Performance Measures

Riccardo Rebonato

EDHEC Business School – EDHEC Risk Institute

1/36
Outline of Lecture

Plan of the Session

Returns By Themselves
Taking Capital Flows into Account

Important Performance Measures

Sharpe Ratio
Treynor Ratio

Non-Gameable Performance Measures

What Do We Mean by Risk?

Next Steps

Appendix – The APT Ratio

2/36
Plan of the Session

In this session we want to

I Understand what different performance measures are
supposed to capture.
I What their theoretical limitations are.
I How performance measures can be gamed, and must therefore
looked at ‘suspiciously’.
I How on can build ‘manipulation-free’ performance measures.

3/36
Returns By Themselves

I Lets’ get started by looking at measures of return.

I In what follows i denotes an asset, a strategy or a portfolio.
The time-t return on i is defined as
Pti − Pt−1
i + Dti
Rti ≡ i
(1)
Pt−1

where Pti is the price of i at time t, and Dti is the cashflow

(dividend, coupon, ...) paid by i at time t.
I So we are splitting the return into
Pti −Pt−1
i
1. capital appreciation (the term i
Pt−1
); and
Dti
2. dividend (the term i
Pt−1
).

4/36
Returns By Themselves

This is valid for a one-period return. What about many periods?

I A bad way to do it is just to compute the many-period return,
i
R a , as the arithmetic mean of 1-period returns:
T
i 1 X i
Ra ≡ Rt (2)
T
t=1

I This is obviously bad if the time intervals are not equal.

I Even if the time step is the same it can be nonsensical (see
example below).
I What else can we do?

5/36
Returns By Themselves

I The one-period logarithmic return is defined by

i,t Pti + Dti

Rlog ≡ log i
(3)
Pt−1

I Since logarithms can be aggregated, this formula allows one

to obtain the return for a period of any length by adding the
returns for the elementary periods.
I This naturally leads to the geometric mean.

6/36
Returns By Themselves
In mathematics, the geometric mean is a mean or average, which indicates the central tendency or typical
value of a set of numbers by using the product of their values (as opposed to the arithmetic mean which
uses their sum). The geometric mean is defined as the nth root of the product of n numbers, i.e., for a set
of numbers x1, x2, ..., xn, the geometric mean is defined as

I The geometric mean (or compound geometric rate of return),

i
R g , allows us to use arithmetic returns to obtain the real rate
of growth of an investment.
I It is given by
" T
# T1
i Y
Rti


Rg ≡ 1+ −1 (4)
t=1

I Let’s see why using arithmetic returns and the compound

geometric rate of return makes sense.

7/36
Returns By Themselves

An extreme example
I Consider an asset whose price changes as follows
I P0 = 100 at t = 0;
I P1 = 200 at t = 1;
I P2 = 100 at t = 2.
Let’s calculate the arithmetic and geometric means.
I Sanity check first: any reasonable measure of two-period
return should give 0 for this example!

8/36
Returns By Themselves
I Let’s start with the arithmetic average. We have
I R1 = P1P−P0 = 200−100
100 = 1.0;
0
I R2 = 100−200
200 = −0.5;
I → R ia = 0.25.
I Next, the geometric mean:
I R ig = [(1 + R1 )(1 + R2 )] 2 − 1;
1

I R ig = [(1 + 1)(1 − 0.5)] 2 − 1;

1

I → R ig = 0.0.
I Last, add the log returns:
1
I Rlog = log PP1i = log 200
100 = 0.693147;
0
2
I Rlog = log PP2i = log 100
200 = −0.693147;
1
1
I Rlog 2
+ Rlog =0
I So, either add the log returns, or calculate the geometric
average of arithmetic returns.

9/36
Returns By Themselves

I This example is extreme (to make a point).

I However, the greater the variation in returns, the bigger the
difference between arithmetic and geometric means.
I The arithmetic mean is always greater or equal to the
geometric mean.
I Note that so far I have implicitly assumed no cash injections
in the portfolio.

10/36
Returns By Themselves – Taking Capital Flows into
Account

I What if the capital committed to a strategy is not constant

over time?
I This is a possible definition of return for any dynamic trading
strategy.
I Let Vt be the value of portfolio at time t, and Ct the cash
injection (positive or negative) at time t.
I Then an approximate formula for the capital weighted return,
RCWR over n periods is given by

VT − V0 − ni=1 Cti
P
RCWR = (5)
V0 + ni=1 T T−ti Cti
P

where Cti is negative if the cashflows is injected.

11/36
Returns By Themselves – Taking Capital Flows into
Account

I You can see that the formula can’t be very good, because
there is no notion of reinvesting/funding the cashflows.
I Its main advantage is that it is an explicit formula.
I A better, but still not perfect, measure is the internal rate of
return, IRR, defined implicitly by
n−1
X Cti VT
V0 = t
+ (6)
(1 + IRR) i (1 + IRR)T
i=1

I Note that the quantities 1

(1+IRR)ti
look like discount factors,
but they are not, because they are not transferable from one
security to the next.

12/36
Important Performance Measures
So far we have looked at returns by themselves.
Let’s now look at performance measures.
I A good place to start is the CAPM model
σi
E [Rti ] = rtf + ρim E [rtm − rtf ] = rtf + βim E [rtm ] (7)
σm
I This can be readjusted to obtain
σi
E [Rti ] − rtf ≡ E [XRti ] = ρim E [rtm − rtf ] (8)
σm
I I can look at this equation as a regression of the excess return
from i against the excess market return (which from now on I
am going to write as xrtm ).
I If the CAPM were true, in this regression there would be no
intercept, and a slope given by σσmi ρim .

13/36
Important Performance Measures

I I can take a purely empirical approach and run a regression of

the excess return against the market return

E [XRti ] = αi + β i E [xrtm ] (9)

I I can look at this equation as a regression of the excess return

from i against the market return.
I If the CAPM were true, in this regression there would be no
intercept, and a slope given by σσmi ρim .
I One can therefore interpret the intercept αi as the share of
return not explained by the market ‘beta’, and attributable to
the manager’s skill.
I It doesn’t tell me anything about how much risk the manager
had to take to earn this alpha.

14/36
Important Performance Measures

How can we take the riskiness of the position into account?

I Here is another useful readjustment of the CAPM model
σi
E [Rti ] = rtf + βim E [xrtm ] = rtf + ρim E [xrtm ] (10)
σm
I This can be written as

σi E [Rti ] − rtf ρim

E [Rti ] − rtf = ρim E [xrtm ] → = E [xrtm ] (11)
σm σi σm
I As we all know, given a strategy or a security, i, the Sharpe
Ratio, SRi is given by

E [Rti ] − rtf
SRi = (12)
σi

15/36
Important Performance Measures
I Therefore we can see that, if the CAPM were true,

E [xrtm ]
SRi = ρim (13)
σm
I This immediately show the limitation of comparing Sharpe
Ratios for different strategies without taking into account of
how these strategies correlate with the market return:
m
I the quantity Eσ[rt ] is common to all securities/strategies;
m
I only the quantity ρim is strategy-dependent.
I So, for instance, a strategy with a low correlation with the
market return should have a low SR.
I And, a strategy with a negative correlation with the market
return should have a negative SR.
I Think: yields of Bunds, US Treasuries, Gilts... and the
‘Greenspan’s put’.
16/36
Important Performance Measures

I These considerations are rarely made when one looks at SRs

to decide which strategy ‘performs better’.
I However, in general, looking at the SR in isolation does not
make a lot of sense.
I It would make more sense
1. in a ‘one-factor world’;
2. when one is looking at long-only equity positions (where
ρim u 1.)
I Can one do better?

17/36
Important Performance Measures
I Let’s go back to

E [Rti ] = rtf + βim E [xrtm ] (14)

I This can be re-adjusted to obtain (in a CAPM world)

E [Rti − rtf ]
Treynor Ratio ≡ = E [xrtm ] (15)
βim
I Note that the RHS is now security-independent.
I If we believe in the CAPM model and can calculate βim
confidently, then we can compare different securities on the
basis of their Treynor Ratio.
I This makes good sense when ρim > 0; careful when ρim = 0 or
ρim < 0: think of current Treasury yields and the Greenspan’s
put!
18/36
Important Performance Measures

I The measures we have looked at so far all imply a CAPM

world.
I We all know about the limitations of CAPM.
I Can one create performance measures that do not rest so
heavily on the validity of the CAPM model?
I There are several ways to do so:
1. either by looking at more ‘successful’ models;
2. or by taking a more ‘fundamental’ approach.
I I deal with the first improvement in the Appendix, where I
introduce the APT model and the APT ratio.
I Next, I am going to look at the second aspect, ie, at the
important topic of non-gameable performance measures.

19/36
Non-Gameable Performance Measures

I We can discuss this (and similar) performance measures at

length, but the measures in this family all have some features
in common:
1. they rely on some pricing model (CAPM, APT) being true;
2. when they make use of a utility function, this is of the
mean-variance type: ‘dislike’ for the variance of the strategy is
primitive, and built into the set-up from the start;
3. they can be ‘gamed’.
I How can one game these measures?

20/36
Non-Gameable Performance Measures

I Here is a very simple example, with the simplest measure of

all, the Sharpe Ratio.
I Consider the strategy of
1. selling an out-of-the-money put;
2. putting the proceeds on deposit;
3. waiting until option expiry.
I With probability π the option ends out-of-the-money. In this
case the SR is infinite.
I With probability 1 − π the option ends in-the-money. In this
case the SR is negative, but finite.
I Hence, the expected SR is infinite.

21/36
Non-Gameable Performance Measures

I The example is contrived, but is shows that dynamic

strategies with leverage and/or very-non-symmetric return
profiles can be manipulated.
I Goetz and Ingersoll show that manipulation of simple
measures is pervasive, and therefore present ‘manipulation
proof’ performance measures.
I The most famous one is the Goetz’s Θ.
I It works as follows.

22/36
Non-Gameable Performance Measures
I Goetz’s Θ is defined as
" T  #
1 X 1 + rt 1−ρ
Θ= log (16)
(1 − ρ)∆t 1 + rtf
t=1

I It can be interpreted as the annualized continuously

compounded excess return certainty equivalent of the
portfolio.
I A risk-free portfolio earning exp[log(1 + rtf ) + Θ∆t)] each
period would have a measured performance of Θ.
I In Eq (16) rt is the non-annualized ∆t-period return from the
strategy, rtf is the non-annualized ∆t-period risk-less rate, ρ is
the risk-aversion coefficient and ∆t is the rebalancing period.1

1
The coefficient ρ should be selected so as to make holding the benchmark
optimal for an uninformed manager. Using a diversified portfolio of stocks as
the benchmark, Goetz et al argue that ρ should be between 2 and 4. 23/36
Non-Gameable Performance Measures

I Goetz’s Θ can be interpreted as follows.

I Expanding to first-order the return one gets

exp[log(1 + rtf ) + Θ∆t)] u (1 + rtf + Θ∆t)) (17)

I This means that a risk-averse investor, with coefficient of risk

aversion ρ, would be prepared to pay an annual fee Θ to an
asset manager to enjoy the benefits of the strategy in
question.

24/36
Non-Gameable Performance Measures

I This measure (and the related Valente measure) have the nice
feature of not being gameable (try and put in the pay-offs the
sold-option scam...)
I It knows about how risk averse the investor is (the SR,
Treynor Ratio and APT Ratio implicitly ‘endorse’ the market’s
degree of risk aversion).
I It can be extended to the case of long/short strategies, and
for unfunded transactions (eg, using swaps).
I However, it has no concept of how the strategy ‘interacts’
with the market portfolio:
I All the wealth of the investor is implicitly invested in the
strategy.

25/36
Next Steps

I Apart from the Goetz’s Θ measure, so far we have always

looked at ‘risk’ as standard deviation.
I However, the same strategy can be perceived as having a very
different risk profile by
1. the trader;
2. the investor;
3. the risk manager;
4. the regulator;
5. the tax payer;...
I All these different agents see the same return profile, but have
different payoff functions.

26/36
Next Steps

I A crude attempt to capture this aspect is made by techniques,

such as VaR, that look at specific percentiles of the return
distribution to make decisions.
I The logic behind VaR as a decision tool is really bad.
I However, the idea of extracting information from the full
distribution of returns is very powerful.
I This leads to the techniques of
1. Empirical Monte Carlo simulation;
2. Parametric Monte Carlo simulation;
3. Historical simulation.
I The key idea is to split the estimation of the return
distribution from the what to do with it.

27/36
Next Steps

I So, (some of) our next steps could be

1. building robust estimates of correlation / covariance matrices;
2. running a really robust Monte Carlo simulation;
3. ’stressing’ a correlation matrix;
4. building a copula...
I What they all have in common is the estimation of the P&L
distribution, from which all risk and performance and utility
measures can be built.
I The logical step is
1. the construction of the high-dimensional joint distribution of
risk factors;
2. the estimation of the P&L sensitivities;
3. The construction of the univariate P&L distribution.

28/36
Appendix – The APT Ratio

I With the APT model one writes

X
Ri (t) = Et [Ri ] + βik fk (t) + i (t) (18)
k=1,n

with
Et [fk (t)] = 0 ∀k (19)
Et [i (t)] = 0 ∀i
covt [i (t) fk (t)] = 0 ∀k, i (20)

29/36
Appendix – The APT Ratio

The APT theorem states the following:

I Given K factors and an asset i, there exist K + 1
asset-independent numbers, λk , not all zero, such that the
expected return of asset i is given by
X
Et [ri ] = λ0 (t) + βik λk . (21)
k=1,K

I The asset-independent quantities λk are usually called market

prices of risk.

30/36
Appendix – The APT Ratio

I The name ‘market price of risk’ is apt.

I To see why, consider first a particular asset, call it asset 0, for
which all the loadings β0k are all zero.
I Such an asset is not exposed to any source of systematic risk
I Therefore it can only earn the risk-less rate.
I Therefore we can write

λ 0 = rf (22)

with rf equal to the riskless rate.

31/36
Appendix – The APT Ratio

I Now we can substitute back Equation (21) into Equation (18)

and collect terms in βik to get
X
Ri (t) = Et [Ri ] + βik fk (t) + i (t) (23)
k=1,n
X X
= rf + βik λk + βik fk (t) +i (t)
k=1,n k=1,n
| {z } | {z }
Et [ri ] systematic random shocks
X  
= rf + βik λ + f (t) + i (24)
(t)
| k {zk } asset-independent
k=1,n

I We are going to use Equation (23) to obtain a performance

measure.

32/36
Appendix – The APT Ratio

I To do so take the riskless rate to the LHS, take expectations,

and you get X
E [Rti − rf ] = βik λk (25)
k
I This can be re-arranged in several ways. For instance

E [R i − rf ]
APT Ratio ≡ P t =1 (26)
k βik λk

I This says that the ratio of the excess return to the sum of the
products of all the security-specific betas and
security-independent market price of risk is the same for all
securities, and equal to 1.

33/36
Appendix – The APT Ratio

I Note what this ratio does.

I The numerator is just the excess return from i.
I The denominator then decomposes the excess return in the
individual compensations for bearing the sources of risk to
which the strategy is exposed.
I The risk components come from the βik = ρσik σi
k

34/36
Appendix – The APT Ratio

I Having a ratio that must be equal to 1 is not very helpful: is

1.02 good? Is 2 good?
I One can re-write the APT ratio as follows:
X X
E [Rti − rf ] = βik λk = βim λm + βik λk (27)
k k6=m

I Now, λm = E [R m ] and therefore

E [Rti − rf ] − k6=m βik λk

P
= E [R m ] (28)
βim

35/36
Appendix – The APT Ratio

I This says that when I

1. subtract from the excess return all the non-market
components, and
2. divide by the ‘market beta’
I I should get the market return for every asset!
I A ‘good deal’ is a strategy for which the adjusted APT ratio is
higher than the market return.
I The ‘units’ now are in percentage points.

36/36