Required Background

for Portfolio Management

KU Leuven

Tamas Vadasz

September 11, 2022

Preliminary draft. Please do not circulate!

Version 0.2.

Contents
1 Modelling financial assets 2
1.1 Asset returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Compounding returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Returns as random variables . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Compounding variances . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Covariance and correlation . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 The variance-covariance matrix . . . . . . . . . . . . . . . . . . . . . . . 8
1.7 (Linear) combination of random variables . . . . . . . . . . . . . . . . . 8

2 Econometrics 10
2.1 Sample moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Linear regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Time series vs. cross-sectional regression . . . . . . . . . . . . . . . . . . 11
2.4 Hypothesis testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 Multivariate regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Analysis 14
3.1 Quick reminder: derivatives . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Maximization problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1
1 Modelling financial assets
This section reviews the most important terms and concepts of modelling financial assets,
and the mathematical/statistical background for portfolio management.

1.1 Asset returns

Financial assets (stock, bond, etc.) are modelled as series of price realizations Pt over
time t. In general, time can be either discrete (t ∈ 0, 1, 2, ..., T ) or continuous (t ∈ [0, T ]),
but in this course we only consider discrete time. So prices are denoted as P0 , P1 , P2 , ....
Returns can be calculated from prices in two ways:

1. Simple return (r):

Pt+1
1 + rt = ⇔ Pt+1 = Pt (1 + rt )
Pt

Often R := 1 + r denotes gross return, in this context r is called net return.

2. Log-return or continuously compounded return (i):

Pt+1
it = ln[ ] = ln Pt+1 − ln Pt ⇔ Pt+1 = Pt eit
Pt

Usual motivation: continuous compounding. P1 = P0 × 1 + i n

→ P0 ei if n → ∞


n

The two returns i and r capture the same price evolution with the relationship i = ln[R].

🤔 A simple return of 10% corresponds to a log-return of ?

i = ln(1.1) ≈ 0.0953 = 9.53%

Simple vs. log-return

• Log-returns are time-additive, simple returns are not.

🤔 You gain 10%, then lose 10% on your portfolio? How much do you
have at the end (i) if these numbers represent log-returns, (ii) if these
numbers represent simple net returns?

• Simple returns are portfolio-additive, but log-returns are not.

This means: if you form a portfolio P by investing a percentage wa to stock A and wb to stock
B, then Rp = wa Ra + wb Rb , but ip 6= wa ia + wb ib !

2
 • Usual advice: use log-returns when you work with time-series models, use simple
returns when you work with cross-sectional data or portfolios of assets.
💭 In practice the difference is usually small, especially for small returns / short horizons. For
this course think always of ‘simple returns’, unless otherwise noted. In practical applications the
difference is important.

• Log-returns can take any value. Simple returns are between [−1, ∞].

❗ Notice that simple returns are logically inconsistent with the “usual” assertion that “returns
are normally distributed”. (Why? Normal distribution has unbounded support.) Formal
asset pricing theory therefore often assumes that log-returns are normally distributed.

1.2 Compounding returns

“Compounding” = compute returns over multiple periods1 .

1. With simple returns:

1 + r02 = (1 + r01 )(1 + r12 )

where rij denotes (compound) returns between periods t = i and t = j.

🤔 A daily return of 0.1% corresponds to an annual return (N = 252,

number of trading days) of ?

ra = (1 + 0.001)252 − 1 = 28.6%

2. With log-returns:
i02 = i01 + i12

When the subsequent returns are the same over multiple periods it follows that

Pt = P0 eit

Usually we work with annualized returns. For example, if for two periods we have
√
r02 = 0.21, then the annualized two-year return is r2 = 1 + 0.21 − 1 = 0.1. The same
goes with log-returns, if i02 = 0.2 then the annualized log-return is i2 = 0.2
2
= 0.1. A
monthly simple return of 1% corresponds to an annualized rate of return of (1.0112 − 1),
while 1% monthly log-return corresponds to 12 × 1 = 12% annualized log-return. This
shows why it’s so much easier to work with log-returns in a time-series context.
1
Here let rtk denote the return between period t and k. When I write rt (just one subscript), I think
of the per period (i.e., annualized) return over a period of length t.

3
 🤔 (The power of compounding): your great-grandfather has invested $100
in a portfolio of US stocks in 1930 which you inherit today. How much money
do you have today? (Make a guess before opening the link!)
a $1000 (i.e. 10 times the original investment)
b $10.000 (i.e. 100 times the original investment)
c $100.000 (i.e. 1000 times the original investment)
d $1.000.000 (i.e. 10 000 times the original investment)

Answer: see this graph. Also, read this classic story about exponential growth.

1.3 Returns as random variables

Random variable: can take any value with the associated probability.

• Discrete: Probability mass function P rob[x̃ = µ]

• Continuous: Cumulative distribution function [CDF]: P rob[x̃ ≤ µ]

❗ We do not know the realization, but we do know the rules which generate those realizations,
that is, the return generating process. Sometimes it is useful to distinguish the random variable
r̃ from the actual realization ri .

• Typical assumption in finance: returns are normally distributed.2

• If i (log-return) is normally distributed, than prices are lognormally distributed:

➡️ a variable X is lognormally distributed if ln(X) is normally distributed.

i ∼ N (µ, σ 2 ) ⇔ ei ∼ LN (µ, σ 2 )

An important relationship: E[R] = eµ+σ

2 /2

As we often say, investors care about return and risk. We proxy these with the
characteristics (first two moments) of the underlying return distribution. Namely,

1. Return = expected value: µ = E[r̃], estimated from realized returns as

T
1X
µ̂ = rt
T t=1

2. Risk = standard deviation (σ) or variance: σ 2 = E[(r̃ − µ)2 ], estimated as

T
1 X
2
σ̂ = (rt − µ)2
T − 1 t=1
2
The great thing about normal distribution: it is completely characterized by two parameters, its mean
µ and its variance σ 2 (or standard deviation σ). In reality, returns are not normal, but nevertheless, this
is often a useful approximation.

4
 What do we mean by investors care? This is the subject of the utility theory
(1st lecture). Under certain restrictive assumptions, one can derive a convenient, very
simple mean-variance utility function, which is often used in education as it leads to nice
tractable solutions.3
1
U (µ, σ; A) = µ − Aσ 2
2
A is a parameter which captures the attitudes towards risk.

1.3.1 Some critical observations

• Can we really say that investors only care about return and risk? What about
social responsibility? What about ethics? What about behavioural biases?
Theory works only under very specific assumptions about human behaviour.

• Can we precisely estimate expected value and variance from historical data?
The underlying technical assumption (usually implicit) is called statistical stationarity.
A time series is stationary if its statistical properties (mean, variance, etc.) do not change
over time. This is generally not true! (Imagine how much COVID-19 has influenced the
statistical properties of the return distributions of stocks or the volatility of prices).

• Can we really say that variance is a sufficient proxy for risk?

Only if the return is normal (or comes from other similar two-parameter family). When
fat tails or other asymmetries are important, this is insufficient.

• Can we really say that returns are normally distributed?

“Experiences with real-world data, however, soon convinces one

that both stationarity and Gaussianity are fairy tails invented for
the amusement of undergraduates.”
(Thomson,1994)

See the pictures below. Figure (a) shows that the S&P500 index moves fairly ran-
domly. However, looking at the daily realizations (plot b) reveals what is called ‘volatility
clustering’ - volatile days tend to follow volatile days, and calm days tend to follow calm
days. This is one example of non-stationarity. Part (c) illustrate the empirical distribu-
tion of returns. It looks random, but when the true random distribution (with the same
expected value and volatility) is superimposed, one can see the famous “fat tails” phe-
nomenon: extreme values are much more likely than what would follow from normality.
3
We will discuss in class what assumptions are needed for such mean-variance preferences to be
accurate.

5
 S&P closing prices [2004−01−02/2018−08−28]
140 S&P returns 2004−01−02 / 2018−08−28
Last 99.739998

120 0.15 0.15

0.10 0.10
100

0.05 0.05
80

0.00 0.00

60
−0.05 −0.05

40 −0.10 −0.10

−0.15 −0.15

Jan 02 2004 Jul 03 2006 Jan 02 2009 Jul 01 2011 Jan 02 2014 Jul 01 2016 Jan 02 2004 Jan 03 2006 Jan 02 2008 Jan 04 2010 Jan 03 2012 Jan 02 2014 Jan 04 2016 Jan 02 2018

(a) S&P 500 price realizations. (b) S&P 500 return realizations.
S&P return histogram S&P return histogram
25

25
20

20
15

15
Density

Density
10

10
5

5
0

0
−0.15 −0.10 −0.05 0.00 0.05 0.10 0.15 0.20 −0.15 −0.10 −0.05 0.00 0.05 0.10 0.15 0.20

ret ret

(c) Histogram of returns. (d) Histogram of returns vs. a normal

distribution. Fat tails.

Figure 1: Illustration of non-normality: S&P price and return realizations.

1.4 Compounding variances

Given i01 and i12 , it is easy to calculate the compound (two-period) return as4

i02 = i01 + i12

Volatility is more complicated. We want to calculate the standard deviation of the sum of
random variables. There is a well-known statistical property of the normal distribution:
if X ∼ N (µX , σX2
) and Y ∼ N (µY , σY2 ) are independent, then

2
X + Y ∼ N (µX + µY ; σX + σY2 )

🤔 Our daily return series exhibit a volatility of 2%. What is the annual
volatility (N = 252) of this time series?
p
σ= N × 0.022 = 31.7%

❗ This is known as the square-root rule. The underlying assumption is that subsequent
return realizations are independent from each other. Such serial independence is a very
strong assumption (↔ efficient market hypothesis).
4
I use log-returns in this section to keep the math simple. This is not crucial.

6
1.5 Covariance and correlation
One rarely invests in just one asset5 . With more than one assets in your portfolio, you
care about the co-movement of those assets. This is the main subject of portfolio theory.

• Covariance: a measure of joint variability of two random variables.

Covariance between returns of A and B: (Notation: Cov[A, B] or σAB )

σAB := E [(r̃A − µA )(r̃B − µB )]

T
1 X
σAB = (rA,t − µA )(rB,t − µB )
N − 1 t=1

(The first expression is a definition for random variables. The second is estimation from a sample.)

Notice that Cov[A, A] = σA2 .

• Correlation coefficient: a normalized version of covariance.

σAB
ρAB =
σA σB

– ρ = 1, perfectly correlated
– ρ = 0, uncorrelated (6= independent!!!)
– ρ = −1, perfectly negatively correlated

❗Pitfalls

• Correlation should not be interpreted as causation.6

• Correlation is a measure of linear dependence. If the true relationship is nonlinear,

correlation can be misleading. A famous example is the Anscombe’s quartet.

• Independence ⇒ uncorrelated, but uncorrelated ; independence7 . Example: x ∼

N (0, 1) and y = x2 .

• Correlation is not invariant under non-linear increasing transformations. (Conse-

quence: correlation is slightly different for log-returns and normal returns.)

• Can be very misleading if the two series’ have a common underlying trend.
5
I hope to convince you by the end of this course to never do that.
6
The rooster’s crow is highly correlated with sunrise, yet roosters likely do not cause the sun to rise.
7
In case of multivariate normal distribution it is actually true that ρ = 0 ⇒ independent).

7
1.6 The variance-covariance matrix
A generalization of covariance to N random variables is the variance-covariance ma-
trix (or just covariance matrix for brevity). It contains the covariance between every
pair of the random variables. Typically is denoted by Σ (large Sigma) .

• The matrix is symmetric because Cov(X, Y ) = Cov(Y, X)

• The diagonal contains the variances, because Cov(X, X) = σx2

• The covariance-matrix is positive semi-definite.

ℹ️ A matrix M is positive semi-definite if and only if x0 M x ≥ 0 for all non-zero x-vector8 . In

practice this property implies that any linear combination of the random variables must
have a non-negative variance. For portfolio management applications this means that any
portfolio must have a non-negative variance (which is a fairly reasonable requirement).

1.7 (Linear) combination of random variables

In portfolio theory we often work with some (linear) combination of random variables9 .
For example, let

X ∼ N (µx , σx ) and Y ∼ N (µy , σy ) and Cov(X, Y ) = σxy

and we want to know the properties of Z = αX + βY . From the properties of the normal
distribution, we can derive that

µz = αµx + βµy
σz2 = α2 σx2 + β 2 σy2 + 2αβσxy

With words: returns are portfolio-additive, but risk is sub-additive.10

More generally, with N normally distributed random variables X1 ...XN combined with
8
A little reminder of matrix-algebra:
XX
x0 M x = xi Mij xj
i j

Understanding matrix-algebra is not required for the exam, although sometimes I’m going to use it.
9
Financial assets are modelled as random variables. A portfolio is a linear combination.
10
Subadditivity of risk is probably the most important element of investment management. Without
this sub-additive property, there would be no Investment Management as a profession, no Finance as a
masters degree program, and you would likely be studying for a philosophy or an art degree.

8
the weights given in vector α := (α1 ...αN )0 , then

µz = α 0 µ =
X
α i µi
i

σz2 = α0 Σα =
XX
αi αj σij
i j

❗This is the expected value and variance of an investment portfolio!

9
2 Econometrics
📖 An excellent book for econometrics with finance applications: Brooks, C. (2019).
Introductory econometrics for finance. Cambridge university press. (Figures below taken
from this book.)

2.1 Sample moments

Estimators. There are two properties of estimators we often refer to:

• Consistency: consistent estimators converge (in probability) to the true value in

large samples.

• Bias: is the expected difference between the estimated value and the true parameter
to be estimated. For unbiased estimators this is zero (even in small samples).

An estimator can be consistent but biased, or can be unbiased but not consistent (rare).
Estimating mean and variance.

• The consistent and unbiased estimators for mean and variance are:
T
1X
µ̂ = rt (1)
T t=1
T
1 X
σ̂ 2 = (rt − µ)2 (2)
T − 1 t=1

(The division by T-1 instead of T is a bias correction).

• These estimators are also random variables. This means they are estimated with
uncertainty, the point estimate may or may not be the true value.
If I estimate the performance of my fund manager as a past average, I obtain a realization
of the sample estimate as a random variable. Is this skill or luck?

2.2 Linear regression

• Describes a linear relationship between variable y and x.

• Univariate linear regression:

yi = α + βxi + εi

E[εi ] = 0, so the model tells the expected value of yi given xi

E[y|xi ] = α + βxi

10
 or unconditionally:
E[y] = α + βE[xi ]

• Estimation technique: ordinary least-squares (OLS estimation)

minimize squared differences between observation and prediction ⇒
PN
i=1 (xi − x) (yi − y) Cov(x, y)
β̂ = PN 2 =
i=1 (xi − x)
σx2

• The Gauss-Markov theorem: OLS is BLUE under the classical assumptions (best
linear unbiased estimator).

• Classical assumptions: (1) the relationship is linear; (2) observations are randomly
sampled; (3) zero conditional mean (E[e|x] = 0); (4) No multicollinearity; (5)
homoscedasticity, no autocorrelation; (6) [optional] the error term is normally dis-
tributed.

• Interpretation of parameters α and β.

2.3 Time series vs. cross-sectional regression

Time-series regression:

• LHS and RHS is a time-series of a variable:

yt = α + βxt + εt

• Example: regressing stock-returns (TS of actual realizations) on a market index.

Cross-sectional regression:

• LHS and RHS variables are contemporary observations of different entities:

yi = α + βxi + εi

• Example: regressing expected stock returns for a universe of stocks (one number for
each stock) on possible explanatory variables (firm characteristics, riskiness, etc.).

2.4 Hypothesis testing

• Estimate of the variance of the error term:

ε̂2i
P
s2 =
T −2

11
• Estimators of the standard errors (of the coefficient):
qP
x2t
s s
SE[α] = p P and SE[β] = pP
T (xt − x)2 (xt − x)2

• Distribution of the estimated coefficients:

α̂ − α β̂ − β
∼ tT2 and ∼ tT2
SE(α̂) SE(β̂)

• Testing the hypothesis that β = β ? (H0 ) against β 6= β ? (H1 )

β̂ − β
−tcrit ≤ ≤ +tcrit
SE(β̂)

(a) Hypothesis testing. (b) The t-distribution (vs. normal).

12
 Example: can investment funds beat the market? Jensen (1968)

• Annual returns on portfolios of mutual funds, OLS regression:

Rjt − Rf t = αj + βj (Rmt − Rf t ) + ujt

Excess return of the portfolio regressed on excess return of the market and a constant.

• Parameter of interest: αj . Null hypothesis H0 : αj = 0. “Jensen’s alpha’’.

2.5 Multivariate regression

yi = α + β1 x1i + β2 x2i + β3 x3i + ... + εi

• Typical trade-off: increasing the number of explanatory variables decreases the

precision of the estimation (wider significance intervals).

• Testing of multiple hypotheses: the F-test.

• The R2 -measure.

13
3 Analysis
3.1 Quick reminder: derivatives
• f 0 (x) > ⇔ f is increasing function.

• f 0 (x) < ⇔ f is decreasing function.

• f 00 (x) > ⇔ f is convex.

• f 00 (x) < ⇔ f is concave.

• f 0 (x) = 0 and f 00 (x) > 0 ⇔ f has a local minimum at x.

• f 0 (x) = 0 and f 00 (x) < 0 ⇔ f has a local maximum at x

3.2 Maximization problem

• The mathematical problem: how to find the value x? , which maximizes a function
f (x)? (potentially, with constraints...)

• Solution:

– First-order condition (FOC)

f 0 (x? ) = 0

∗ Either a maximum, or a minimum. ‘Necessary condition’.

∗ Intuition: f 0 (x) changes sign around the critical value.
– Second-order condition
f 00 (x? ) < 0

Q: Let a, b > 0. Find the maximizer of f (x) = ax − 12 bx2

A:
a
f 0 (x) = a − bx ⇒ x? =
b
f 00 (x) = −b < 0 ⇒ x? is maximizer.

14