Decision Making under Uncertainty

Microeconomics for Finance

Juan F. Imbet

Master in Finance, 1st Year - Paris Dauphine University - PSL

Choosing Between Uncertain Payments

Probability of two states of the world: Good (G) and Bad (B). E.g. 0.5 each. What do you prefer?

Payments, Good state Payments, Bad state
Asset 1 1500 1000
Asset 2 2000 500
Asset 3 2000 1000

Choosing Between Uncertain Payments

Asset 3 payments state by state dominate those of assets 1 and 2.

⇒ Obviously, 3 is preferred

Asset 1 vs. Asset 2 is not obvious: more in state B but less in state G.

Similar problem. Which basket the consumer does prefer: 2 apples and 3 bananas or 3 apples and 2 bananas?

⇒ We must quantify the agents’ preferences ideally with a utility function.

Lotteries

A risky payment (or lottery) is a random variable \(\tilde{a}\) with random realizations in each state of nature \(a_1, a_2, \ldots, a_N\) with probabilities \(\pi_1, \pi_2, \ldots, \pi_N\), such that \(\sum_{n=1}^N \pi_n = 1\).

Examples include:

  • Cash flows from an investment project depending on the state of the economy.
  • Returns from a financial asset depending on market conditions.
  • Insurance payoffs depending on whether an accident occurs or not.
  • Derivatives payoffs depending on the underlying asset price at maturity.

Preferences over Lotteries

We want to define a relation of order (i.e, a preference) over the set of risky payments (lotteries). Denote \(\tilde{a}\) and \(\tilde{b}\) two risky payments (lotteries).

  • \(\tilde{a}\) is preferred to \(\tilde{b}\), denoted \(\tilde{a} \succ \tilde{b}\), if the agent prefers \(\tilde{a}\) to \(\tilde{b}\).
  • \(\tilde{a}\) is indifferent to \(\tilde{b}\), denoted \(\tilde{a} \sim \tilde{b}\), if the agent is indifferent between \(\tilde{a}\) and \(\tilde{b}\).
  • \(\tilde{a}\) is not preferred to \(\tilde{b}\), denoted \(\tilde{a} \preceq \tilde{b}\), if the agent does not prefer \(\tilde{a}\) to \(\tilde{b}\).

Rational Preferences

We say that the relation of order \(\succeq\) is rational if it satisfies:

  1. Completeness: For any lotteries \(\tilde{a}, \tilde{b}\), either \(\tilde{a} \succeq \tilde{b}\) or \(\tilde{b} \succeq \tilde{a}\) (or both).
  2. Transitivity: For any lotteries \(\tilde{a}, \tilde{b}, \tilde{c}\), if \(\tilde{a} \succeq \tilde{b}\) and \(\tilde{b} \succeq \tilde{c}\), then \(\tilde{a} \succeq \tilde{c}\).
  3. Continuity: For any lotteries \(\tilde{a}, \tilde{b}, \tilde{c}\), if \(\tilde{a} \succeq \tilde{b} \succeq \tilde{c}\), then there exists \(\alpha \in [0,1]\) such that \(\alpha \tilde{a} + (1-\alpha) \tilde{c} \sim \tilde{b}\).
  4. Independence: For any lotteries \(\tilde{a}, \tilde{b}, \tilde{c}\) and any \(\alpha \in (0,1)\), if \(\tilde{a} \succeq \tilde{b}\), then \(\alpha \tilde{a} + (1-\alpha) \tilde{c} \succeq \alpha \tilde{b} + (1-\alpha) \tilde{c}\).

Utility Function

  • If the relation of order \(\succeq\) is rational, then there exists a utility function \(U\) such that, for any lotteries \(\tilde{a}, \tilde{b}\),

    \[\tilde{a} \succeq \tilde{b} \Leftrightarrow U(\tilde{a}) \geq U(\tilde{b})\]

  • The utility function is not unique. If \(V\) is another utility function representing the same preferences, then there exists an increasing transformation \(f\) such that

    \[V(\tilde{a}) = f(U(\tilde{a}))\]

Properties of the Utility Function

We want to determine a utility function, \(U\), “reasonable”, which generates a relation of order (i.e, a preference) over the set of risky payments (lotteries).

Back to our example, the utility function should depend positively on the payment realization in each state.

If \(\tilde{a}\) is a risky payment with realizations \(a^{G}\), \(a^{B}\), then \[U(\tilde{a}) = U(a^{G}, a^{B}) \quad \text{with} \quad \frac{\partial U}{\partial a^{G}} > 0, \quad \frac{\partial U}{\partial a^{B}} > 0\]

Properties of the Utility Function

How to take into account the probabilities of each state of the world?

Payments, Good state Payments, Bad state
Asset 1 1500 1000
Asset 2 2000 500
  • If \(\pi^{G} = 0.99\) and \(\pi^{B} = 0.01\), we should prefer asset 2.

  • Denote u(x) the utility obtained from a certain payment x

  • → a good “candidate” is an expected utility function

    \[U(\tilde{a}) = \pi^G u(a^G) + \pi^B u(a^B) = \mathbb{E}[u(\tilde{a})]\]

Properties of the Utility Function

Payments, Good state Payments, Bad state
Asset 1 1500 1000
Asset 2 2000 500
  • If \(\pi^{G} = \pi^{B} = 0.5\), both assets have same mean 1250. Which one do you prefer?

  • If you are risk averse, \(\mathbb{E}[u(\tilde{a}_1)] > \mathbb{E}[u(\tilde{a}_2)].\)

  • → If we still consider the “candidate” \[\mathbb{E}[u(\tilde{a})] = \pi^G u(a^G) + \pi^B u(a^B),\] Which condition on u?

Expected Utility

John von Neumann and Oskar Morgenstern have determined conditions under which agents’ preferences for risky payments (lotteries) can be described by an expected utility function such that

\[\mathbb{E}[u(\tilde{a})] = \sum_{n=1}^N \pi_n u(a_n)\]

with \(u\) concave and increasing, \(u' > 0\) and \(u'' < 0\).

More generally, if a risky payment \(\tilde{a}\) has a distribution \(d\Phi\),

\[\mathbb{E}[u(\tilde{a})] = \int_{\mathbb{R}} u(a) \, d\Phi(a)\]

  • The function \(u\) is called the von Neumann-Morgenstern utility function. It is unique up to an increasing affine transformation. E.g. \(u\) and \(v = \alpha u + \beta\), with \(\alpha > 0\) and \(\beta \in \mathbb{R}\), represent the same preferences.

Risk Aversion

Under these conditions, agents are risk averse. In particular, with Jensen inequality for a concave function, we have

\[\mathbb{E}[u(\tilde{a})] \leq u(\mathbb{E}[\tilde{a}]).\]

Agents would prefer to obtain the mean of payment with certainty rather than the risky payment.

How to quantify risk aversion?

The level of concavity, measured by the second derivative, \(u''\), is not a good measure. Indeed, if \(v = \alpha u + \beta\), with \(\alpha > 1\)

\(v\) generates the same relation of order among risky payments than u does. Yet \(v'' < u'' < 0\).

⇒ A measure of risk aversion must be invariant with any affine transformation of the utility function.

Measuring Risk Aversion

There are two classic measures of risk aversion

\[R_A(y) = -\frac{u''(y)}{u'(y)} = \text{coefficient of absolute risk aversion}\]

and

\[R_R(y) = -y \frac{u''(y)}{u'(y)} = \text{coefficient of relative risk aversion}\]

where \(y\) is the payment (the income) of the agent.

Both are positive: \(R_A(y) > 0, R_R(y) > 0\).

Interpreting Risk Aversion

Consider a risk averse agent who receives a risky payment \(y + \tilde{\epsilon}\), with mean \(y\) and variance \(\mathrm{Var}[\tilde{\epsilon}] = \sigma^{2}\).

The certainty equivalent of this payment is \(y - \Delta\) such that

\[u(y - \Delta) = \mathbb{E}[u(y + \tilde{\epsilon})]\]

\(\Delta\) is a risk premium: How much the agent is willing to reduce his mean payment to get rid of uncertainty.

Interpreting Risk Aversion

Assuming that \(\Delta\) and \(\epsilon\) are small, \(\Delta \ll 1\) and \(\sigma^{2} \ll 1\), give an approximation of \(\Delta\).

\[u(y - \Delta) \approx u(y) - \Delta u'(y)\] \[\mathbb{E}[u(y + \tilde{\epsilon})] \approx \mathbb{E}[u(y)] + \mathbb{E}[\tilde{\epsilon} u'(y)] + \frac{1}{2}\mathbb{E}[\tilde{\epsilon}^2 u''(y)] = u(y) + \frac{\sigma^2}{2} u''(y)\]

\[\Rightarrow \Delta = -\frac{u''(y)}{u'(y)} \cdot \frac{\sigma^2}{2} = R_A(y) \cdot \frac{\sigma^2}{2}\]

Interpreting Risk Aversion

When the risky payment is \(y + y \tilde{\epsilon}\), define \(\delta\) as \[u(y - \delta y) = \mathbb{E}[u(y + y \tilde{\epsilon})].\] Similarly, one can give a first order approximation of \(\delta\) under the assumption \(\delta << 1\) and \(\sigma^{2} << 1\).

\[u(y - \delta y) \approx u(y) - \delta y u'(y)\] \[\mathbb{E}[u(y + y \tilde{\epsilon})] \approx u(y) + \frac{y^2 \sigma^2}{2} u''(y)\]

\[\Rightarrow \delta = -\frac{y u''(y)}{u'(y)} \cdot \frac{\sigma^2}{2} = R_R(y) \cdot \frac{\sigma^2}{2}\]

Whether uncertainty is absolute or relative (w.r.t \(y\)), the certainty equivalent depends on the coefficient of absolute or relative risk aversion.

Usual Utility Functions

  • CARA (“Constant Absolute Risk Aversion”)

    \[ u(y) = -\frac{1}{\nu} e^{-\nu y} \]

  • CRRA (“Constant Relative Risk Aversion”)

    \[ u(y) = \frac{y^{1-\gamma}}{1-\gamma} \]

    and, in addition, \[u(y) = \ln(y)\] for \(\gamma = 1\).

Risk Neutrality

When \(u\) is affine, the agent is risk neutral:

\[ u(y) = \alpha y + \beta \]

He only cares about his expected revenue.

Specifically,

\[ R_A(y) = R_R(y) = 0 \quad \forall y \]

Risk Aversion and Portfolio Choice

Consider an agent with an expected utility function à la von Neumann-Morgenstern, who invests at \(t=0\) and, at \(t=1\), receives the pay-off of its investment and “consume” it (get the utility out of the payment).

The agent can invest his initial wealth \(y_0\) in the following way:

  • invests an amount \(w\) in a risky asset (the S&P 500 for instance) with random net return \(\tilde{r}\), and the rest in the risk free asset with (certain) net return \(r_f\).

His final wealth \(\tilde{y}_1\) at \(t=1\): \[ \tilde{y}_1 = (y_0 - w) \times (1 + r_f) + w \times (1 + \tilde{r})= y_0 \times (1 + r_f) + w \times (\tilde{r} - r_f) \]

Risk Aversion and Portfolio Choice

Objective function of the agent:

\[ \max_w U(w) = \max_w \mathbb{E}[u(\tilde{y}_1)] = \max_w \mathbb{E}[u(y_0 (1 + r_f) + w (\tilde{r} - r_f))] \]

Unconstrained optimization of a concave function,

\[ u'' < 0 \Rightarrow U''(w) = \mathbb{E}[(\tilde{r} - r_f)^2 u''(y_0 (1 + r_f) + w (\tilde{r} - r_f))] < 0 \]

The optimal amount \(w*\) is given by the first order condition:

\[ U'(w) = 0 \]

Risk Aversion and Portfolio Choice

If the investor is risk neutral:

\[u(y) = \alpha y + \beta\]

\[\max_w \mathbb{E}[u(\tilde{y}_1)] \Leftrightarrow \max_w y_0 (1 + r_f) + w (\mathbb{E}[\tilde{r}] - r_f).\]

Then,

\[\mathbb{E}[\tilde{r}] > r_f \Rightarrow w^* = +\infty\] \[\mathbb{E}[\tilde{r}] < r_f \Rightarrow w^* = -\infty\]

Risk Aversion and Portfolio Choice

Same problem, different formulation: The agent buys z units of risky assets, at price p, with pay-off at t=1 equal to \(\tilde{a}\).

\[\tilde{y}_1 = (y_0 - p \cdot z) \times (1 + r_f) + z \times \tilde{a}\] \[= y_0 \times (1 + r_f) + z \times (\tilde{a} - (1 + r_f)p)\]

First order condition: \[\mathbb{E}[(\tilde{a} - (1 + r_f)p) u'(y_0 (1 + r_f) + z^* (\tilde{a} - (1 + r_f)p))] = 0\]

Let’s approximate the function around \(z=0\).

\[ \begin{aligned} U'(z) & = \mathbb{E}[(\tilde{a} - (1 + r_f)p) u'(y_0 (1 + r_f))] \\ & + z \mathbb{E}[(\tilde{a} - (1 + r_f)p)^2 u''(y_0 (1 + r_f))] + o(z) \end{aligned} \]

The first order condition is only approximately zero at \(z=0\) if the first term is zero, i.e. \[\mathbb{E}[(\tilde{a} - (1 + r_f)p)] = 0 \Leftrightarrow \frac{\mathbb{E}[\tilde{a}]}{1+r_f} = p\]

  • If the investor gets in expectation the risk-free return, he does not invest in the risky asset: \(z^* = 0\).

Exercise 1: Certainty Equivalent (Basic)

An investor has a utility function \(u(y) = \sqrt{y}\) and faces a lottery: - with probability 0.5, she receives €100; - with probability 0.5, she receives €400.

Question:
Compute her expected utility and certainty equivalent (CE).

Solution

\[ \mathbb{E}[u(\tilde{y})] = 0.5\sqrt{100} + 0.5\sqrt{400} = 0.5(10 + 20) = 15 \]

Certainty equivalent is found by \(u(CE) = 15\):

\[ \sqrt{CE} = 15 \Rightarrow CE = 225 \]

Expected value = €250 → Risk premium = €25.

Exercise 2: Certainty Equivalent (with CRRA Utility)

Let \(u(y) = \frac{y^{1-\gamma}}{1-\gamma}\) with \(\gamma = 2\) and a lottery paying: - €80 with probability 0.6
- €200 with probability 0.4

Question:
Compute the certainty equivalent and risk premium.

Solution

\[ \mathbb{E}[u(\tilde{y})] = 0.6\frac{80^{-1}}{-1} + 0.4\frac{200^{-1}}{-1} = 0.6(-\tfrac{1}{80}) + 0.4(-\tfrac{1}{200}) = -0.0095 \]

\[ u(CE) = -0.0095 \Rightarrow \frac{CE^{-1}}{-1} = -0.0095 \Rightarrow CE = 105.26 \]

Expected value = €128 → Risk premium = €22.74.

Exercise 3: Portfolio Choice (Analytical)

Initial wealth \(y_0 = 100\).
Risk-free rate \(r_f = 5\%\).
Risky asset return \(\tilde{r}\) takes:

  • \(r_G = 20\%\) with probability 0.5
  • \(r_B = -10\%\) with probability 0.5

Utility: \(u(y) = \ln(y)\).

Question:
Find the optimal risky investment \(w^*\).

Solution

Final wealth:

\[ \tilde{y}_1 = 100(1.05) + w(\tilde{r} - 0.05) \]

Expected utility:

\[ U(w) = 0.5\ln(105 + 0.15w) + 0.5\ln(105 - 0.15w) \]

First-order condition:

\[ U'(w) = 0.5\frac{0.15}{105 + 0.15w} - 0.5\frac{0.15}{105 - 0.15w} = 0 \]

\[ \Rightarrow 105 + 0.15w = 105 - 0.15w \Rightarrow w^* = 0 \]

→ With symmetric returns around the risk-free, optimal exposure = 0.

Exercise 4: Portfolio Choice (Approximation)

Same setup as before, but now \(\mathbb{E}[\tilde{r}] = 0.10\), \(\sigma = 0.20\), \(r_f = 0.05\), and \(u(y)=\ln(y)\).

Question:
Approximate the optimal risky share \(w^*\) using a mean–variance approximation.

Solution

  • Approximate \(u(y)\) around \(\mathbb{E}[\tilde{y}_1]\):

\[ u(y) \approx u(\mathbb{E}[\tilde{y}_1]) + u'(\mathbb{E}[\tilde{y}_1])(y - \mathbb{E}[\tilde{y}_1]) + \frac{1}{2}u''(\mathbb{E}[\tilde{y}_1])(y - \mathbb{E}[\tilde{y}_1])^2 \] - Expected utility: \[ \mathbb{E}[u(\tilde{y}_1)] \approx u(\mathbb{E}[\tilde{y}_1]) + \frac{1}{2}u''(\mathbb{E}[\tilde{y}_1])\mathrm{Var}[\tilde{y}_1] \]

  • Final wealth: \[ \tilde{y}_1 = 100(1.05) + w(\tilde{r} - 0.05) \]
  • Mean and variance: \[ \mathbb{E}[\tilde{y}_1] = 105 + 0.05w, \quad \mathrm{Var}[\tilde{y}_1] = w^2 \sigma^2 \]

  • Objective function: \[ \max_w U(w) \approx \max_w \ln(105 + 0.05w) - \frac{1}{2}\frac{1}{(105 + 0.05w)^2} w^2 \sigma^2 \]

  • First-order condition: \[ U'(w) = \frac{0.05}{105 + 0.05w} - \frac{w \sigma^2}{(105 + 0.05w)^2} = 0 \]

  • Solution: \[ \Rightarrow w^* \approx \frac{0.05(105 + 0.05w^*)}{\sigma^2} \approx \frac{0.05 \times 105}{0.20^2} = 131.25 \]

Exercise 5 (Advanced): Optimal Insurance under Risk Aversion

An individual has initial wealth \(y=100\) and faces a loss \(\tilde{L}\):

  • €0 with probability 0.9
  • €40 with probability 0.1

They can buy insurance coverage \(I\) that fully compensates the loss if it occurs, at a premium \(\pi\) per unit of coverage.
Utility: \(u(y) = \ln(y)\).

Questions: 1. Write expected utility as a function of \(I\).
2. For which premium \(\pi\) is full insurance (\(I=40\)) optimal?

Solution

Expected utility:

\[ U(I) = 0.9\ln(y - \pi I) + 0.1\ln(y - \pi I - L + I) \]

Differentiate:

\[ U'(I) = 0.9\frac{-\pi}{y - \pi I} + 0.1\frac{-\pi + 1}{y - \pi I - L + I} = 0 \]

At full insurance \(I=L=40\):

\[ 0.9\frac{-\pi}{y - \pi L} + 0.1\frac{-\pi + 1}{y - \pi L} = 0 \Rightarrow -\pi(0.9 + 0.1) + 0.1 = 0 \Rightarrow \pi = 0.1 \]

→ The actuarially fair premium (\(\pi = \text{probability of loss}\)) makes full insurance optimal.