Appendix: Why the Mean-Variance Objective?
Why is the mean–variance (MV) objective function \[ \max_w \mathbf{w}^T \mathbf{\mu} - \frac{\gamma}{2} \mathbf{w}^T \mathbf{\Sigma} \mathbf{w} \] a valid approximation of investor preferences?
Here is one justification of the Mean-Variance optimization problem using the Taylor Series Approximation. Refer to Levy and Markowitz (1979) for more on this topic.
The Goal
We start with an investor who wants to maximize the expected utility of their final wealth, \(E[U(W)]\). We want to show that it can be approxmiated by the optimization problem: \(\max_w \mathbf{w}^T \mathbf{\mu} - \frac{\gamma}{2} \mathbf{w}^T \mathbf{\Sigma} \mathbf{w}\).
Step 1: The Setup
Let
- \(W\) be the random variable representing final wealth
- \(\mu_W = E[W]\) be the expected final wealth
- \(\sigma^2_W = E[(W - \mu_W)^2]\) be the variance of final wealth
- \(U(W)\) be a utility function where \(U'>0\) (more wealth is good) and \(U''<0\) (risk aversion)
We define the Certainty Equivalent (CE), the guaranteed cash amount that provides the same utility as the risky wealth \(W\). \[U(CE) = E[U(W)]\]
Since utility increases with wealth, maximizing Expected Utility is equivalent to maximizing the Certainty Equivalent (CE).
Step 2: Approximating the “Risky” Side (\(E[U(W)]\))
We approximate the utility of the random wealth \(W\) using a second-order Taylor Series expansion around the expected wealth \(\mu_W\).
\[U(W) \approx U(\mu_W) + U'(\mu_W)(W - \mu_W) + \frac{1}{2}U''(\mu_W)(W - \mu_W)^2\]
Now, we apply the expectation operator \(E[\cdot]\) to the entire equation:
\[E[U(W)] \approx E[U(\mu_W)] + E[U'(\mu_W)(W - \mu_W)] + E\left[\frac{1}{2}U''(\mu_W)(W - \mu_W)^2\right]\]
Since \(\mu_W\), \(U(\mu_W)\), and its derivatives are constants, they move outside the expectation:
\[E[U(W)] \approx U(\mu_W) + U'(\mu_W)\underbrace{E[W - \mu_W]}_{0} + \frac{1}{2}U''(\mu_W)\underbrace{E[(W - \mu_W)^2]}_{\sigma^2_W}\]
Note: The middle term is zero because the expected deviation from the mean is always zero.
Result A: \[E[U(W)] \approx U(\mu_W) + \frac{1}{2}U''(\mu_W)\sigma^2_W\]
Step 3: Approximating the “Certainty” Side (\(U(CE)\))
We also approximate the utility of the Certainty Equivalent \(CE\). We expand this around \(\mu_W\) as well. (This is valid assuming the risk premium is small, so \(CE\) is close to \(\mu_W\)).
\[U(CE) \approx U(\mu_W) + U'(\mu_W)(CE - \mu_W)\]
(We stop at the first order here because \(CE\) is a deterministic constant, not a random variable, and we assume the difference \((CE-\mu_W)\) is small enough that the squared term is negligible.)
Result B: \[U(CE) \approx U(\mu_W) + U'(\mu_W)(CE - \mu_W)\]
Step 4: Equating and Solving for CE
Recall our definition: \(U(CE) = E[U(W)]\). Therefore, Result A = Result B.
\[U(\mu_W) + U'(\mu_W)(CE - \mu_W) \approx U(\mu_W) + \frac{1}{2}U''(\mu_W)\sigma^2_W\]
Subtract \(U(\mu_W)\) from both sides:
\[U'(\mu_W)(CE - \mu_W) \approx \frac{1}{2}U''(\mu_W)\sigma^2_W\]
Divide by \(U'(\mu_W)\) to isolate \(CE\):
\[CE - \mu_W \approx \frac{1}{2} \frac{U''(\mu_W)}{U'(\mu_W)} \sigma^2_W\]
\[CE \approx \mu_W + \frac{1}{2} \frac{U''(\mu_W)}{U'(\mu_W)} \sigma^2_W\]
Step 5: Introducing the Risk Aversion Parameter
We substitute the Arrow-Pratt coefficient of absolute risk aversion, defined as \(A = -\frac{U''(\mu_W)}{U'(\mu_W)}\).
\[CE \approx \mu_W - \frac{1}{2} A \sigma^2_W\]
Step 6: Converting Wealth to Portfolio Returns
The equation above is in terms of wealth. We must convert it to portfolio returns.
Let \(W_0\) be initial wealth and \(R_p\) be the random portfolio return with mean \(\mu_p\) and variance \(\sigma^2_p\). Final wealth is: \(W = W_0(1 + R_p)\). We can compute the mean and variance of final wealth \(W\):
- Mean: \(\mu_W = W_0(1 + \mu_p) = W_0 + W_0 \mu_p\)
- Variance: \(\sigma^2_W = W_0^2 \sigma^2_p\)
Substitute these into the CE equation:
\[CE \approx (W_0 + W_0 \mu_p) - \frac{1}{2} A (W_0^2 \sigma^2_p)\]
Step 7: The Final Optimization
The investor wants to maximize \(CE\). In an optimization problem, we can remove additive constants (\(W_0\)) and divide by positive constants (\(W_0\)) without changing the location of the maximum.
Maximize CE: \[\max \left[ W_0 + W_0 \mu_p - \frac{1}{2} A W_0^2 \sigma^2_p \right]\]
Remove additive constant \(W_0\): \[\max \left[ W_0 \mu_p - \frac{1}{2} A W_0^2 \sigma^2_p \right]\]
Divide by scaling factor \(W_0\): \[\max \left[ \mu_p - \frac{1}{2} (A \cdot W_0) \sigma^2_p \right]\]
Define Relative Risk Aversion (\(\gamma\)): We define \(\gamma = A \cdot W_0\). This is the coefficient of Relative Risk Aversion.
Substitute Vector Notation: \(\mu_p = \mathbf{w}^T \mathbf{\mu}\) and \(\sigma^2_p = \mathbf{w}^T \mathbf{\Sigma} \mathbf{w}\)
We arrive at the final optimization problem:
\[\max_{\mathbf{w}} \quad \mathbf{w}^T \mathbf{\mu} - \frac{\gamma}{2} \mathbf{w}^T \mathbf{\Sigma} \mathbf{w}\]