target
In this chapter
Intermediate
Mean–variance optimisation uses expected return and variance to choose a portfolio from a set of allowed holdings. If you want a particular expected return, the method finds the allowed mix with the smallest variance at that target. Before calculating risk, check that the allowed weights can reach the return at all.
Write down the inputs and the rules
Start with a forecast and a volatility for each asset over the same period. Include a covariance for each pair. Covariance tells you whether the returns tend to deviate from their averages together. The optimiser needs that information because changing a weight changes the asset’s contribution and its interaction with other holdings.
The example uses three hypothetical funds. All forecasts describe one-year nominal USD simple total returns including income, before fees and taxes. The weights refer to the opening portfolio and stay fixed through the period. There is no borrowing, shorting or unallocated cash. The target is an exact 8.00% expected return. An “at least” target would allow a different set of portfolios. Dropping the return target entirely gives the global minimum-variance portfolio under those same weight limits.
| Fund | Expected return | Volatility (pp) | Variance (pp²) |
|---|---|---|---|
| A | 4% | 5 | 25 |
| B | 8% | 10 | 100 |
| C | 12% | 15 | 225 |
Assume all three pairwise covariances are zero. That keeps the arithmetic short. In a real optimisation, the pairwise cross terms matter whenever covariances are nonzero. All the inputs are estimates, so it helps to consider the risk–return trade-off before putting weight on small differences between portfolios.
Worked case. Find the least variance at exactly 8%
Let \(w_A,w_B,w_C\) be the opening weights as decimals. Because the weights sum to one, use 8% as your starting point. Each dollar shifted from B to A reduces the forecast by four points per dollar of portfolio weight. The same shift to C adds four points. To hit 8% exactly, set \(w_A=w_C=t\) and \(w_B=1-2t\), with \(0\le t\le0.50\). Every allowed mix at 8% is now on this one scale.
With three assets there are three pairs, A–B, A–C and B–C. For each pair, its cross term is twice the product of its two weights and its covariance. Here those three terms are zero, leaving the general rule for uncorrelated assets, \(\sigma_p^2=\sum_{i=1}^{3}w_i^2\sigma_i^2\). In this rule \(\sigma_p^2\) is portfolio variance, \(i\) runs through the three funds, \(w_i\) is an opening weight and \(\sigma_i^2\) is that fund’s variance. In pp², the portfolio variance \(V(t)\) is
\[ \begin{aligned} V(t)&=25t^2+100(1-2t)^2+225t^2\\ &=650t^2-400t+100\\ &=650\left(t-\frac{4}{13}\right)^2+\frac{500}{13}. \end{aligned} \]
The last form shows exactly where variance is lowest. The square cannot be negative and is zero at \(t=4/13\). That fits the interval. The weights are \(4/13,\ 5/13,\ 4/13\), about 30.77%, 38.46% and 30.77%. Variance is \(500/13\) pp², giving volatility of 6.20 pp. Holding just B also gives an expected 8%, with volatility of 10 pp. The mix improves on B because the assets’ modelled moves provide offsets.
Apply a weight cap and check the best attainable return
Worked case. Put a 25% limit on C
Now the mandate allows at most 25% in C. At the exact 8% target, \(w_A=w_C=t\) still holds, so the feasible range is \(0\le t\le0.25\). The unconstrained optimum, \(4/13\), is outside it. On the remaining interval, the square in the formula is smallest at \(t=0.25\), the closest feasible point. That gives 25% A, 50% B and 25% C, with variance of \(25(0.25)^2+100(0.50)^2+225(0.25)^2=40.625\) pp² and volatility of 6.37 pp. The target survives. The cap makes the best attainable volatility higher at that return.
What if the mandate also calls for a 9.25% target? Get the highest feasible return by putting the maximum 25% in C and the rest in B, the next-highest forecast. The result is \(0.25(12\%)+0.75(8\%)=9.00\%\). The target is outside the opportunity set. Taking a larger allocation in C would break the cap, so changing the variance estimate cannot make 9.25% feasible with these forecasts.
Check how much of the answer comes from the estimates
Each set of inputs defines its own answer. If B’s forecast falls, equal A and C weights may no longer hit the target. If the funds start moving together, the cross terms can raise the volatility of an allocation that looked attractive. Recalculate with plausible alternative forecasts, covariances and constraints. With more assets, you can enter the budget, target and limits into a quadratic optimiser. It minimises an expression containing squared weights and products of weights. Check the returned weights against the original rules, then recompute return and variance with the unrounded weights. If the solver says the target is infeasible, compute the lowest and highest expected returns the rules allow before changing the risk inputs. Variance counts large gains as well as losses, so a mandate based on avoiding a particular shortfall needs a measure that asks about that shortfall directly. In strategic asset allocation, these estimates help you choose long-term weights subject to the actual mandate.
Practice questions
Choose one answer per question. Round only the final result where needed.
1. Use the target to connect weights
Three assets have expected returns of 3%, 7% and 11%. A fully invested long-only mix targets exactly 7%. Which relationship must hold?
2. Check a cap before minimising
The forecasts are 4%, 8% and 12%. The highest-return asset is capped at 40%. What is the maximum expected return without borrowing or shorting?
3. Find the constrained point
Variance at a fixed target is \(V(t)=200(t-0.40)^2+25\), where \(t\) is a decimal weight. Only \(0\le t\le0.30\) is allowed. Which value is best?
4. Test feasibility with a lower bound
Fund A is expected to return 4% and B 10%. A must have at least a 50% weight. Is an exact 8% target feasible in a fully invested long-only portfolio?
5. Read the optimiser’s objective
Two allowed portfolios both target 9%. One has variance 49 pp² and the other 64 pp². Which does variance minimisation prefer?
6. Change a covariance estimate
At the same positive weights and individual variances, an off-diagonal covariance estimate rises. What happens to the variance of that particular mix?