Understanding Risk-Aversion through Utility Theory

Ashwin Rao

ICME, Stanford University

February 3, 2020

Ashwin Rao (Stanford) Utility Theory February 3, 2020 1 / 14

Intuition on Risk-Aversion and Risk-Premium

Let’s play a game where your payoff is based on outcome of a fair coin
You get $100 for HEAD and $0 for TAIL
How much would you pay to play this game?
You immediately say: “Of course, $50”
Then you think a bit, and say: “A little less than $50”
Less because you want to “be compensated for taking the risk”
The word Risk refers to the degree of variation of the outcome
We call this risk-compensation as Risk-Premium
Our personality-based degree of risk fear is known as Risk-Aversion
So, we end up paying $50 minus Risk-Premium to play the game
Risk-Premium grows with Outcome-Variance & Risk-Aversion

Ashwin Rao (Stanford) Utility Theory February 3, 2020 2 / 14

Specifying Risk-Aversion through a Utility function

We seek a “valuation formula” for the amount we’d pay that:

Increases one-to-one with the Mean of the outcome
Decreases as the Variance of the outcome (i.e.. Risk) increases
Decreases as our Personal Risk-Aversion increases
The last two properties above define the Risk-Premium
But fundamentally why are we Risk-Averse?
Why don’t we just pay the mean of the random outcome?
Reason: Our satisfaction to better outcomes grows non-linearly
We express this satisfaction non-linearity as a mathematical function
Based on a core economic concept called Utility of Consumption
We will illustrate this concept with a real-life example

Ashwin Rao (Stanford) Utility Theory February 3, 2020 3 / 14

Law of Diminishing Marginal Utility

Ashwin Rao (Stanford) Utility Theory February 3, 2020 4 / 14

Utility of Consumption and Certainty-Equivalent Value

Marginal Satisfaction of eating cookies is a diminishing function

Hence, Accumulated Satisfaction is a concave function
Accumulated Satisfaction represents Utility of Consumption U(x)
Where x represents the uncertain outcome being consumed
Degree of concavity represents extent of our Risk-Aversion
Concave U(·) function ⇒ E[U(x)] < U(E[x])
We define Certainty-Equivalent Value xCE = U −1 (E[U(x)])
Denotes certain amount we’d pay to consume an uncertain outcome
Absolute Risk-Premium πA = E[x] − xCE
πA E[x]−xCE xCE
Relative Risk-Premium πR = E[x] = E[x] =1− E[x]

Ashwin Rao (Stanford) Utility Theory February 3, 2020 5 / 14

Certainty-Equivalent Value

Ashwin Rao (Stanford) Utility Theory February 3, 2020 6 / 14

Calculating the Risk-Premium
We develop mathematical formalism to calculate Risk-Premia πA , πR
To lighten notation, we refer to E[x] as x̄ and Variance of x as σx2
Taylor-expand U(x) around x̄, ignoring terms beyond quadratic
1
U(x) ≈ U(x̄) + U 0 (x̄) · (x − x̄) + U 00 (x̄) · (x − x̄)2
2
Taylor-expand U(xCE ) around x̄, ignoring terms beyond linear
U(xCE ) ≈ U(x̄) + U 0 (x̄) · (xCE − x̄)
Taking the expectation of the U(x) expansion, we get:
1
· U 00 (x̄) · σx2
E[U(x)] ≈ U(x̄) +
2
Since E[U(x)] = U(xCE ), the above two expressions are ≈. Hence,
1
U 0 (x̄) · (xCE − x̄) ≈ · U 00 (x̄) · σx2
2

Ashwin Rao (Stanford) Utility Theory February 3, 2020 7 / 14

Absolute & Relative Risk-Aversion
From the last equation on the previous slide, Absolute Risk-Premium
1 U 00 (x̄) 2
πA = x̄ − xCE ≈ − · 0 · σx
2 U (x̄)
00
We refer to function A(x) = − UU 0 (x)
(x)
as the Absolute Risk-Aversion
1
· A(x̄) · σx2
πA ≈
2
In multiplicative uncertainty settings, we focus on variance σ 2x of xx̄
x̄
In multiplicative settings, we also focus on Relative Risk-Premium πR
πA 1 U 00 (x̄) · x̄ σx2 1 U 00 (x̄) · x̄
πR = ≈− · · = − · · σ 2x
x̄ 2 U 0 (x̄) x̄ 2 2 U 0 (x̄) x̄

00
We refer to function R(x) = − UU 0(x)·x
(x) as the Relative Risk-Aversion
1
πR ≈ · R(x̄) · σ 2x
2 x̄

Ashwin Rao (Stanford) Utility Theory February 3, 2020 8 / 14

Taking stock of what we’re learning here

We’ve shown that Risk-Premium can be expressed as the product of:

Extent of Risk-Aversion: either A(x̄) or R(x̄)
Extent of uncertainty of outcome: either σx2 or σ 2x
x̄

We’ve expressed the extent of Risk-Aversion as the ratio of:

Concavity of the Utility function (at x̄): −U 00 (x̄)
Slope of the Utility function (at x̄): U 0 (x̄)
For optimization problems, we ought to maximize E[U(x)] (not E[x])
Linear Utility function U(x) = a + b · x implies Risk-Neutrality
Now we look at typically-used Utility functions U(·) with:
Constant Absolute Risk-Aversion (CARA)
Constant Relative Risk-Aversion (CRRA)

Ashwin Rao (Stanford) Utility Theory February 3, 2020 9 / 14

Constant Absolute Risk-Aversion (CARA)
1−e −ax
Consider the Utility function U(x) = a for a 6= 0
−U 00 (x)
Absolute Risk-Aversion A(x) = U 0 (x) = a
a is called Coefficient of Constant Absolute Risk-Aversion (CARA)
For a = 0, U(x) = x (meaning Risk-Neutral)
If the random outcome x ∼ N (µ, σ 2 ),

 1−e −aµ+ a22σ2
E[U(x)] = a for a 6= 0
µ for a = 0

aσ 2
xCE = µ −
2
aσ 2
Absolute Risk Premium πA = µ − xCE =
2
2
For optimization problems where σ is a function of µ, we seek the
2
distribution that maximizes µ − aσ2
Ashwin Rao (Stanford) Utility Theory February 3, 2020 10 / 14
A Portfolio Application of CARA
We are given $1 to invest and hold for a horizon of 1 year
Investment choices are 1 risky asset and 1 riskless asset
Risky Asset Annual Return ∼ N (µ, σ 2 )
Riskless Asset Annual Return = r
Determine unconstrained π to allocate to risky asset (1 − π to riskless)
Such that Portfolio has maximum Utility of Wealth in 1 year
−aW
With CARA Utility U(W ) = 1−ea for a 6= 0
Portfolio Wealth W ∼ N (1 + r + π(µ − r ), π 2 σ 2 )
From the section on CARA Utility, we know we need to maximize:
aπ 2 σ 2
1 + r + π(µ − r ) −
2
So optimal investment fraction in risky asset
µ−r
π∗ =
aσ 2

Ashwin Rao (Stanford) Utility Theory February 3, 2020 11 / 14

Constant Relative Risk-Aversion (CRRA)
x 1−γ −1
Consider the Utility function U(x) = 1−γ for γ 6= 1
−U 00 (x)·x
Relative Risk-Aversion R(x) = U 0 (x) =γ
γ is called Coefficient of Constant Relative Risk-Aversion (CRRA)
For γ = 1, U(x) = log(x). For γ = 0, U(x) = x − 1 (Risk-Neutral)
If the random outcome x is lognormal, with log(x) ∼ N (µ, σ 2 ),

 e µ(1−γ)+ σ22 (1−γ)2 −1
E[U(x)] = 1−γ for γ 6= 1
µ for γ = 1
σ2
xCE = e µ+ 2
(1−γ)

xCE σ2 γ
Relative Risk Premium πR = 1 − = 1 − e− 2
x̄
2
For optimization problems where σ is a function of µ, we seek the
2
distribution that maximizes µ + σ2 (1 − γ)
Ashwin Rao (Stanford) Utility Theory February 3, 2020 12 / 14
A Portfolio Application of CRRA (Merton 1969)

We work in the setting of Merton’s 1969 Portfolio problem

We only consider the single-period (static) problem with 1 risky asset
Riskless asset: dRt = r · Rt · dt
Risky asset: dSt = µ · St · dt + σ · St · dzt (i.e. Geometric Brownian)
We are given $1 to invest, with continuous rebalancing for 1 year
Determine constant fraction π of Wt to allocate to risky asset
To maximize Expected Utility of Wealth W = W1 (at time t = 1)
Constraint: Portfolio is continuously rebalanced to maintain fraction π
So, the process for wealth Wt is given by:

dWt = (r + π(µ − r )) · Wt · dt + π · σ · Wt · dzt

W 1−γ −1
Assume CRRA Utility U(W ) = 1−γ , 0 < γ 6= 1

Ashwin Rao (Stanford) Utility Theory February 3, 2020 13 / 14

Recovering Merton’s solution (for this static case)
Applying Ito’s Lemma on log Wt gives us:
Z t Z t
π2σ2
log Wt = (r + π(µ − r ) − ) · du + π · σ · dzu
0 2 0

π2σ2 2 2
⇒ log W ∼ N (r + π(µ − r ) −,π σ )
2
From the section on CRRA Utility, we know we need to maximize:
π 2 σ 2 π 2 σ 2 (1 − γ)
r + π(µ − r ) − +
2 2

π2σ2γ
= r + π(µ − r ) −
2
So optimal investment fraction in risky asset
µ−r
π∗ =
γσ 2
Ashwin Rao (Stanford) Utility Theory February 3, 2020 14 / 14