Portfolio Management

MGMT 648 · October 10

Kerry Back

Sat Oct 10, 2026

The Sharpe ratio

Excess return per unit of standard deviation.

Definition

Sharpe ratio=average return−risk-free ratestandard deviation of return \text{Sharpe ratio} \;=\; \frac{\text{average return} - \text{risk-free rate}} {\text{standard deviation of return}}

The numerator is the average excess return: what the investment earned beyond a Treasury bill. The denominator is the standard deviation of that same excess return, which equals the standard deviation of the return itself when the bill rate is known in advance.

Can be regarded as reward per unit risk.

Cash

Cash is invested at the money market rate rfr_f. Suppose half of each dollar invested is put in stocks earning rr and the other half is in the money market. The portfolio return and the excess return over the money market rate are

rp=12r+12rf,rp−rf=12(r−rf).r_p = \frac{1}{2}r + \frac{1}{2}r_f\,, \qquad r_p - r_f = \frac{1}{2}(r - r_f)\,.

The expected excess return and the standard deviation are each half those of rr, so the Sharpe ratio is the same as that of rr. This fact is true for any division between stocks and the money market.

Leverage

If we invest $100 and borrow another $100 at the rate rfr_f to hold $200 of stocks, then the portfolio return and excess return are:

rp=2r−rf,rp−rf=2(r−rf).r_p = 2r - r_f\,, \qquad r_p - r_f = 2(r - r_f)\,.

The expected excess return and the standard deviation are each twice those of rr, so the Sharpe ratio, measured against the borrowing rate, is again the same. This is true for any amount borrowed.

Simplification

To simplify, we will assume borrowing and saving at the T-bill rate. In reality, brokerages charge more than that for borrowing.

Some ETF data

Returns

Total returns on three sheets: weekly, monthly, and annual.

Sample

January 2005 through October 2026. GLD has the shortest history of these funds; it began trading in November 2004.

Risk-free rate

The 3-month Treasury, FRED series DGS3MO. It averaged 1.8% over the sample and pays 4.2% today.

Download: etf_returns.xlsx. Every optimization that follows uses the monthly sheet; the weekly sheet is what the estimation-error slides discuss.

One portfolio, three leverages

Each point is a combination of SPY and Treasury bills. The line starts at the bill rate and its slope is the Sharpe ratio, so raising the Sharpe ratio tilts the whole line up and gives a higher average return at every level of risk.

We’re assuming borrowing at the T-bill rate. In reality, there will be a 1% or 2% spread. That changes the slope, bending the line down, beyond the “all in” point.

Two funds: IEF and SPY

The two funds

IEF

iShares 7–10 Year Treasury Bond ETF. Holds Treasury notes with seven to ten years left to maturity and sells them as they age out of that range. No credit risk; the price moves when yields move.

SPY

SPDR S&P 500 ETF Trust. Holds the 500 stocks of the S&P 500 in index proportions. Listed in 1993, the first US exchange-traded fund.

Excess returns, not raw returns

Every mean and standard deviation from here on is estimated from excess returns: returns minus the T-bill return for that period.

IEF

Over the sample, IEF averaged 3.1% a year, and T-bills averaged 1.8%, so the premium was 1.3%. It makes more sense to extrapolate the 1.3% premium to the future than to extrapolate the 3.1% mean return, because T-bills today are at 4.2%: the expected return is 5.5%.

Treasuries and the S&P 500

Average excess return Standard deviation Sharpe ratio Expected return
IEF, 7–10 year Treasuries 1.3% 6.5% 0.20 5.5%
SPY, S&P 500 9.6% 14.8% 0.65 13.8%

The correlation between the two is −0.07-0.07.

Monthly excess returns, 261 months, annualized. Expected return is today’s 4.2% plus the average excess return, and is the vertical axis of every picture that follows.

Maximizing the Sharpe ratio

Choose the weights on the funds to make the Sharpe ratio as large as possible. The weights should sum to 1 (fully invested in the two ETFs with no leverage).

Prompt

Find the portfolio of IEF and SPY that maximizes the Sharpe ratio. Use the monthly returns in etf_returns.xlsx to estimate means, standard deviations, and the correlation. Use a risk-free rate of 0.042/12. Do you need any other information from me?

I expect it to ask whether you want to allow short sales. Say “yes.” When it finishes, ask for a plot, if it didn’t give you one automatically.

Every mix of the two

The minimum-variance portfolio

The left tip of the curve: the weights that make the portfolio’s standard deviation as small as possible. Nothing is said about its average return, and that missing condition is what separates it from the problem on the next slide.

82% IEF

18% SPY

5.8% standard deviation

At 5.8% it carries less risk than IEF alone, which is 6.5%, and earns more over bills: 2.8% against 1.3%.

The tangency portfolio

45% IEF

55% SPY

0.69 Sharpe ratio

Neither fund on its own reaches 0.69. SPY alone is 0.65 and IEF alone is 0.20.

A third fund: GLD

Gold against the other two

Average excess return Standard deviation Sharpe ratio Expected return
IEF 1.3% 6.5% 0.20 5.5%
SPY 9.6% 14.8% 0.65 13.8%
GLD 9.6% 17.3% 0.56 13.9%

Correlations: GLD with SPY 0.08, GLD with IEF 0.31, IEF with SPY −0.07-0.07.

What does GLD add?

GLD and SPY have the same average excess return, 9.6%, and GLD is the more volatile of the two: 17.3% against 14.8%.

What GLD adds

Its correlation with the 45/55 portfolio of IEF and SPY is 0.19. SPY’s correlation with that same portfolio is 0.94.

The criterion

An asset is worth holding when its excess return is high relative to its covariance with the portfolio, not relative to its own variance.

Exercise

Use AI to find the minimum variance portfolio and the tangency portfolio using the monthly returns to estimate means, variances, and covariances. Assume the risk-free rate is 0.042/12. If asked about short sales, tell AI to allow them (it won’t matter for this example).

Ask for a plot.

The new weights

17% IEF

50% SPY

33% GLD

The Sharpe ratio rises in sample from 0.69 to 0.83. The minimum-variance portfolio hardly moves — 5.8% either way, with well under 1% in GLD — so the gain is in the tangency portfolio and not in the left tip.

In sample, because these are the weights that fit this sample best. What they earn on returns they were not fitted to comes at the end of this half.

Three funds

How precisely we know the inputs

Confidence intervals for means and for standard deviations.

Annualizing means and standard deviations

A year’s return is the sum of 52 weekly returns, exactly in logs and closely enough in levels. If those 52 returns are independent of one another, the variance of their sum is the sum of their variances:

σannual2=52σweekly2soσannual=52σweekly \sigma^2_{\text{annual}} = 52\,\sigma^2_{\text{weekly}} \qquad\text{so}\qquad \sigma_{\text{annual}} = \sqrt{52}\,\sigma_{\text{weekly}}

Means add the same way, μannual=52μweekly\mu_{\text{annual}} = 52\,\mu_{\text{weekly}}, so the mean scales with the number of periods and the standard deviation with its square root.

The independence assumption

SPY’s weekly excess returns have a lag-1 autocorrelation of −0.09-0.09, so part of each week’s move is reversed the week after. Some of the weekly variance cancels inside a month, and the variance of a monthly return comes to 3.06 times the variance of a weekly return rather than 4.33.

17.6% annualized from weekly

14.8% annualized from monthly

17.1% from annual returns

The third number breaks the pattern, and two things are behind it. Monthly returns are mildly positively autocorrelated, +0.01+0.01 at lag 1, which pushes the annual variance above 12 times the monthly. But that accounts for only about a seventh of the gap. The rest is that 21 observations cannot pin down a standard deviation at all.

This is why every optimization in this session runs on monthly returns. A weekly covariance matrix scaled by 52 overstates the annual risk of whatever reverses week to week, and the optimizer prices that overstatement into its weights.

Confidence intervals, from annual data

With TT annual excess returns drawn from a distribution with standard deviation σ\sigma, the sample average r‾\bar r has standard error σ/T\sigma / \sqrt{T}, and the 95% interval for the expected excess return is

r‾±1.96σT \bar r \;\pm\; 1.96\,\frac{\sigma}{\sqrt{T}}

At σ=17%\sigma = 17\% and T=21T = 21 years the standard error is 3.7%, so the interval runs 7.3 points either side of the average. Halving that width takes four times the years; getting it to a point either side would take about 1,100.

The mean at three frequencies

Returns Observations Average excess Standard deviation Standard error 95% interval
Annual 21 10.30% 17.08% 3.73% 3.00% to 17.61%
Monthly 252 0.789% 4.274% 0.269% 0.262% to 1.317%
Weekly 1,095 0.191% 2.445% 0.074% 0.046% to 0.336%

The annual row is already annual. Multiply the monthly row by 12 and the mean is 9.5% a year, the interval 3.1% to 15.8%; multiply the weekly row by 52 and the mean is 9.9%, the interval 2.4% to 17.5%. There is very little difference. More frequent observations do not help in estimating annual means.

The standard deviation at three frequencies

Returns Observations Standard deviation 95% interval
Annual 21 17.08% 13.07% to 24.67%
Monthly 252 4.274% 3.931% to 4.684%
Weekly 1,095 2.445% 2.346% to 2.552%

Multiply the monthly row by 12\sqrt{12} and the standard deviation is 14.8% with a confidence interval of 13.6% to 16.2%. More frequent observations help a great deal in estimating standard deviations.

Annualizing the weekly row (multiply by 52\sqrt{52}) produces an estimate that is too high (17.6%), because it ignores the weekly reversals.

Correlations

There is a correlation to be estimated for each pair of assets. That makes N(N−1)/2N(N-1)/2 correlations for NN assets. For the S&P 500, we have

500 means

500 variances

124,750 correlations

Three funds took nine numbers. The S&P 500 takes 125,750, each one carrying a confidence interval.

Treat the output of a mean-variance optimizer with a lot of skepticism, particularly for a large number of assets.

Short sales

Negative weights, margin, and the rebate.

Nothing requires a weight to be positive

A negative wiw_i is a short sale.

  1. Borrow the shares from another investor, through your broker.
  2. Sell them. The proceeds stay in the account as collateral for the loan.
  3. Buy the shares back later and return them.

You owe the lender any dividends paid while you are short, so what you pay over the life of the position is the total return on the stock.

Regulation T

The Federal Reserve sets initial margin at 50% of market value for a purchase and 50% on top of the proceeds for a short sale.

The constraint

Required equity is half the long position plus half the short position, so long + short can be at most twice your equity.

A 130/30 fund

130 long and 30 short is 160 of gross exposure against 100 of equity, inside the Reg T limit of 200.

FINRA sets the maintenance requirement afterwards: 25% of a long position and usually 30% of a short one.

The short interest rebate

The lender holds your sale proceeds as cash collateral and invests them. The interest paid back to you is the rebate.

General collateral

Easy-to-borrow shares. The loan fee is a few basis points, so the rebate is close to the full market rate.

Specials

Hard-to-borrow shares. The fee can exceed the market rate and the rebate turns negative.

Retail accounts

The broker often keeps the interest on the proceeds and the rebate is zero.

Everything that follows assumes a 100% rebate: the proceeds earn the full risk-free rate.

What a long-short strategy earns

Invest $1. Go long LL dollars and short SS dollars. With a full rebate, you earn rfr_f on 1−L+S1-L + S.

Set wL=Lw_L=L (the fraction invested in LL) and set wS=−Sw_S=-S (a negative number representing your investment in SS). Then you earn rfr_f on 1−wL−wS1-w_L-w_S. The portfolio return and excess return are

rp=(1−wL−wS)rf+wLrL+wSrSrp−rf=wL(rL−rf)+wS(rS−rf) r_p \;=\; (1-w_L \;-\; w_S)r_f \;+\; w_Lr_L \;+\; w_Sr_S \qquad r_p - r_f \;=\; w_L (r_L - r_f) \;+\; w_S(r_S - r_f)

This is the same arithmetic as for two long positions, just with wSw_S being negative.

Another example: Add real estate and international stocks

Two more funds

Add these to IEF, SPY and GLD.

Ticker Fund
VNQ Vanguard Real Estate
EFA iShares MSCI EAFE, developed markets outside North America

Ask for the tangency portfolio with short sales allowed, using the monthly returns to estimate means, variances, and covariances. Then ask for it when short sales are disallowed. Use a risk-free rate of 0.042/12.

Constraining the weights

In sample, constraints hurt performance. But constraints can produce better out of sample behavior. Constraints can reduce “overfitting” to a particular random sample.

There are other methods to reduce overfitting: push the expected return and risk estimates all closer to each other or shrink all weights towards market weights.

For strategic asset allocation &mdash deciding on a baseline allocation to asset classes &mdash people never allow short sales.

One more example: Industry sectors

Nine sector funds

Ask for tangency portfolios without short sales for these funds, using the monthly returns in sector_returns.xlsx to estimate means, variances, and covariances. Use a risk-free rate of 0.042/12. Interpret the weights &mdash why high allocations to some sectors and low allocations to others.

Ticker Sector Ticker Sector
XLB Materials XLP Consumer staples
XLE Energy XLU Utilities
XLF Financials XLV Health care
XLI Industrials XLY Consumer discretionary
XLK Technology