Implementing Mean-Variance Theory

Vincent Milhau
vincent.milhau@edhec.edu

Academic year 2024-2025

1 / 49
The Estimation Problem
▶ The tangency portfolio depends on the covariance matrix and
the expected excess returns:

𝚺−1 𝛍̃
𝐰tan =
𝟏′ 𝚺−1 𝛍̃

▶ These parameters are not observed and must be estimated

from past data, so we have an estimated tangency portfolio:

𝚺̂ −1 𝛍̂̃
𝐰̂ tan =
𝟏′ 𝚺̂ −1 𝛍̂̃

𝚺̂ = estimated covariance matrix;

𝛍̂̃ = estimated expected excess returns
▶ How good is this portfolio when we hold it out of the sample
from which the parameters were estimated?

2 / 49
Outline

Scientific Versus Heuristic Diversification

The Sample Average Returns

The Sample Covariance Matrix

Improved Covariance Matrix Estimators

Introducing Weight Constraints

Diversifying Without Expected Returns

3 / 49
Mean-Variance Optimization Versus Equal Weighting
▶ Reference paper: DeMiguel, Garlappi and Uppal (2009). “Optimal
Versus Naive Diversification: How Inefficient is the 1/𝑁 Portfolio
Strategy?” Review of Financial Studies.
▶ The tangency portfolio is estimated with the sample covariance
matrix and the sample average returns.
▶ The portfolio is rebalanced every month based on parameters
estimated from the past 120 months of data.

▶ We compare it with the equally weighted portfolio, whose

weights are
1
𝑤𝑖 = for every 𝑖 = 1, 2, ..., 𝑁
𝑁

4 / 49
Mean-Variance Optimization Versus Equal Weighting
(Con’t)
▶ Excerpt from Table 3 of DeMiguel et al. (2009):

▶ The equally weighted portfolio has higher Sharpe ratio than the
mean-variance efficient portfolio in 5 out of 6 investment
universes!
▶ “mv (in sample)” is the tangency portfolio estimated with the
full sample of returns (unrealistic!)
▶ Equal weighting (“naive diversification”) performs often much
better than mean-variance (“scientific diversification”)...
5 / 49
How to Improve Mean-Variance Efficient Portfolios?
▶ Conclusion: Estimation errors penalize severely mean-variance
efficient portfolios.
▶ Approach 1: Improve inputs (expected return and covariance
estimates):
▶ Why do sample estimates perform poorly?
▶ Search for better parameter estimates.

▶ Approach 2: Regularize outputs (weights):

▶ Introduce weight constraints in mean-variance optimization to
make efficient portfolios closer to equally weighted.
▶ Approach 3: Reduce the number of parameters to estimate:
▶ Use parsimonious diversification methods without expected
returns, or without correlations, or without volatilities.

6 / 49
Outline

Scientific Versus Heuristic Diversification

The Sample Average Returns

The Sample Covariance Matrix

Improved Covariance Matrix Estimators

Introducing Weight Constraints

Diversifying Without Expected Returns

7 / 49
Outline

Scientific Versus Heuristic Diversification

The Sample Average Returns

The Sample Covariance Matrix

Improved Covariance Matrix Estimators

Introducing Weight Constraints

Diversifying Without Expected Returns

8 / 49
Sample Mean
▶ A natural estimate of the expected return 𝜇𝑖 is the sample mean
of past returns.

Definition
The sample mean of returns for asset 𝑖 is

𝑇
1
𝜇̂ 𝑖 = 𝑟𝑖̄ = ∑𝑟
𝑇 𝑡=1 𝑡,𝑖

▶ How good is this as an estimate of the true expected return 𝜇𝑖 ?

9 / 49
The Sample Mean is Unbiased
Property
Assume that, for all date 𝑡,

𝔼 [𝑟𝑡,𝑖 ] = 𝜇𝑖

Then, 𝜇̂ 𝑖 is an unbiased estimator of 𝜇𝑖 .

▶ Interpretation: the average of sample means across multiple

independent samples with the same size 𝑇 converges to the
true expected return as the number of samples grows to infinity
(for fixed 𝑇).

10 / 49
Confidence Interval for the Expected Return
Property
Assume that the returns

𝑟0,𝑖 , 𝑟1,𝑖 , ..., 𝑟𝑡,𝑖 , ...

are independent and identically distributed with mean 𝜇𝑖 and

2
variance 𝜎𝑖 .
Then, an asymptotic 95% confidence interval for the expected return
𝜇𝑖 estimated from 𝑇 returns is [ℓb𝑖 , ub𝑖 ], where

𝜎̂ 𝑖 𝜎̂ 𝑖
ℓb𝑖 = 𝜇̂ 𝑖 − 1.96 , ub𝑖 = 𝜇̂ 𝑖 + 1.96
√𝑇 √𝑇

and 𝜎̂ 𝑖 is the sample volatility of returns.

▶ Interpretation: the true expected return lies within the

confidence bounds with a probability that converges to 95% as
the sample size 𝑇 grows to infinity.
11 / 49
Empirical Application
Implementation Exercise

The workbook Expected_returns.xlsx contains monthly quotes

for 10 US stocks from 12.2018 to 12.2023.
1. Calculate the sample average of simple returns, the
sample volatility and the 95% confidence bounds.
2. Calculate the annualized sample averages and confidence
bounds, defined as

𝑟ann = [1 + 𝑟]12 − 1

where 𝑟 stands for 𝜇̂ 𝑖 , ℓb𝑖 or ub𝑖 .

Solution in Expected_returns_solution.xlsx.

12 / 49
Annualized Returns with Confidence Bounds

▶ Confidence bounds are very wide.

▶ Past average returns provide little information about expected
returns.

13 / 49
Estimating Expected Returns in Practice
▶ Approach 1: postulate some risk - expected return relationship.
▶ Scientific Beta Efficient Maximum Sharpe Ratio Indices assume
that expected excess returns are proportional to semivolatilities;
▶ TOBAM’s Maximum Diversification Approach assumes that
expected excess returns are proportional to volatilities (see
foundation paper by Choueifaty and Coignard (2008). “Towards
Maximum Diversification”. Journal of Portfolio Management);
▶ Use a multifactor model, where expected returns are functions of
betas (but this may still lead to unrealistic estimates).
▶ Approach 2 (beyond the scope of this course): incorporate prior
views about expected returns.
▶ The Black-Litterman model (Black and Litterman (1992), “Global
Portfolio Optimization”. Financial Analysts Journal) mixes investor’s
views with expected returns inferred from the market portfolio.
▶ See workbook Black-Litterman.xlsx.

▶ Approach 3: diversify without expected returns!

14 / 49
Outline

Scientific Versus Heuristic Diversification

The Sample Average Returns

The Sample Covariance Matrix

Improved Covariance Matrix Estimators

Introducing Weight Constraints

Diversifying Without Expected Returns

15 / 49
Sample Covariance
▶ Reminder: the covariance between the returns on two assets 𝑖
and 𝑗 is
ℂov[𝑟𝑖 , 𝑟𝑗 ] = 𝔼 [[𝑟𝑖 − 𝜇𝑖 ] [𝑟𝑗 − 𝜇𝑗 ]]

▶ A natural estimate is obtained by replacing expectations with

the sample means.

Definition
Consider two assets 𝑖 and 𝑗. The sample covariance is

𝑇
1
𝜎̂ 𝑖𝑗 = ∑ [𝑟 − 𝑟 ̄ ] [𝑟 − 𝑟 ̄ ]
𝑇 𝑡=1 𝑡,𝑖 𝑖 𝑡,𝑗 𝑗

where 𝑟𝑖̄ and 𝑟𝑗̄ are the sample means.

▶ Notations for the sample covariance matrix: 𝐒 or 𝚺.̂

16 / 49
Bias of Sample Covariance
Property
Consider two assets 𝑖 and 𝑗. If, for any two dates 𝑠 and 𝑡,

ℂov [𝑟𝑡,𝑖 , 𝑟𝑡,𝑗 ] = 𝜎𝑖𝑗 (constant covariance),

ℂov [𝑟𝑡,𝑖 , 𝑟𝑠,𝑗 ] = 0 (zero time series covariance),

𝔼 [𝑟𝑡,𝑗 ] = 𝜇𝑗 , 𝔼 [𝑟𝑡,𝑖 ] = 𝜇𝑖 (constant expected returns),

then 𝜎̂ 𝑖𝑗 is a biased estimator of 𝜎𝑖𝑗 :

𝑇 −1
𝔼 [𝜎̂ 𝑖𝑗 ] = 𝜎𝑖𝑗
𝑇

▶ An unbiased estimator is
𝑇
𝑇 1
𝜎̂ 𝑖𝑗 = ∑ [𝑟 − 𝑟 ̄ ] [𝑟 − 𝑟 ̄ ]
𝑇 −1 𝑇 − 1 𝑡=1 𝑡,𝑖 𝑖 𝑡,𝑗 𝑗

17 / 49
Sample Covariance in Excel
▶ See workbook Covariance_matrix.xlsx.

▶ The lower part of the biased covariance matrix can be obtained

with the Data Analysis Tool.

▶ Alternatively, we can write a formula with Excel functions:

▶ Covariance normalized by 𝑇: COVARIANCE.P;
▶ Covariance normalized by 𝑇 − 1: COVARIANCE.S;
▶ Function OFFSET.

18 / 49
Confidence Interval for the True Covariance
Property
Consider two assets 𝑖 and 𝑗, and assume that the pairs of returns

(𝑟1,𝑖 , 𝑟1,𝑗 ), (𝑟2,𝑖 , 𝑟2,𝑗 ), ..., (𝑟𝑡,𝑖 , 𝑟𝑡,𝑗 ), ...

are independent and identically distributed.

Then, an asymptotic confidence interval at the 95% confidence level
for 𝜎𝑖𝑗 is [ℓb𝑖𝑗 , ub𝑖𝑗 ], where

𝜋̂ 𝑖𝑗 𝜋̂ 𝑖𝑗
ℓb𝑖𝑗 = 𝜎̂ 𝑖𝑗 − 1.96√ , ub𝑖𝑗 = 𝜎̂ 𝑖𝑗 + 1.96√
𝑇 𝑇
𝑇 2
1
𝜋̂ 𝑖𝑗 = ∑ [[𝑟 − 𝑟 ̄ ] [𝑟 − 𝑟 ̄ ] − 𝜎̂ 𝑖𝑗 ]
𝑇 𝑡=1 𝑡,𝑖 𝑖 𝑡,𝑗 𝑗

and 𝜎̂ 𝑖𝑗 is the sample covariance.

19 / 49
Impact of Sampling Frequency on Precision
▶ In the Python notebook estimation.ipynb, we calculate the
“precision”, defined as the absolute sample covariance divided
by the bandwidth of the confidence interval:

|𝜎̂ 𝑖𝑗 |
ub𝑖𝑗 − ℓb𝑖𝑗

▶ The narrower the confidence interval, the greater the precision.

▶ We also calculate the precision of the sample average return.

20 / 49
Empirical Application
▶ Distribution of precision figures across the 10 × 11/2 = 55
covariances for 10 stocks:
Precision of sample covariance
4.0

3.5

3.0

2.5

2.0

1.5

1.0

0.5
Monthly Weekly Daily

▶ Increasing the sampling frequency (monthly → weekly → daily)

leads to a more accurate estimate of the sample covariance but
has no noticeable impact on the precision of the sample
average.

21 / 49
How Many Covariances Do We Have to Estimate?
Property
For a 𝑁-asset investment universe, the total number of covariances
and variances to estimate is
𝑁[𝑁 + 1]
2

▶ In detail, there are 𝑁[𝑁 − 1]/2 covariances between distinct

assets, plus 𝑁 variances.

22 / 49
The Curse of Dimensionality
▶ The number of covariances increases faster than the number of
assets.
Number of assets Number of covariances
𝑁 𝑁[𝑁 + 1]/2
5 15
50 1,275
500 125,250
2,000 20,010,000

▶ If we use 2 years of weekly data to estimate the covariance

matrix of 500 assets, we only have

2 × 52 × 500 = 52, 000

datapoints to estimate 125,250 parameters.

23 / 49
Which Estimation Errors Matter More? Expected
Returns or Covariances?
▶ Kan and Zhou (2007). “Optimal Portfolio Choice with Parameter
Uncertainty”. Journal of Financial and Quantitative Analysis.

▶ 𝑁 = universe size .

▶ 𝑇 = sample size .

▶
Estimation errors in expected returns
have in general more impact
than errors in covariances .
▶ But errors in covariances
have a growing impact as
𝑁/𝑇 increases.

24 / 49
Summary
▶ The estimated tangency portfolio is affected by estimation
errors.
▶ Although they are unbiased, sample average returns are very
inaccurate estimates of the true expected returns.
▶ The sample covariance matrix is subject to the curse of
dimensionality as 𝑁/𝑇 is large.

25 / 49
Outline

Scientific Versus Heuristic Diversification

The Sample Average Returns

The Sample Covariance Matrix

Improved Covariance Matrix Estimators

Introducing Weight Constraints

Diversifying Without Expected Returns

26 / 49
Structured Estimator 1: The Single Index Model
▶ Sharpe (1963). “A Simplified Model For Portfolio Analysis”.
Management Science.
▶ Assume that returns are generated by the one-factor model:

𝑟𝑡,𝑖 = 𝑐𝑖 + 𝛽𝑖 𝐹𝑡 + 𝜀𝑡,𝑖

𝐹𝑡 = common risk factor;

𝜀𝑡,𝑖 = residual (idiosyncratic return)
with the conditions:

𝔼 [𝜀𝑡,𝑖 ] = 0 residuals are centered

ℂov [𝜀𝑡,𝑖 , 𝐹𝑡 ] = 0 residuals are uncorrelated from the factor

ℂov [𝜀𝑡,𝑖 , 𝜀𝑡,𝑗 ] = 0 residuals are uncorrelated in cross section

27 / 49
Covariance Matrix Decomposition
▶ In the single factor index model, the covariance matrix of
returns can be decomposed as

𝚺 = 𝜎𝐹2 𝛃𝛃′ + 𝚺𝜀

where
2
𝜎𝜀,1 0 ⋯ 0
𝛽1 2
0 𝜎𝜀,2 ⋯ 0
𝛃 = [⋯], 𝚺𝜀 = [ ]
⋯
𝛽𝑁 2
0 0 ⋯ 𝜎𝜀,𝑁

and 𝜎𝐹2 is the factor variance.

▶ Extension to a model with 𝐾 factors:

𝚺 = 𝛃𝛀𝛃′ + 𝚺𝜀

𝛀 = factor covariance matrix;

𝛃 = matrix of betas
28 / 49
Estimating the Covariance Matrix with the Single
Index Model
Implementation Exercise

The workbook Covariance_matrix.xlsx contains the monthly re-

turns on 10 US stocks from 12.2018 to 12.2023. We want to esti-
mate the covariance matrix with the single index model, taking
the market factor of Ken French’s library as the common risk
factor. In the tab “Single index”,
1. Calculate the variance of the factor.
2. Calculate the beta and the R-squared of the returns on
each stock with respect to the common factor, by using
the Excel functions SLOPE and RSQ.
3. Calculate the total variance of each stock with the
function STDEV.P.
4. Calculate the systematic covariance matrix, 𝜎𝐹2 𝛃𝛃′ .
5. Calculate the single-index covariance matrix, as given in
slide 28.
Solution in workbook Covariance_matrix_solution.xlsx. 29 / 49
Structured Estimator 2: The Constant Correlation
Model
▶ Elton and Gruber (1973). “Estimating the Dependence Structure
of Share Prices”. Journal of Finance.
▶ The model described here is named “mean model” in EG73: we
assume that all pairs of assets have the same correlation.
▶ We have 𝑁 variances to estimate plus 1 correlation, hence 𝑁 + 1
parameters.
▶ See workbook Covariance_matrix.xlsx, where we estimate the
common correlation as the average sample correlation
(excluding the self correlations).

30 / 49
Bias versus Robustness Tradeoff
▶ Structured estimators are more parsimonious than the sample
one:
Number of independent parameters to estimate
Universe size Sample covariance Single index Constant
matrix model correlation
𝑁 𝑁[𝑁 + 1]/2 2𝑁 + 1 𝑁+1
5 15 11 6
50 1,275 101 51
500 125,250 1,001 501
2,000 20,010,000 4,001 2,001
hence they tend to be more robust, i.e., less sensitive to the
choice of the sample.
▶ However, they are biased if the assumption about the structure
of returns is wrong, while the sample estimator is unbiased.
▶ Next level (beyond the scope of this course): get the best of
both worlds by combining the sample estimator with a
structured one.
▶ This is called linear shrinkage.
▶ Ledoit and Wolf (2003). “Improved estimation of the covariance
matrix of stock returns with an application to portfolio selection”. 31 / 49
Comparing Estimators
▶ A comparative study: Coqueret and Milhau (2014). “Estimating
Covariance Matrices for Portfolio Optimization”. Scientific Beta
White Paper.
▶ Methodology:
1. Estimate the covariance matrix over the 𝑇 past weeks and get an
estimate 𝚺;̂
2. Calculate the global minimum variance portfolio and hold it for 3
months:
𝚺̂ −1 𝟏
𝐰gmv = ;
𝟏′ 𝚺̂ −1 𝟏
3. Re-estimate the covariance matrix and loop through steps 1 and 2
until the sample end;
4. Report the volatility of the GMV strategy.

▶ The lower the OOS volatility, the better the estimation method.

32 / 49
Investment Universes
▶ Several investment universes are considered to check
robustness of the findings and analyze the impact of the
universe size (𝑁) .
▶ Definition of universes in CM14:

33 / 49
How Do the Structured Estimators Perform? (Con’t)
▶ Excerpt of Table 6 in CM14 (volatilities of estimated GMV
portfolios):

▶ FEW: single factor estimate with equally weighted index; FCW:

single factor estimate with cap-weighted index.
▶ ID: equally weighted portfolio.
▶ When 𝑁/𝑇 is “large” , the structured estimators perform
better than the sample one .
34 / 49
Outline

Scientific Versus Heuristic Diversification

The Sample Average Returns

The Sample Covariance Matrix

Improved Covariance Matrix Estimators

Introducing Weight Constraints

Diversifying Without Expected Returns

35 / 49
Short Positions in Mean-Variance Efficient Portfolios
▶ See the workbook Weight_constraints.xlsx, where the tangency
portfolio is
345.0% A
−341.1% B
𝐰tan = [ 125.2% ] C
114.1% D
−143.2% E

▶ Total leverage (absolute sum of short positions): 484.3%.

▶ Short positions may be costly to implement because of
collateral requirements.
▶ To calculate the tangency portfolio subject to a long-only
constraint, we must solve

𝜇𝑝 (𝐰) − 𝑅𝑓
max subject to 𝐰 ′ 𝟏 = 1 and 𝐰 ≥ 𝟎
𝐰 𝜎𝑝 (𝐰)

36 / 49
Tangency Portfolio Under Long-Only Constraints
▶ Tick the box “Make Unconstrained Variables Nonnegative” in the
Solver to remove short positions.

Implementation Exercise

In the workbook Weight_constraints.xlsx, calculate the tangency

portfolio (maximum Sharpe ratio portfolio) with the Solver sub-
ject to the long-only constraint (and the budget constraint).
Calculate the expected excess returns, the volatilities and the
Sharpe ratios of the two tangency portfolios. What is the op-
portunity cost of imposing a long-only constraint in Sharpe ra-
tio maximization?
Solution in Weight_constraints_solution.xlsx.

37 / 49
Effect of Long-Only Constraints on the Tangency
Portfolio
▶ Jagannathan and Ma (2003). “Risk Reduction in Large Portfolios:
Why Imposing the Wrong Constraints Helps”. Journal of Finance.
▶ In large universes (e.g., 500 stocks), long-only constraints
improve the out-of-sample Sharpe ratio for all covariance
matrix estimators. Excerpt from Table VIII in JM03:

▶ The long-only constraints have a cost in sample.

▶ But the equally weighted portfolio is hard to beat!
38 / 49
Effect of Long-Only Constraints on the Minimum
Variance Portfolio
▶ Long-only constraints also reduce the out-of-sample volatility
of the estimated global minimum variance portfolio, at least if
we use the sample covariance matrix.
▶ Excerpt from Table VI in JM03 (“C” = long-only):

▶ If we use an improved estimator (3-factor or Ledoit’s shrinkage),

the constraints have much less effect.
▶ An upper bound of 2% (“D”) has little effect or a risk-increasing
effect.
39 / 49
Outline

Scientific Versus Heuristic Diversification

The Sample Average Returns

The Sample Covariance Matrix

Improved Covariance Matrix Estimators

Introducing Weight Constraints

Diversifying Without Expected Returns

40 / 49
Alternative Diversification Methods
▶ See the Benchmark Builder of Scientific Beta:
▶ Tangency portfolio * ;
▶
Global minimum variance
;
(GMV) portfolio
▶
Equally weighted portfolio,
or portfolio that minimizes
;
the dispersion of weights
across constituents
▶
GMV portfolio assuming
;
equal volatilities

▶ Inverse volatility portfolio ;

▶ Div. Risk Weighted + Max Decon. + Max Decor. + Max Sharpe * .
* Requires expected return estimates.
41 / 49
From Tangency Portfolio to Global Minimum Variance
and Equally Weighted
▶ Reminder: the tangency portfolio maximizes the Sharpe ratio

𝜇𝑝 (𝐰) − 𝑅𝑓
𝜎𝑝 (𝐰)

subject to the budget constraint 𝐰 ′ 𝟏 = 1.

▶ If all expected returns are equal, then the tangency portfolio
coincides with the global minimum variance portfolio.
▶ If we assume further that all covariances (including variances)
are equal, then the tangency portfolio is the equally weighted
portfolio.
▶ We can use the GMV or EW portfolio as a prior for the tangency
portfolio if we assume that all constituents have the same
expected return.

42 / 49
Another Diversification Method: Risk Parity
▶ Maillard, Roncalli and Teiletche (2010). “The Properties of
Equally Weighted Risk Contribution Portfolios”. Journal of
Portfolio Management.
▶ Motivation: An equally weighted portfolio of stocks and bonds is
perfectly diversified in terms of dollars, but most of the variance
comes from stocks, which are much more volatile than bonds.

Weights Risk contributions

B
S B
S

▶ How should we weight the constituents so that they contribute

equally to portfolio variance?
43 / 49
Risk Contributions Within a Portfolio
▶ Consider an equally weighted portfolio of stocks and bonds with
volatilities
𝜎1 = 20%, 𝜎2 = 5%
and a correlation 𝜌12 = 2%.
▶ Calculate the portfolio variance:
𝜎𝑝2 = 𝑤12 𝜎12 + 𝑤22 𝜎22 + 2𝑤1 𝑤2 𝜎12
▶ Re-arrange terms to have a contribution of stocks and a
contribution of bonds:
2 2
𝜎𝑝2 = ⏟⏟⏟⏟⏟⏟⏟⏟⏟
𝑤1 [𝑤1 𝜎1 + 𝑤2 𝜎12 ] + ⏟⏟⏟⏟⏟⏟⏟⏟⏟
𝑤2 [𝑤2 𝜎2 + 𝑤1 𝜎12 ]
𝑐1 𝑐2

▶ Calculate risk contributions (see workbook

Risk_contributions.xlsx):
Stock contribution (𝑐1 ) 0.01005
Bond contribution (𝑐2 ) + 0.000675
Portfolio variance (𝜎𝑝2 ) 0.010725
44 / 49
Risk Parity with 2 Assets
▶ Reminder: portfolio variance with 2 constituents:

2 2
𝜎𝑝2 = 𝑤1 [𝑤1 𝜎1 + 𝑤2 𝜎12 ] + 𝑤2 [𝑤2 𝜎2 + 𝑤1 𝜎12 ]
⏟⏟⏟⏟⏟⏟⏟⏟⏟ ⏟⏟⏟⏟⏟⏟⏟⏟⏟
𝑐1 𝑐2

▶ To have equal risk contributions, we want

𝑐1 = 𝑐2
𝑤1 + 𝑤2 = 1

Definition
The risk parity portfolio is such that all constituents have the same
contribution to the portfolio variance.

45 / 49
Risk Parity with 2 Assets (Con’t)
▶ It can be shown (left as an exercise) that the risk parity portfolio
of 2 assets is the inverse volatility portfolio

1 1
𝑤1 = 𝑧 × , 𝑤2 = 𝑧 ×
𝜎1 𝜎2

with
−1
1 1
𝑧=[ + ]
𝜎1 𝜎2
▶ Exercise: calculate the stock and bond weights in the risk parity
portfolio if the stock volatility is 𝜎1 = 20% and the bond
volatility is 𝜎2 = 5%.

46 / 49
Risk Parity with 3 Assets
▶ Portfolio variance with 3 assets:
𝜎𝑝2 = 𝑤12 𝜎12 + 𝑤22 𝜎22 + 𝑤32 𝜎32 + 2𝑤1 𝑤2 𝜎12 + 2𝑤1 𝑤3 𝜎13 + 2𝑤2 𝑤3 𝜎23
▶ Re-arrange terms to have contributions of assets 1, 2 and 3:
2
𝜎𝑝2 = 𝑤1 [𝑤1 𝜎1 + 𝑤2 𝜎12 + 𝑤3 𝜎13 ]
⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟
𝑐1
2
+ 𝑤2 [𝑤2 𝜎2 + 𝑤1 𝜎12 + 𝑤3 𝜎23 ]
⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟
𝑐2
2
+ 𝑤3 [𝑤3 𝜎3 + 𝑤1 𝜎13 + 𝑤2 𝜎23 ]
⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟
𝑐3

▶ To have equal risk contributions (risk parity), we want

𝑐1 = 𝑐2 = 𝑐3 and 𝑤1 + 𝑤2 + 𝑤3 = 1
▶ There is no mathematical expression for the weights, so we must
use a numerical solution (see Activity).
47 / 49
Summary and Perspectives
▶ Mean-variance optimization is conceptually simple and
mathematically tractable, at least if we allow for short
positions:
▶ Fund separation theorems are available.

▶ Implementation is difficult because of estimation errors in the

covariance matrix and the expected returns.
▶ Possible remedies:
▶ Improve covariance matrix estimates by assuming some structure
in returns.
▶ Next level: linear shrinkage.
▶ Improve expected return estimates?
▶ Very difficult!
▶ Incorporate prior views, e.g. with the Black-Litterman model.

48 / 49
Summary and Perspectives (Con’t)
▶ Possible remedies (con’t):
▶ Impose long-only constraints.
▶ Next level: impose norm constraints to have more diversified
portfolios (DeMiguel, Garlappi, Nogales and Uppal (2009). “A
Generalized Approach to Portfolio Optimization: Improving
Performance by Constraining Portfolio Norms”. Management Science).
▶ Deviate from Sharpe ratio maximization (GMV, equally weighted,
risk parity...)
▶ Proxy the tangency portfolio as a combination of
factor-replicating portfolios.

49 / 49