Estimation

Nuno Clara

Nuno Clara 1 / 39

Investments (2024)
Estimation

▶ Portfolio optimization, choosing managers requires estimating

properties of returns
▶ How good are those estimates?
▶ How does estimation error affect optimization?

Nuno Clara 2 / 39

Investments (2024)
 ▶ Basic goal: go through what we’ve done so far (portfolio choice,
max SR), think about estimation (also important for next classes on
regressions, beating the market ...)
1. How does estimation affect optimal portfolios?
▶ How much data do you need?
▶ Robust versions of M-V optimization

Nuno Clara 3 / 39

Investments (2024)
An example: Health Services Industry

Exc Mean = 0.120

Std. Deviation = 0.262
SR = 0.458

Nuno Clara 4 / 39

Investments (2024)
An example: Health Services Industry

Exc Mean = 0.076

Std. Deviation = 0.254
SR = 0.297

Nuno Clara 5 / 39

Investments (2024)
An example: Health Services Industry

Exc. Mean = 0.083

Std. Deviation = 0.235
SR = 0.352

Nuno Clara 6 / 39

Investments (2024)
Measurement

▶ Measurement is hard:
▶ You need the expected return going forward
▶ How long should the sample be?
▶ Was the sample you looked at weird?

Nuno Clara 7 / 39

Investments (2024)
 How much data do you need?

Nuno Clara 8 / 39

Investments (2024)
 ▶ The standard error measures the standard deviation of the error in
an estimate
▶ Estimate is within ±1.96 standard errors of the truth 95% of the
time
▶ If you have no other information, then 95% confidence interval
(point estimate ±1.96 standard errors) is the range where there’s a
95% chance the true parameter lies

Nuno Clara 9 / 39

Investments (2024)
Standard error of the mean

▶ The standard error for an estimate of a mean like a mean return - is:
σ
r) = √
SE (¯
T
where T is the number of time periods you have

Nuno Clara 10 / 39

Investments (2024)
Standard error of the mean

▶ Consider total US stock market between 1928-2022:

▶ Tyears = 95
▶ Average excess return of stocks over rf : r¯annual = 0.082
▶ with a standard deviation of return of: σannual = 0.197
▶ Thus a historical Sharpe ratio of: SR = 0.416
▶ Then:
σannual 0.197
SE = √ = √ = 2.02%
T 95
▶ What is the confidence interval for the estimated average excess
return?
8.2% ± 1.96 × 2.02% = [4.23% − 12.16%]

Nuno Clara 11 / 39

Investments (2024)
What is the t-statistic?

▶ Recall: The t-statistics is calculated as the estimated mean divided

by the standard error of the mean. We have mean of 8.2% and SE
of 2.02%, in this case:

t − stat = 8.2/2.02 = 4.06

▶ Generally, if t-stat is higher than 1.96 or lower than -1.96 then the
mean is statistically different from 0.

Nuno Clara 12 / 39

Investments (2024)
Standard error of the mean

▶ For the mean, higher-frequency data does not help

▶ Suppose we use monthly data

σmonthly
SE (12¯
rmonthly ) = 12SE (¯
rmonthly ) = 12 p
12Tyears
√
σannual / 12 σannual
= 12 p =p
12Tyears Tyears

▶ Monthly data used to estimate annual returns gives the exact same
standard error as annual data!
▶ It’s not the number of observations that matters, it’s the total length
of the time series

Nuno Clara 13 / 39

Investments (2024)
Standard error of the standard deviation

▶ For Normally distributed data (similar results more generally):

σannual
SE (σannual ) = p
2Tyears

▶ Consider S&P 500: Tyears = 95, σannual = 0.197. Then

SE = 0.0143,
▶ This is pretty tight - 95%CI = [0.17 to 0.22]

Nuno Clara 14 / 39

Investments (2024)
Standard error of the standard deviation

▶ Suppose we go from annual to monthly

√ √ √
 
σmonthly
SE 12 × σmonthly = 12×SE (σmonthly ) = 12 p =
| {z } 2 × 12Tyears
Annualized SD
√
√ σannual / 12 1 σ
= 12 p =√ p annual
2 × 12Tyears 12 2 × Tyears
| {z }
SE from annual data
▶ Going to monthly data helps substantially reduces SE by factor of 3.5

Nuno Clara 15 / 39

Investments (2024)
Estimating means vs standard deviation

Figure: 5-year rolling windows

Nuno Clara 16 / 39

Investments (2024)
Estimating means vs standard deviation

Figure: 10-year rolling windows

Nuno Clara 17 / 39

Investments (2024)
Estimating means vs standard deviation

Figure: 20-year rolling windows

Nuno Clara 18 / 39

Investments (2024)
Estimation of means and standard deviations

▶ Conclusions:
1. Means require a huge amount of data to estimate accurately, at least
for equities
▶ Same for Sharpe ratio - depends only on sample length (1/ Tyears )
p

2. Standard deviations can be measured very accurately if you use

high-frequency data
▶ Same result holds for covariances
▶ Caution: this assumes means and standard deviations are stable over
time (they’re not...)

Nuno Clara 19 / 39

Investments (2024)
Estimation issues

▶ Historical data isn’t always a good guide

▶ Are we conffident means and covariances are stable?
▶ Past high returns might mean valuations too high ) future expected
returns are low
▶ A given time period is less representative for more volatile assets
▶ More volatility means we need more data
▶ Common way for salespeople to try to fool you
▶ Don’t just use past data; use your own forecasts of the future

Nuno Clara 20 / 39

Investments (2024)
 ▶ MV portfolios sometimes called ”error-maximizing portfolios”
▶ Small changes in inputs can lead to extremely different portfolios
▶ Errors are easily magnified - garbage in/garbage out, but worse
▶ If there’s even a little garbage, MV will focus on it
▶ E.g. one Sharpe ratio is overestimated
▶ Examples: assets with high correlations or high Sharpe ratios
▶ You can do this in Excel

Nuno Clara 21 / 39

Investments (2024)
 Popular alternatives to full
mean-variance

Nuno Clara 22 / 39

Investments (2024)
Alternatives to mean-variance optimization

▶ Want to maximize the SR, but we might not totally trust our
estimates
▶ What can we do?
1. Minimum-variance
2. Risk parity
3. Equal weights
▶ Common feature: these methods don’t estimate means

Nuno Clara 23 / 39

Investments (2024)
Alternatives to mean-variance optimization (I)

▶ Minimum-variance portfolios:
▶ Always take the farthest-left point on the M-SD frontier
▶ Optimal when all expected returns are the same
▶ That is, use it when you don’t think you can forecast returns
▶ Still requires knowing all covariances

Nuno Clara 24 / 39

Investments (2024)
Alternatives to mean-variance optimization (II)

▶ Risk parity (variances)

▶ Each asset gets weight proportional to 1/σi2
▶ Contribution to risk from each asset is equalized
▶ Optimal when covariances are zero and expected returns are all the
same
▶ Use it when you can’t forecast returns or correlations
▶ Good: need lots of data for means
▶ Bad: correlations go to 1 in crashes

Nuno Clara 25 / 39

Investments (2024)
Alternatives to mean-variance optimization (II)

▶ Risk parity (standard deviations)

▶ Each asset gets weight proportional to 1/σi
▶ Optimal when covariances are zero and Sharpe Ratios are all the
same
▶ In practice very close to risk parity with variances

Nuno Clara 26 / 39

Investments (2024)
Alternatives to mean-variance optimization (III)

▶ Equal weights:
▶ Every asset gets a weight of 1/N
▶ This is not the same as value weighting - puts more weight on small
asset classes
▶ Optimal if returns uncorrelated and means proportional to variances
rj = kσj2 )
(¯
▶ So use 1/N across factors/asset classes (uncorrelated things, ideally)
▶ Requires no estimation
▶ Very robust; easy to implement

Nuno Clara 27 / 39

Investments (2024)
Weights in the HMC example

Nuno Clara 28 / 39

Investments (2024) Figure: Weights in HMC example

Cummulative returns four asset classes

Figure: Assets: U.S. government bonds, U.S. corporate bonds, U.S. stocks, and
international stocks

Nuno Clara 29 / 39

Investments (2024)
Simulations

▶ One way to test portfolio strategies: simulate returns

▶ Assume means, standard deviations, correlations; computer can
simulate samples of data
▶ Then imagine a person who looks at the simulated data (which is
noisy), estimates means, SDs, etc., chooses a portfolio
▶ Give them 42 years of monthly data
▶ Does it matter whether we give them monthly or annual data?
▶ How do MV, risk parity, minimum-variance, and 1/N do?

Nuno Clara 30 / 39

Investments (2024)
Mean-Variance Portfolios

Figure: Mean-Variance Porfolios

▶ Blut dots are Mean-variance portfolios

▶ Compare with the tangency portfolio ...

Nuno Clara 31 / 39

Investments (2024)
Risk-Parity Portfolios

Figure: Red dots are Risk-Parity Portfolios

Nuno Clara 32 / 39

Investments (2024)
1/N portfolio

Figure: Red dots are 1/N portfolio

Nuno Clara 33 / 39

Investments (2024)
Minimum Variance portfolio

Figure: Red dots are Minimum Variance portfolio

Nuno Clara 34 / 39

Investments (2024)
Overall Compare

Figure: Comparison of different strategies

Nuno Clara 35 / 39

Investments (2024)
Sharpe Ratios

Figure: Comparison SR of different strategies

Nuno Clara 36 / 39

Investments (2024)
Estimates

▶ Suppose you made a mistake and your expected return on one asset
is way too high (e.g. you were excited about computers in 1999)
▶ What will Mean-Variance optimization do to its weight?
▶ What will risk parity do?
▶ What will equal weights do?

Nuno Clara 37 / 39

Investments (2024)
Estimates

▶ Suppose you made a mistake and your variance on one asset is way
too low (e.g. you thought mortgage-backed securities would always
be safe)
▶ What will Mean-Variance optimization do to its weight?
▶ What will risk parity do?
▶ What will equal weights do?

Nuno Clara 38 / 39

Investments (2024)
Conclusion

▶ Estimation for mean-variance optimization:

▶ Means are hardest to estimate
▶ Simple implementable alternatives: minimum-variance, risk parity,
equal weights (these are good 401k choices)

Nuno Clara 39 / 39

Investments (2024)