PORTFOLIO

CONSTRUCTION
THEORY
Laurent E. Calvet
Topic 1
 The role of portfolio management

Portfolio management consists of managing one’s (individual, company,

institution) investments, by allocating wealth to investable assets (stocks,
bonds, cash, mutual funds, commodities, real estate, etc.), in order to
achieve the “best” outcome within a stipulated time frame.

Why is portfolio management useful to individuals/customers?

• Individuals invest now for a future date.
• If we knew the future, it would be an easy task, but most financial
investments entail uncertainty.

What is the task of the portfolio manager?

• Optimize the investor’s portfolio, based on available information.

15/10/2022 2
 Portfolio management in 3 steps

Step 1: Collect data and model future returns

• Ideally, we would like to know future asset returns with certainty. In

practice, however, future returns are random.

• We infer the probability distribution of future returns from past

returns by computing invariant quantities, which will hopefully behave
similarly in the future.

• Asset pricing models (e.g., the CAPM) are useful to infer future returns.

15/10/2022 3
 Portfolio management in 3 steps

Step 2: Optimize the portfolio

• Optimal portfolio allocation determined according to investment
objectives and distribution of future returns.
• Portfolio models & quantitative investment strategies are widely
implemented in the industry.

Step 3: Evaluate portfolio performance

• Performance measures, such as the average portfolio return.
• Risk measures, such as portfolio volatility.
• Risk-adjusted performance: Sharpe ratio.

15/10/2022 4
 Course outline
Portfolio choice
• Markowitz model

Asset pricing models

• CAPM: theory and empirical tests
• Arbitrage Pricing Theory (APT)
• Choice of factors

The Efficient Market Hypothesis

• Weak, semi-strong and strong forms
• Empirical challenges

Performance evaluation

15/10/2022 5
 Schedule
• Lecture 1: Sept 5-6, 2022

• Lecture 2: Sept 12-13, 2022

• Lecture 3: Sept 26-27, 2022

• Lecture 4: Oct 3-4, 2022

• Lecture 5: Oct 17-18, 2022

+ 3 Tutorials

15/10/2022 6
 Instructors

• Laurent E. Calvet, Professor of Finance

[email protected]

• Xuan Feng, PhD student

[email protected]

• Alessandro Gastaldello, PhD student

[email protected]

15/10/2022 7
 Grading and references

Grading
• 70% final exam (closed-book exam)
• 30% tutorials

References (not required)

• Bodie Kane and Marcus, Investments.
• Elton, Gruber, Brown, and Goetzmann, Portfolio Theory and Investment
Analysis.

15/10/2022 8
 Topic 1: Mean-variance analysis

Fundamentals of financial analysis

• Prices and returns
• Mean and volatility of returns
• Returns, mean and volatility of portfolios

Markowitz model: the mean-variance frontier

• Two-asset case
• Multi-asset case
• Implementation issues

15/10/2022 9
 Asset returns

Consider a dividend-paying stock and a time series of 𝑇 + 1 (closing)

prices, 𝑃0, 𝑃1, 𝑃2, 𝑃3, … , 𝑃! .

The corresponding dividend payments between time 𝑡 − 1 and time 𝑡 are

denoted by 𝐷𝑖𝑣 "#$;" .

The total return accounts for capital gains (return attributed to price
movements) and income (dividend payments):

𝑃" + 𝐷𝑖𝑣 "#$;" − 𝑃"#$ 𝑃" + 𝐷𝑖𝑣 "#$;"

𝑅" = = −1
𝑃"#$ 𝑃"#$

15/10/2022 10
 Adjusted prices

Working with time series of dividends, along with the time series of prices,
can be inconvenient. For this reason, some data providers (for example
Yahoo! Finance) also report the time series of adjusted prices.

Adjusted prices (𝑨𝒅𝒋𝑷𝒕 ) are fictitious and account for dividend payments.
They are defined by:
𝐴𝑑𝑗𝑃" 𝑃" + 𝐷𝑖𝑣 "#$;"
=
𝐴𝑑𝑗𝑃"#$ 𝑃"#$

Total returns are easily computed from adjusted prices:

𝑃" + 𝐷𝑖𝑣 "#$;" − 𝑃"#$ 𝐴𝑑𝑗𝑃"

𝑅" = = −1
𝑃"#$ 𝐴𝑑𝑗𝑃"#$

15/10/2022 11
 Example

Closing adjusted prices of a stock:

AdjPt = 4.51
AdjPt-1 = 4.50

The total return is:

Rt = AdjPt / AdjPt-1 – 1 = 4.51 / 4.50 – 1 = 0.2222%

15/10/2022 12
 Mean return
• We observe the historical returns 𝑅$ , 𝑅' ,…, 𝑅! of an asset.
• We assume that the expected return (that is, the population
mean of the return distribution) is time invariant.
• The expected value on a future single-period return, 𝑅!($ ,
can be estimated by the arithmetic average of historical
returns:

!
1
𝑅5 = 6 𝑅" → E 𝑅$
𝑇
")$

where E denotes the population mean.

15/10/2022 13
 Mean return (cont.)
The arithmetic mean of arithmetic returns is often annualized, by multiplying
the return average with the frequency (number of periods / year) of the data:
(
𝑓
𝐴𝑛𝑛𝑅 = ' 𝑅%
𝑇
%&'
Typical values for 𝑓 are:
• Daily data: 𝑓 = 252 (average number of US trading days in a calendar year: 52
weeks x 5 trading days/week - 8 public holidays)
• Weekly data: 𝑓 = 52
• Monthly data: 𝑓 = 12
• Annual data: 𝑓 = 1 (no annualization needed!)

This method provides a first-order approximation of the expected 1-year

return.
15/10/2022 14
 Return variance and volatility
• Returns are random, which is a source of risk to investors. Variance and
volatility are measures of risk.

• Knowing that the returns of a security have the value 𝑅* with a probability
𝑝* (𝑖 = 1, … , 𝑁), the variance of returns is given by:
+
𝜎 ) 𝑅 = Var 𝑅 = E 𝑅 − E 𝑅 ) = ' 𝑝* 𝑅* − E 𝑅 ),

*&'
where:
+
E 𝑅 = ' 𝑝* 𝑅* .
*&'

• The standard deviation (or volatility) is the square root of the variance:
𝜎 𝑅 = Std 𝑅 = Var 𝑅
15/10/2022 15
 Historical volatility
Given the time series of historical returns, 𝑅" ( 𝑡 = 1, … , 𝑇 ), an
estimator of volatility, called historical volatility, is given by the
sample standard deviation of historical returns:

!
1 '
𝜎; 𝑅" = 6 𝑅" − 𝑅"
𝑇−1
")$

where:
!
1
𝑅" = 6 𝑅"
𝑇
")$

15/10/2022 16
 Historical volatility: Annualization
• Historical volatility is often annualized.

• We multiply the historical volatility of single-period returns 𝜎; 𝑅" by

the square root of the frequency of the data:

𝐴𝑛𝑛𝜎7 𝑅% = 𝜎7 𝑅% 8 𝑓

• Why the square root of f ?

15/10/2022 17
 Return on

a portfolio of assets

15/10/2022 18
 Covariance and correlation of returns
• When building portfolios, we are not just interested in the distribution of the returns
of each individual asset. We need information on the joint distribution of returns.

• Assume that the returns on two assets, 𝑅' and 𝑅) , take the joint values 𝑅',* , 𝑅),*
with a probability 𝑝* (𝑖 = 1, … , 𝑁).

• The covariance is defined by:

%

𝜎!" = Cov 𝑅! , 𝑅" = E 𝑅! − E 𝑅! 𝑅" − E 𝑅" = * 𝑝# 𝑅!,# − E 𝑅! 𝑅",# − E 𝑅"

#$!

• The correlation is given by:

Cov 𝑅! , 𝑅"
𝜌!" = Corr 𝑅! , 𝑅" =
𝜎 𝑅! 𝜎 𝑅"
Note that −1 ≤ Corr 𝑅' , 𝑅) ≤ 1.

15/10/2022 19
 Historical covariance and correlation
• We observe the historical returns 𝑅$," and 𝑅'," (𝑡 = 1, … , 𝑇 ) on two
assets.

• The sample covariance,

'
𝜎7') = ∑(%&' 𝑅',% − 𝑅' 𝑅),% − 𝑅) ,
(-'

is an unbiased estimator of the covariance.

• The historical covariance is often annualized as follows:

𝐴𝑛𝑛𝜎7') = 𝜎7') 8 𝑓

• The historical correlation is estimated by the sample correlation:

𝜎7')
𝜌7') =
𝜎7' 𝜎7)

15/10/2022 20
 Variance-covariance matrix

• Asset covariances are often represented in matrix form.

• For example, for 3 assets, the covariance matrix 𝚺 is defined by:

𝜎$' 𝜎$' 𝜎$,

𝚺 = 𝜎'$ 𝜎'' 𝜎',
𝜎,$ 𝜎,' 𝜎,'

15/10/2022 21
 Return on a portfolio of assets
Consider a portfolio 𝑃 of 𝑁 assets:

• The portfolio weights associated to each asset are

𝑤',%-' , 𝑤),%-' ,…, 𝑤+,%-'
at the end of period 𝑡 − 1.

• The returns on the assets are

𝑅',% , 𝑅),% ,…, 𝑅+,%
between 𝑡 − 1 and 𝑡.

• The return on portfolio 𝑃 between dates 𝑡 − 1 and 𝑡, is then given by:

𝑹𝑷,𝒕 = 𝒘𝟏,𝒕-𝟏 𝑹𝟏,𝒕 + 𝒘𝟐,𝒕-𝟏 𝑹𝟐,𝒕 + ⋯ + 𝒘𝑵,𝒕-𝟏 𝑹𝑵,𝒕

15/10/2022 22
 Return on a portfolio of assets (cont.)
• We stack returns and weights into column vectors of size 𝑁:
𝑅',% 𝑤',%-'
𝑅 𝑤),%-'
𝑹% = ),% , 𝒘%-' = ⋮
⋮
𝑅+,% 𝑤+,%-'

• The return of the portfolio can be compactly rewritten as:

𝑅3,% = 𝒘4%-' 𝑹%
where 𝒘-"#$ is the transpose of the vector 𝒘"#$ .

• If the asset returns, 𝑅*," , are jointly normal, then

the portfolio returns, 𝑅.," , are also normally distributed.

15/10/2022 23
 Portfolio of 2 assets
Consider a portfolio 𝑃 of 2 assets with the following characteristics:
• Expected returns: 𝜇' = E 𝑅' and 𝜇) = E 𝑅)
• Standard deviations of returns: 𝜎' = Std 𝑅' and 𝜎) = Std 𝑅)
• Covariance of returns: 𝜎') = Cov 𝑅' , 𝑅)
• Correlation of returns: 𝜌') = Corr 𝑅' , 𝑅) = !!"#
!
" #

• Portfolio weights: 𝑤' and 𝑤)

The return on the portfolio is 𝑅. = 𝑤$ 𝑅$ + 𝑤' 𝑅'

The expected return is therefore 𝜇. = 𝐸 𝑅. = 𝑤$ 𝜇$ + 𝑤' 𝜇'

The portfolio variance, 𝜎.' = Var 𝑅. , is given by:

𝜎.' = 𝑤$' 𝜎$' + 𝑤'' 𝜎'' + 2𝑤$ 𝑤' 𝜎$' = 𝑤$' 𝜎$' + 𝑤'' 𝜎'' + 2𝑤$ 𝑤' 𝜌$' 𝜎$ 𝜎'
15/10/2022 24
 Portfolio of N assets
Consider a portfolio 𝑃 containing 𝑁 assets with the following characteristics:
• 𝑁×1 vector of expected returns: 𝛍 = E 𝐑
• 𝑁×𝑁 covariance matrix: 𝚺
• 𝑁×1 vector of portfolio weights: 𝐰

𝜇' 𝜎'' 𝜎') 𝜎'+ 𝑤'

𝜇) 𝜎)' 𝜎)) ⋯ 𝜎)+ 𝑤)
𝛍= ⋮ , 𝚺= ⋮ ⋮ ⋱ ⋮ , 𝐰= ⋮
𝜇+ 𝜎+' 𝜎+) ⋯ 𝜎++ 𝑤+

The return on the portfolio is 𝑅. = 𝑤$ 𝑅$ + ⋯ + 𝑤/ 𝑅/

The expected return of the portfolio return is: 𝜇. = ∑/ 𝑤 𝜇

*)$ * * = 𝐰 -𝛍

The portfolio variance is 𝜎.' = ∑/ ∑ /

𝑤 𝑤 𝜎
*)$ 0)$ * 0 *0 = 𝐰 -
𝚺𝐰
15/10/2022 25
 Markowitz model

with

2 risky assets

15/10/2022 26
 The Markowitz model (1952)
The model is based on the following hypotheses.

• Investors are rational and risk averse.

- They prefer portfolios with high average return and low risk.

• In the Markowitz model, investors choose portfolios based on:

- expected return (first moment of portfolio return),
- variance (second moment of portfolio return).

15/10/2022 27
 Markowitz model with 2 assets
• We consider a portfolio of two assets, with respective weights 𝑤$ and 𝑤' .
• Because of the portfolio constraint, 𝑤$ + 𝑤' = 1, there is only one degree of
freedom for the choice of the weights:
𝑤' = 1 − 𝑤$

• Portfolio expected return and variance:

𝜇. 𝑤$ = 𝑤$ 𝜇$ + 1 − 𝑤$ 𝜇'

𝜎.' 𝑤$ = 𝑤$' 𝜎$' + 1 − 𝑤$ ' 𝜎'' + 2𝑤$ 1 − 𝑤$ 𝜎$ 𝜎' 𝝆𝟏𝟐

• The portfolio variance depends crucially on the correlation of the assets.

15/10/2022 28
 Special case: perfect correlation
The variance of the portfolio is:
𝜎.' 𝑤$ = 𝑤$' 𝜎$' + 1 − 𝑤$ ' 𝜎'' + 2𝑤$ 1 − 𝑤$ 𝜎$ 𝜎' 𝜌$'

Assume for now that only long positions are allowed, 0 ≤ 𝑤$ , 𝑤' ≤ 1.

The correlation between the returns of the two assets, 𝜌$' , plays a key role in
the risk-return characteristics of the portfolio. Two particular values are:
• 𝜌$' = +1: perfect positive correlation,
• 𝜌$' = −1: perfect negative correlation.

In these two cases, the portfolio variance 𝜎.' 𝑤$ can be written as the square
of a binomial: 𝑎 ± 𝑏 ' = 𝑎' + 𝑏 ' ± 2𝑎𝑏.
15/10/2022 29
 Perfect positive correlation
• When 𝜌$' = +1, the portfolio variance simplifies to:

𝜎.' 𝑤$ = 𝑤$' 𝜎$' + 1 − 𝑤$ ' 𝜎'' + 2𝑤$ 1 − 𝑤$ 𝜎$ 𝜎' = 𝑤$ 𝜎$ + 1 − 𝑤$ 𝜎' '

• Portfolio volatility is:

345! 4$
𝜎. 𝑤$ = 𝑤$ 𝜎$ + 1 − 𝑤$ 𝜎' 𝜎. 𝑤$ = 𝑤$ 𝜎$ + 1 − 𝑤$ 𝜎'

where the absolute value can be dropped when 0 ≤ 𝑤$ ≤ 1

(no short sales allowed).

15/10/2022 30
 Perfect positive correlation (cont.)
Expected return and volatility are
both linear in 𝑤$ . 𝜇$

𝜇 = 𝑤' 𝜇' + 1 − 𝑤' 𝜇) 𝜇# 𝑆#

O 3
𝜎3 = 𝑤' 𝜎' + 1 − 𝑤' 𝜎)

Eliminate 𝑤$ in order to write 𝜇. as 𝑆"

𝜇"
a function of 𝜎. :

𝜎3 − 𝜎)
𝜇3 = 𝜇) + 𝜇' − 𝜇)
𝜎' − 𝜎)

The portfolio expected return is a

𝜎" 𝜎# 𝜎$
linear function of portfolio volatility.

15/10/2022 31
 Perfect negative correlation
• If 𝜌$' = −1, the portfolio variance can be rewritten as:
𝜎3) 𝑤' = 𝑤' 𝜎' − 1 − 𝑤' 𝜎) )

• Portfolio volatility is:

𝜎3 𝑤' = 𝑤' 𝜎' − 1 − 𝑤' 𝜎)

• The expected portfolio return is:

5# 𝜇3 = 𝑤' 𝜇' + 1 − 𝑤' 𝜇) 5 65
For 𝑤' ≥ : O ⇒ 𝜇3 = 𝜇) + $ # 𝜇' − 𝜇)
5" 65# 𝜎3 = 𝑤' 𝜎' − 1 − 𝑤' 𝜎) 5" 65#

5# 𝜇3 = 𝑤' 𝜇' + 1 − 𝑤' 𝜇) 5 -5

For 𝑤' < : O ⇒ 𝜇3 = 𝜇) − $ # 𝜇' − 𝜇)
5" 65# 𝜎3 = 1 − 𝑤' 𝜎) − 𝑤' 𝜎' 5" 65#

15/10/2022 32
 Perfect negative correlation (cont.)
When 𝜌$' = −1, we have:
𝜇$
𝜎$ + 𝜎"
𝜇" + 𝜇 − 𝜇" , 𝜇$ ≥ 𝜇)$
𝜎# + 𝜎" # 𝜇#
𝜇$ = 𝜎$ − 𝜎" 𝑆#
𝜇" − 𝜇 − 𝜇" , 𝜇$ < 𝜇)$
𝜎# + 𝜎" #
𝜇)$

6% 𝜇" 𝑆"
where 𝜇S. = 𝜇' + 𝜇$ − 𝜇' .
6! (6%

𝜎" 𝜎# 𝜎$

15/10/2022 33
 Imperfectly correlated assets
We stick to a situation where short sales are not allowed: 0 ≤ 𝑤$ , 𝑤' ≤ 1.
When assets are imperfectly correlated, the return/volatility curve is a
hyperbola contained between the segments obtained for 𝜌$' = −1 and 𝜌$' = 1.
𝜇$
We obtain the equation of the hyperbola
by solving for 𝑤$ in the first equation
below and substituting it into the second
𝑆# equation.
𝜇#
𝜇)$ 𝜇$ = 𝑤# 𝜇# + 1 − 𝑤# 𝜇"
+
𝜎$" = 𝑤#" 𝜎#" + 1 − 𝑤# " 𝜎"" + 2𝑤# 1 − 𝑤# 𝜎# 𝜎" 𝜌#"
𝜇"
𝑆"
Try it!

𝜎" 𝜎# 𝜎$
15/10/2022 34
 Example: Diversification effect

What are the expected returns and

volatilities of assets D and E?

You can get them from the graph:

𝜇7 = 8%, 𝜎7 = 12%
𝜇8 = 13%, 𝜎8 = 20%

For a given expected return and portfolio

weights between 0 and 1, the portfolio
become less risky when 𝜌 decreases
(diversification effect).

15/10/2022 35
 Minimum variance portfolio
We compute the weight 𝑤$ that minimizes the portfolio variance 𝜎.' 𝑤$ :

𝜎.' 𝑤$ = 𝑤$' 𝜎$' + 1 − 𝑤$ ' 𝜎'' + 2𝑤$ 1 − 𝑤$ 𝜎$ 𝜎' 𝜌$'

We write the first-order condition:

d𝜎.' 𝑤$
= 2𝑤$ 𝜎$' − 2 1 − 𝑤$ 𝜎'' + 2 1 − 2𝑤$ 𝜎$ 𝜎' 𝜌$' = 0
d𝑤$

The solution is:

'
∗
𝜎' − 𝜎$ 𝜎' 𝜌$'
𝑤$ = '
𝜎$ + 𝜎'' − 2𝜎$ 𝜎' 𝜌$'

15/10/2022 36
 Markowitz model

with

N risky assets

15/10/2022 37
 General problem
• Consider an individual who can invest in 𝑁 assets, for which the vector of
expected returns 𝛍 and the covariance matrix 𝚺 are known.

• The investor can choose the portfolio weights 𝒘 = 𝑤$ , 𝑤' ,…, 𝑤/ - .

• The investor is subject to the constraint that the sum of portfolio

weights is equal to 1 :
𝑤$ + 𝑤' + ⋯ + 𝑤/ = 𝒘- 𝟏 = 1,
where 𝟏 = 1 1 … 1 - is an 𝑁×1 column vector of ones.

• The investment objective is to minimize the portfolio variance, 𝜎.' 𝒘 ,

so that the return of the portfolio, 𝜇. 𝒘 , is equal to the target, 𝜇.̅

15/10/2022 38
 Analytical solution
The agent solves:
min w¢Σw Variance
w
minimization
s.t.
Target portfolio
w¢μ = µ and w¢1 = 1 Portfolio
expected return
constraint

When there are no further constraints on the portfolio weights, the

solution is available in closed form:
𝑎8 𝜇̅ − 𝑎) -' 𝑎' − 𝑎) 𝜇̅ -'
𝒘∗ = 𝚺 𝝁+ 𝚺 𝟏
𝑎' 𝑎8 − 𝑎)) 𝑎' 𝑎8 − 𝑎))

where 𝑎' , 𝑎) and 𝑎8 are three constant scalar quantities:

𝑎' = 𝝁4 𝚺 -' 𝝁 > 0, 𝑎) = 𝝁4 𝚺 -' 𝟏, 𝑎8 = 𝟏4 𝚺 -' 𝟏 > 0

It can be shown that 𝑎' 𝑎8 − 𝑎)) ≥ 0.

15/10/2022 39
 Analytical solution
𝑎, 𝜇̅ − 𝑎' #$ 𝑎$ − 𝑎' 𝜇̅ #$
𝒘∗ = 𝚺 𝝁+ 𝚺 𝟏
𝑎$ 𝑎, − 𝑎'' 𝑎$ 𝑎, − 𝑎''

The optimal portfolio is the linear combination of two funds.

• The first fund is a risk-adjusted performance-seeking portfolio.

• The second fund is a risk-minimizing portfolio.

The weight of each fund in the optimal portfolio depends on the

[.
target return 𝝁

15/10/2022 40
 Proof (optional)
The investor minimizes 𝒘- 𝚺𝒘 subject to the constraints 𝒘′𝝁 = 𝜇̅ and 𝒘- 𝟏 = 1.

The Lagrangian is:

ℒ 𝒘 = 𝒘- 𝚺𝒘 − 𝜆 𝒘- 𝝁 − 𝜇̅ − 𝜙 𝒘- 𝟏 − 1

The first derivative w.r.t. 𝒘 is zero:

2𝚺𝒘 − 𝜆𝝁 − 𝜙𝟏 = 0

Assume that 𝚺 is invertible and solve for the optimal 𝒘:

∗
𝜆 #$ 𝜙 #$
𝒘 = 𝚺 𝝁+ 𝚺 𝟏
2 2
We next solve for 𝜆 and 𝜙.

15/10/2022 41
 Proof (optional)
Substitute 𝒘∗ into the constraints 𝒘- 𝝁 = 𝜇̅ and 𝒘- 𝟏 = 1:

𝜆 - #$ 𝜙 - #$
𝝁 𝚺 𝝁 + 𝝁 𝚺 𝟏 = 𝜇̅
2 2
𝜆 - #$ 𝜙 - #$
𝟏 𝚺 𝝁+ 𝟏 𝚺 𝟏=1
2 2

Finally, we solve for the Lagrange multipliers 𝜆 and 𝜙:

𝜆 𝑎, 𝜇̅ − 𝑎' 𝜙 𝑎$ − 𝑎' 𝜇̅
= ' , =
2 𝑎$ 𝑎, − 𝑎' 2 𝑎$ 𝑎, − 𝑎''
where:
𝑎$ = 𝝁- 𝚺 #$ 𝝁, 𝑎' = 𝝁- 𝚺 #$ 𝟏, 𝑎, = 𝟏- 𝚺 #$ 𝟏

15/10/2022 42
 Minimum variance portfolio: Example

Consider the special case of 3 uncorrelated assets:

1
0 0
𝜎')
𝜎') 0 0
1
𝚺= 0 𝜎)) 0 ⇒ 𝚺 -' = 0 0
𝜎))
0 0 𝜎8) 1
0 0
𝜎8)
1 1
0 0
𝜎') 𝜎')
1 1 1
Y=
⇒ 𝒘 𝚺 -' 𝟏 = 0 0 1 = )
𝜎)) 𝜎)
1
1 1
0 0
𝜎8) 𝜎8)
15/10/2022 43
 Minimum variance portfolio: Example
$ $ $
• The three portfolio weights are 𝑤
a$ = ,𝑤
a' = ,𝑤
a, = .
6!% 6%% 6&%

• The weight of each asset is inversely proportional to its variance. Assets

with lower risk are over-weighted relative to those with higher risk.

• Asset expected returns do not impact the risk-minimizing portfolio.

• In general, 𝑤
a$ , 𝑤
a' , and 𝑤
a, do NOT add up to 1. They are rescaled to form
the mean-variance frontier portfolio, so that, once they are added to the
rescaled performance-seeking portfolio, the total weights add up to 1.

• When assets are correlated, the expressions for the portfolio weights are
more complicated.

15/10/2022 44
 Performance-seeking portfolio
Consider the continued example of 3 uncorrelated assets,
with volatilities 𝜎$ , 𝜎' , 𝜎, , and expected returns 𝜇$ , 𝜇' , 𝜇, :

1 𝜇!
0 0
𝜎!" 𝜎!"
𝜇! 𝜇"
1
. = 𝚺 '! 𝝁 = 0
𝒘 0 𝜇" = "
𝜎"" 𝜇( 𝜎"
1 𝜇(
0 0 𝜎("
𝜎("

15/10/2022 45
 Performance-seeking portfolio (cont.)
:! :% :&
• The three weights are thus 𝑤
a$ = , 𝑤
a' = , 𝑤
a, = .
6!% 6%% 6&%

• The weight of each asset is proportional to its expected return (performance-

seeking) and inversely proportional to its variance (risk-adjustment). Assets
with high performance/low risk are over-weighted with respect to those with
low performance/high risk.

• If assets are correlated, the weights of assets in the performance-seeking

portfolio have more complicated expressions.

15/10/2022 46
 Mean-variance efficient frontier
We want to determine the equation of the mean-variance portfolio frontier.
The optimal portfolio weights 𝒘∗ is a function of the target portfolio return 𝜇.̅
Compute the portfolio variance. After some simple calculations:

𝑎 𝜇̅ ) − 2𝑎 𝜇̅ + 𝑎
8 ) '
𝜎3) = 𝒘∗ 4 𝚺𝒘∗ =
𝑎' 𝑎8 − 𝑎))

The portfolio variance 𝜎.' is a quadratic function of the target return 𝜇.̅
In the volatility 𝜎. - return 𝜇̅ space, the efficient frontier is a hyperbola.

𝜇̅

𝜎$
15/10/2022 47
 Additional trading constraints
• Until now, the investor can take any (long or short) position.

• Often the investor faces constraints on the positions they can take.

- Short-sales constraints on some assets:

𝑤* ≥ 0 for some 𝑖 ∈ {1. . 𝑁}

- Exposure constraints: the exposure (long or short) on some assets

could be limited to a certain amount 𝐾 (e.g. 𝐾 = 20%):
−𝐾 ≤ 𝑤* ≤ 𝐾 for some 𝑖 ∈ {1. . 𝑁}

15/10/2022 48
 Additional trading constraints (cont.)
• These constraints are incorporated in the mean-variance optimization
problem.

• For example, in the case of short-sale constraints on all the assets,

the problem is:
min 𝒘4 𝚺𝒘
𝒘
𝒘4 𝝁 = 𝜇̅
𝒘4 𝟏 = 1
𝑤* ≥ 0 ∀𝑖 ∈ 1. . 𝑁

• In the tutorial, you will learn how to compute the numerical solution.

15/10/2022 49
 Effect of short-sales constraints
• We illustrate below the mean-variance frontier of a set of commodities with long-
only positions.
• The effect of weight constraints is in general to reduce ex-ante efficiency
(based on expected return and variance), which is why the long-only frontier is
“inside” the unconstrained frontier.
• The endpoints of the long-only solution correspond to the asset with the lowest
return and the asset with the highest return.

15/10/2022 50
 Markowitz model:

Practical issues

15/10/2022 51
 In-sample vs. out-of-sample analysis

Example

• You observe the monthly returns on 20 stocks over the 2001 to 2020
period (20 years).

• You want to study the performance of an investment strategy over

the period 2011-2020.

• The investment strategy, characterized by the portfolio weights,

depends on a set of parameters Θ (e.g.: expected returns and
volatilities) estimated on the stock return data.

15/10/2022 52
 In-sample vs. out-of-sample analysis

• If you estimate the vector of parameters Θ using data after Jan 2011:
- You are using information unavailable to a real-world investor to
build your portfolio, thus performing an in-sample analysis.
- This produces a look-ahead bias.

• If you estimate the vector of parameters Θ using only data before

Jan 2011, you are instead performing an out-of-sample analysis.

15/10/2022 53
 Backtesting

Test of investment strategies

• Assessing the performance of a portfolio strategy using historical

data and performing an out-of-sample analysis is also called
backtesting.

• Such a test is meant to replicate the behavior of a real investor.

• Does this eliminate all look-ahead biases?

Not always: investment strategies are often designed to perform
well (“look good”) in backtests. Since past performance is not
always indicative of future performance, this can be dangerous!

15/10/2022 54
 Parameter uncertainty
• Asset return parameters can be inferred:
- from past data
- from expert views

• This induces estimation errors in the model inputs used to

determine the optimal portfolio:
- expected returns
- volatilities
- correlations

• Estimation error has a major impact on the choice of portfolio

weights and on portfolio performance.

15/10/2022 55
 Parameter uncertainty
• The academic literature has analyzed the impact of estimation error in
different portfolio models.
Chopra and Ziemba (1993), Kan and Zhou (2007), DeMiguel, Garlappi and
Uppal (2009)

• These papers show that the out-of-sample behavior of mean-variance

efficient portfolios is severely compromised by estimation error.

• In particular, the estimation error in expected returns strongly affects

the performance of the mean-variance portfolio.

• Estimation error in volatilities and correlations (covariance matrix) has

a milder but non-negligible impact.

15/10/2022 56
 Practice Questions

15/10/2022 57
 Problem 1
Distribution of return on Stock XYZ

n Rate of Probability of
Return Occurrence
1 12% 0.18
2 10% 0.24
3 8% 0.29
4 4% 0.16
5 -4% 0.13
Total 1.00

Compute the expected return, variance and standard deviation.

15/10/2022 58
 Solution to Problem 1
E(RXYZ) = 0.18*12% + 0.24*10% + 0.29*8% + 0.16*4% + 0.13*(-4%) = 7%

Var(RXYZ) = 0.18*(12% - 7%)2 + 0.24*(10% - 7%)2 + 0.29*(8% - 7%)2 +

+ 0.16*(4% - 7%)2 + 0.13*(-4% - 7%)2 = 24.12(%%)

Std(RXYZ) = sqrt(24.12%%) = 4.91%

where the symbol %% corresponds to a multiplier equal to 10-4.

n Rate of Probability of
Return Occurrence
1 12% 0.18
2 10% 0.24
3 8% 0.29
4 4% 0.16
5 -4% 0.13
Total 1.00
15/10/2022 59
 Problem 2
Consider stocks XYZ and ABC

n Rate of Rate of Probability

Return Return of
XYZ ABC Occurrence
E(RXYZ) = 7%
1 12% 21% 0.18
E(RABC) = 10%
2 10% 14% 0.24
Var(RXYZ) = 24.12%%
3 8% 9% 0.29
Var(RABC) = 53.64%%
4 4% 4% 0.16
5 -4% -3% 0.13
Total 1.00

Compute the expected return, variance and standard deviation.

15/10/2022 60
 Solution to Problem 2
n Rate of Probability of E(RXYZ) = 7%
Return Occurrence
E(RABC) = 10%
1 12% 0.18
Var(RXYZ) = 24.12%%
2 10% 0.24
Var(RABC) = 53.64%%
3 8% 0.29
4 4% 0.16
5 -4% 0.13
Total 1.00

Cov(RXYZ, RABC) = 0.18(12%-7%)(21%-10%) + 0.24(10%-7%)(14%-10%)+

+ 0.29(8%-7%)(9%-10%) + 0.16(4%-7%)(4%-10%)
+ 0.13(-4%-7%)(-3%-10%) = 33.96%%

Corr(RXYZ, RABC) = Cov(RXYZ, RABC) / ( Std(RXYZ) * Std(RABC) ) = 94.41%

15/10/2022 61