Financial Engineering and Risk Management

Implementation difficulties with mean-variance

Martin Haugh Garud Iyengar

Columbia University
Industrial Engineering and Operations Research
 Implementation details
Parameter estimation: mean return µ and covariance V
Statistical errors in estimation.
Never have enough data: V has d(d + 1)/2 independent parameters.
Portfolio is very sensitive to estimation errors.

How does one get negative exposures in optimal mean-variance portfolios?

Short the assets: unlimited downside! And often not feasible.
leveraged Exchange Traded Funds (ETFs): Be very careful!

Is variance a good risk measure in practice?

Works for Normal (or elliptic) random variables
Need higher moments to capture deviation from Normality
Tail risk measures

2
 Parameter estimation
The true parameters (µ, V ) are never known!

Use Historical returns r (1) , r (2) , . . ., r (N ) to compute estimates for

PN
mean: µ(est) = N1 t=1 r (t)
PN
covariance matrix: V (est) = N 1−1 t=1 (r (t) − µ(est) )(r (t) − µ(est) )>

15
Each dot is an estimate for
the mean vector with N =
10

true mean
60 months of simulated data.
5
estimated mean

Estimated mean can be quite

µ2

“far” from the true mean.

−5

95% confidence ellipse True mean lies in the 95%

−10
confidence ellipse with prob-
ability 0.95.
−15
−2 −1 0 1 2 3 4 5 6 7 8
µ1

3
 Does parameter error matter?
6.5
Estimated mean and
6 covariance from
N = 60 months of
5.5
simulated data
5
Estimated frontier =
return (%)

4.5
frontier for estimated
parameters
4
Realized frontier = true
3.5
True frontier mean and volatility of
3
Estimated frontier the estimated frontier
Realized frontier portfolios
2.5
2.5 3 3.5 4 4.5 5 5.5 6
volatility (%)

Why is parameter error so serious in portfolio selection?

4
 Many independent samples!
6.5
Estimated mean and
6 covariance from
N = 60 months of
5.5
simulated data
5
Estimated frontier =
return (%)

4.5
frontier for estimated
parameters
4

Realized frontier = true

3.5
mean and volatility of
3
the estimated frontier
portfolios
2.5
2 2.5 3 3.5 4 4.5 5 5.5 6 6.5
volatility (%)

The efficient frontier is extremely unstable.

5
 Why is parameter error so serious?
Two identical assets with mean µ, variance σ 2 , and correlation ρ = 0
Optimal investment x ∗ = (0.5, 0.5)

Suppose estimates
(est)
µ1 = µ +  (positive error)
(est)
µ2 = µ −  (negative error)
Average error in the estimates = 0 ... estimator is good on average

Optimal investment: overweight asset 1. Precisely the wrong thing to do!

Realized performance is lower than expected for the overweighted asset and
better than expected for the underweighted asset! Performance becomes worse as
more shorting is allowed.

Ron Michaud: “Mean-variance results in error-maximizing

investment-irrelevant portfolios.”

6
 Efficient frontier with no-short sales constraint
3.6

3.5

3.4

3.3
return (%)

3.2

3.1

2.9
True frontier
Estimated frontier
2.8
Realized frontier

2.7
3 3.02 3.04 3.06 3.08 3.1 3.12 3.14 3.16 3.18 3.2
volatility (%)

The “corner” in the set {x : x ≥ 0} makes solutions more sensitive to estimates!

7
 Efficient frontier with leverage constraint
30

True frontier
Estimated frontier
25 Realized frontier

20
return (%)

15

10

0
0 5 10 15 20 25 30
volatility (%)

Pd
Leverage constraint: i=1 |xi | ≤ 2
8
 Efficient frontier with robust constraints
25

True frontier
Estimated frontier
Realized frontier
20

15
return (%)

10

0
0 5 10 15 20 25 30
volatility (%)

Let Sm = confidence region

Robust target return constraint: minµ∈Sm µ> x ≥ r.


9
 Methods to improve parameter estimates
Shrinkage methods: James and Stein (1961), Ledoit and Wolf (2004)
 P (est)
(sh) (est) d
“Shrink” to the global mean: µi = αµi + (1 − α) d1 j=1 µj
 P (est)
d
“Shrink” to identity matrix: V (sh) = αV (est) + (1 − α) d
1 2
i=1 σi I

Use subjective views to improve estimates: Black-Litterman method

Non-parametric nearest neighbor-like methods

Let r denote the currently observed return
Let S = {t : kr − rt k ≤ β} = times when historical returns were close to r
Predict that future return belongs to the set {rt+1 : t ∈ S}

10
Financial Engineering and Risk Management
Negative exposures and leveraged ETFs

Martin Haugh Garud Iyengar

Columbia University
Industrial Engineering and Operations Research
 Short positions
Short positions can result in superior returns.

But short positions are very risky!

Long positions have limited downside
The lowest price of an asset is zero.
The largest amount one can lose is the amount invested.

Short positions have unlimited downside

Short positions are created by selling assets borrowed from a broker.
The asset has to be re-purchased and returned to the broker at a later date.
The price of the asset can become arbitrarily large.
Potential loss from the short position can be arbitrarily large

Need a product that has negative exposure and limited liability.

2
 Exchange Traded Funds
Exchange Traded Funds (ETFs) are exchange traded products that track the
returns on stock indices, bond indices, commodities, currencies, etc.

Source: Bloomberg.com

3
 Leveraged ETFs
Produce daily returns that are a multiple of the daily returns on the index.
Bull ETFs: return β × the daily return on index. β = 2, 3
Inverse ETFs: return −β × the daily return on index. β = 1, 2, 3

Source: Bloomberg.com
An inverse ETF is a product that gives negative exposure to the index with
limited liability – all one stands to lose is the initial investment.
4
 Details on the return on an ETF
The return on an ETF is the return on the underlying index compounded daily.
(ETF)
RT = (1 + r1 )(1 + r2 ) · · · (1 + rT )

where rt is the net daily return on the underlying index. Just like a stock.

The return on a leveraged ETF with leverage factor β

(β-ETF)
RT = (1 + βr1 )(1 + βr2 ) · · · (1 + βrT )

where rt is the net daily return on the underlying index.

ETFs are constructed by investing in derivatives written on the index.

The daily compounding has consequences that are not immediately obvious.

5
 Impact of daily compounding
Index: Current value of index is 100.
Day 1: Index goes up to 105. r1 = 5%
Day 2: Index goes back down to 100. r2 = −4.76%

ETF: Current price of an ETF following the index is 100.

Day 1: ETF price goes up to 100(1 + r1 ) = 105
Day 2: ETF price goes down to 105(1 + r2 ) = 100

Bull ETF: Current price of a 2×ETF following the index is 100.

Day 1: ETF price goes up to 100(1 + 2r1 ) = 110
Day 2: ETF price goes down to 105(1 + 2r2 ) = 99.52 ... lose money!

Inverse ETF: Current price of an Inverse ETF following the index is 100.
Day 1: ETF price goes down to 100(1 − r1 ) = 95
Day 2: ETF price goes up to 105(1 − r2 ) = 99.52 ... lose money!
End up in the same place as the 2×ETF!
6
 Volatility and ETF returns
The gross return from a static leveraged position in the underlying index is
βST − (β − 1)S0 (1 + rT )
S0
where r is the funding rate and β is the leverage ratio.

The gross return from investing in the β-Leveraged ETF is approximately

 β Pn−1 Si+1
2
ST (β 2 −β)
(1−β)rT − fT − ln S
e 2 i=0 i
S0
where f is the expense ratio of the fund and n is the number of daily
observations between times 0 and T . (So Sn = ST .)
2
 2
−β) Pn−1 Si+1
The term − (β 2 i=0 ln Si is the loss due to volatility: short vol

ETFs are generally designed for short-term plays on an index or sector, and
should be used that way. Over long-periods leveraged ETFs do not work as one
may expect, especially in volatile markets.
7
Leveraged ETFs during volatile markets
DIG: ProShares Ultra Oil & Gas (2× ETF)
DUG: ProShares UltraShort Oil & Gas (−2× ETF)

8
Financial Engineering and Risk Management
Beyond variance

Martin Haugh Garud Iyengar

Columbia University
Industrial Engineering and Operations Research
 Problems with variance
Appropriate for Normal and other elliptic distributions

Does not capture larger deviations from the mean

Symmetric measure: equally penalizes the deviation above/below mean

2
 Value at Risk
Definition. The value at risk VaRp (L) of random loss L at the confidence level
p ∈ (0, 1) is defined as
VaRp (L) := pth -quantile of L ≈ FL−1 (p)
where FL is the CDF of the random loss L.

0.2
VaR is a “tail” risk mea-
0.18
sure
0.16
95% VaR
0.14 VaRp is increasing in p
0.12

0.1
VaR0.99 (L) ≥ VaR0.95 (L)
0.08

0.06

0.04
Probability = 0.05
0.02

0
0 2 4 6 8 10 12 14 16 18 20
loss

3
 Conditional Value at Risk
Definition. The conditional value at risk CVaRp (L) of random variable L at the
confidence level p ∈ (0, 1) is defined as
R∞
xf (x)dx
VaRp (L) L
CVaRp (L) = E [L | L ≥ VaRp (L)] =
P(L ≥ VaRp (L))
where fL is the density of the random loss L.

0.2
CVaR is also a “tail” risk
0.18
measure
95% VaR
0.16

0.14
CVaRp (L) ≥ VaRp (L)
0.12

95% CVaR
0.1
CVaRp is increasing in p
0.08

0.06
Other names: Tail condi-
0.04
Probability = 0.05 tional expectation and Ex-
0.02
pected Shortfall
0
0 2 4 6 8 10 12 14 16 18 20
loss

4
 VaR and CVaR for Normal distribution
Suppose L ∼ N (µ, σ 2 ).

VaRp (L) = µ + σΦ−1 (p)

where Φ is the CDF of a standard Normal with mean 0 and volatility 1.

 Z 1 
1 −1
CVaRp (L) = µ + σ Φ (β)dβ
1−p p

where Φ is the CDF of a standard Normal with mean 0 and volatility 1.

VaRp and CVaRp for a Normal random variable completely defined by mean µ
and volatility σ. Surprise ?

5
 VaR and CVaR from samples
Suppose L1 , L2 , . . . , LN are N independently and identically distributed (IID)
samples from the Loss L

Let L(1) , L(2) , . . . , L(N ) denote the order-statistics of the N samples

L(1) = smallest value among all samples
L(2) = the next larger value ...
L(N ) = the largest value
L(1) , L(2) , . . . , L(N ) are the samples sorted in increasing order.

Let Kp = dpN e. E.g. p = 0.95 and N = 1000 implies Kp = 950.

VaRp (L) ≈ L(Kp ) = Kp -th term in the sorted samples

N
1 X
CVaRp (L) ≈ L(k)
(1 − p)N
k=Kp

= sum of the largest N − Kp + 1 samples divided by (1 − p)N

6
 Impact of return distribution
Experimental setup:
Computed the Sharpe optimal portfolio for the data in the spreadsheet.
Generated N = 10, 000 samples of loss −rx for three different distributions
(a) Multivariate Normal with mean µ and covariance V
ν−2


(b) Multivariate t with ν = 12 degrees of freedom, mean µ, covariance ν
V,
(c) Mixture of Normals 0.75N (µ, (0.76)2 V ) + 0.25N (µ, (1.5)2 V )

Used samples to estimate VaR and CVaR

The Students’s t distribution has fatter tails than Normal particularly when the
degrees of freedom are small. Expect VaR and CVaR to be larger.

The Normal distribution with covariance matrix (1.5)2 V has all volatilities 50%
higher. The mixture models a situation where there is 25% chance of very high
volatility. Expect VaR and CVaR to be larger.

7
Loss histogram for Normal returns

20

18 95% VaR = 1.67

16

14

12
95% CVaR = 3.15
10

0
−30 −20 −10 0 10 20 30
loss (%)

8
Loss histogram for t returns

35

30

95% VaR = 2.36

25

20
95% CVaR = 4.58

15

10

0
−30 −20 −10 0 10 20 30
loss (%)

9
Loss histogram for mixture of Normals

50

45

40

95% VaR = 1.93

35

30

25

20
95% CVaR = 5.60
15

10

0
−30 −20 −10 0 10 20 30
loss (%)

10
 Pros and Cons of VaR
Pros
Captures the tail behaviour of portfolio losses
Can be robustly estimated from data: not very susceptible to outliers

Cons:
Only sensitive to the p-th quantile and not the distribution beyond
Incentive for “tail stuffing”
VaR is not sub-additive: diversification can increase VaR

11
 Pros and Cons of CVaR
Pros
Captures the tail behaviour of portfolio losses beyond the p-th quantile
Sub-additive: diversification reduces CVaR
Mean-CVaR portfolio selection problems can be solve very efficiently

Cons:
CVaR is defined by an expectation: can be sensitive to outliers

We will return to the topic of VaR and CVaR in the risk management module.

12