Portfolio Construction

Introduction
Information analysis ignored real world issues.
We now confront those issues directly, especially:
Constraints
Transactions costs
Things are much easier analytically if these issues
didnt exist.

We use portfolio construction in actual

implementation, as well as backtesting of new
investment strategies.

Objective Function/Approach
We have focused on the objective function:
T
T
U P P2 h

V
PA
PA h

PA

Some investors have raised issues with this

approach:
Sensitivity to errors in inputs: error maximization
Alternative risk definitions

We will briefly discuss alternatives

Optimization and Data Errors

How do estimation errors in and V impact
portfolio construction?
forecast true
2
2forecast true

Relative to the true solutions, we will:

Overweight stocks with >0, <0.
Forecast higher alpha and lower risk than justified.

This may still be the best construction approach.

We can adjust expectations for size of the bias.

Mullers Portfolio Construction Test

Muller (1993) compared four alternative methods in
the following lab setting:
Follow S&P 500 stocks monthly from 1984 through 1987,
with annual rebalance. (1987 treatment unusual)
Generate panels of alphas with fixed IC (0.10)

n t IC IC
n t 1 IC 2 n Zn t

Methods:

Screen 1: Top N stocks, equal weighted

Screen 2: Top N stocks, cap weighted
Stratification: Top J stocks in each BARRA industry, match weight
Quadratic programming, long-only, no position>10%.

Test Results: Observed IRs

Date

Jan84

Jan85

Jan86

May87

Average
StandardDeviation
Maximum
Minimum

Risk
Aversion

SCREENI

High
Medium
Low

High
Medium
Low

High
Medium
Low

High
Medium
Low

SCREENII

1.10
0.95
0.73

0.78
0.74
0.50

1.17
0.69
0.60

1.43
1.01
0.66

0.86
0.27
1.43
0.50

STRAT

1.30
2.24
1.31

1.47
0.53
0.15

0.91
0.98
0.99

2.04
1.48
1.17

1.10
0.79
2.24
0.53

Compare IR results to theory (2.24)

QP

0.63
0.64
0.69

1.98
1.29
0.83

0.69
0.33
0.51

2.82
2.60
2.17

1.27
0.89
2.82
0.33

2.16
1.89
1.75

0.98
1.68
1.49

2.08
2.29
2.51

2.14
1.76
1.82

1.88
0.40
2.51
0.98

Bias in Optimization
Mullers analysis shows quadratic
optimization superior to alternative methods
in generating high realized IR portfolios.
In addition, he estimated a typical risk bias
of ~20% for his optimized portfolios. So if
the predicted risk was 3%, the realized risk
averaged about 3.6%.
This depends on the portfolios of interest,
and on the accuracy of the risk model.

Defining Risk

We have discussed this before. Two studies look at how risk

definitions can impact portfolio construction.
Kahn and Stefek (1996) study of asset selection: downside risk
forecasts for equities reduce to standard deviation forecast plus noise,
with standard deviation forecast based only on history. Even investors
with preferences defined by alternative risk measures are better served
by mean/variance analysis.
Grinold (1999) compares mean/variance and scenario-based
approaches to asset allocation in the institutional (benchmark-aware
setting). The two approaches lead to similar portfolios, though the
scenario-based approaches require considerably more work. He
further shows that this approach seldom leads to option investing, even
though those are the investments most requiring the analysis.

Back to Quadratic Optimization

While prior tests show this approach superior to
others on average, it is still susceptible to errors in
input variables.
We can treat this problem in two ways:
Controlling alphas
Adding constraints.

We tend to prefer controlling alphas over adding

constraints, as this provides more transparency into
what we are doing. Both approaches can lead to
the same answer.

Controlling Alphas
Scale
Two equivalent views:
n IC n zn

Std ~ IC

IR T V 1

If input alphas are inconsistent with our

expectations for IC or IR, we can rescale them.
What is a typical equity alpha?

Trimming Outliers
In addition to alpha scale, we can
specifically focus on extreme values.
Here is one ad hoc approach:

3
delete

3 5

Neutralization
Remove unintentional industry, sector, or
market timing bets.
This is different from scaling and trimming.
Neutralization isnt always the right answer.
What economic information does the signal
contain?

Insights into Neutralization

We can think about neutralization in terms
of alphas or portfolios. The connection:

2 V
h PA
Examples

Insight into Neutralization

Benchmark Neutrality: Here are two
possible choices:
B e
B

Both transformations guarantee that the

benchmark has zero alpha.
How do they differ?
What is happening to the optimal portfolio in
each case?

Factor Neutralization
Start with a factor model and factor
portfolios:
r X b u
1

T
1
T
b X
X
X
r1

H rT

Plus our optimal connection between alphas

and portfolios:
2 V
h PA 2 X
F x PA 2

h
PA

Goal
Separate and h into common factor and specific
components:
CF SP
h PA hCF h SP
Such that:

CF 2 V
hCF 2 X
F x PA

SP 2 V
h SP 2 hPA
Challenge: This separation isnt unique. There
are an infinite number of portfolios with active
factor exposures xPA.

Separating the Portfolio

To uniquely define a separation, we stipulate that
hCF is the minimum risk portfolio with factor
exposures xPA. We construct this then from factor
portfolios:

hCF H x PA

We can then show that:

h
X
H
SP I
1

This approach can lead to pinpoint control over

portfolio factor exposures.

Back to Benchmark Neutrality

Our two approaches differed in how we
hedged out the beta of the optimal portfolio.
If we use B e we are hedging with
the minimum risk, fully invested portfolio.
If we use B we are hedging
with the benchmark, the minimum risk =1
portfolio.

Portfolio Construction
Actual optimization isnt as simple as:
T
T
U h

V
PA
PA
PA h

We also have constraints:

Full investment:
Long-only
Position size:
Factor exposure:

And transactions costs.

hTP e 1, hTPA
e 0
hP n 0, n

hPA n hPA n, max

xPA j xPA j , max

Observations on Constraints and Costs

These are equivalent to alpha adjustments.
How?

Some can significantly impact the portfolio.

Now we will develop our intuition for
constraints by analyzing the surprisingly
large impact of the long-only constraint.

Long-only Constraint
This constraint involves the benchmark:
hP n 0 hPA n hB n

Lets develop our intuition with a simple

model:
n IC n zn

IR n zn

IR zn
hPA n
2 n N

We also know that:

IR
2

Simple Model of Long-Only Constraint

Hence we have:
P zn
hPA n
n N
So for an equally-weighted benchmark:
n
1
hPA n zn
N
P N

The constraint is more binding, the higher the

portfolio active risk, the lower the stock residual
risk, and the more stocks in the benchmark

Minimum Score
SensitivitytoPortfolioActiveRisk
0

MinimumScore

1%
2%
3%
4%

5%
6%
7%
8%

4
50

150

250

350

450

550

NumberofAssets

650

750

850

950

This isnt just about negative scores

The long-only constraint interacts with the
full investment constraint.
Active short positions fund active long
positions.

If we limit short positions, this will impact

long positions too.
Example: N=1,000
Compare long-only and long-short
implementations.

Implications for Long Positions

Long/ShortandLongOnlyActivePositions:EqualWeightedBenchmark
2.0%
1.5%
1.0%

ActiveHolding

0.5%
0.0%

LongOnly
Long/Short

0.5%
1.0%
1.5%
2.0%
2.5%
0

100

200

300

400

500
AssetRank

600

700

800

900

Estimating More Realistic Impact

We can only derive analytic results in
overly simplified examples.
Inequality constraints dont lend themselves to
analytic results.
The results depend on benchmark weights.
Assuming equal weighting is not realistic.
We will now analyze the long-only constraint in
more realistic detail. This will require
simulations.

The Framework
We will start with a more realistic N stock cap
weighted benchmark.
More on this shortly.

We will generate random alphas according to:

IR zn
n
N

for many choices of N and P. For each such

choice, we will generate 900 panels of alphas. We
will then analyze subsequent ex ante portfolio
alpha and risk on average over those 900 samples.

Realistic Benchmarks
Capitalization Weighted
Every one is different (S&P 500, Russell
1000, MSCI EAFE, )
But they are typically not far from lognormally distributed.
We will use the log-normal distribution, fit
to typical benchmarks.

Efficient Frontier
10.0%
9.0%
8.0%

ForecastActiveReturn

7.0%
Long/Short

6.0%

N=50
N=100

5.0%

N=250
N=500

4.0%

N=1000

3.0%
2.0%
1.0%
0.0%
0.0%

1.0%

2.0%

3.0%

4.0%

5.0%

ForecastActiveRisk

6.0%

7.0%

8.0%

9.0%

Empirical Fit
We fit the following functional form to the
efficient frontier:
1 N
1 P 1
P , N IR

1 N

0.57
N 53 N

We can also calculate the transfer coefficient:

TC P , N

P , N / P

IR

1 P

1 N

P 1 N

Transfer Coefficient

Sensitivities
Chapter 15 in the book displays the
sensitivity of the transfer coefficient to
average residual risk and correlation
between size and volatility. It also discusses
(p. 434) that changing the lognormal model
coefficient c over the range of interest does
not affect results.
Most important other impact of the longonly constraint is to induce unwanted size
exposures.

Size Bias

Size Bias
The typical US active equity manager has a
smallcap bias. This helps explain why.