4.1 Diversification and the Efficient Frontier
Why mixing risky assets can lower risk: two-asset portfolio maths and the role of correlation, n assets with a covariance matrix and $w^{\mathsf T}\Sigma w$, the minimum-variance portfolio by Lagrange multipliers, the efficient frontier, and the risk that diversification can never remove.
📍 The situation
Kavya, 31, works in Hyderabad and has ₹6,00,000 invested in a single stock — call it Alpha Bank Ltd — because she knows the bank and its shares "have done well". A colleague suggests she move part of it into a pharmaceutical company, Beta Pharma Ltd, "to diversify". Kavya is puzzled: Beta Pharma is also a risky stock, and its expected return looks lower. How can adding a second risky stock, with a lower return, make her portfolio better? Her cousin, who works at a fund house, adds that she should really be thinking in terms of an equity index fund, a debt fund and gold, and talks about the "minimum-variance portfolio" and the "efficient frontier".
This section turns those words into calculations. We first work out exactly what happens to return and risk when two assets are mixed, and see that the answer depends on one number — their correlation. With three or more assets the same calculation is a vector–matrix product, $w^{\mathsf T}\Sigma w$; a Lagrange multiplier then finds the least risky mix, and the best mixes for every target return trace the efficient frontier. Finally we see why diversification has a limit: some risk is shared by all stocks and no amount of spreading removes it. That limit tells Kavya roughly how many stocks it takes to diversify — and why owning ten mutual funds may not help at all.
All expected returns, volatilities and correlations on this page are illustrative assumptions chosen to make the arithmetic realistic — they are not historical data, forecasts or recommendations. In practice these inputs are estimated from past data and are uncertain; §4.3 shows how sensitive optimal portfolios are to them. Companies, funds and people are hypothetical, and nothing here is investment advice.
Two-Asset Portfolios: Return, Risk and Correlation
A portfolio is described by the fraction of money in each asset. Those fractions — the weights — are what an investor chooses, so every formula in this chapter is a function of the weights.
A portfolio of $n$ assets with total value $V$ holds amount $V_i$ in asset $i$. Its weights are $w_i=V_i/V$, so $w_1+w_2+\cdots+w_n=1$. A negative weight is a short position (an asset borrowed and sold); a portfolio with all $w_i\ge0$ is long-only. If asset $i$ returns $r_i$ over a period, the portfolio returns $$r_p=w_1r_1+w_2r_2+\cdots+w_nr_n,$$ because each rupee in asset $i$ grows by $r_i$.
We write $\mu_i=E[r_i]$ for an asset's expected return, $\sigma_i$ for its volatility (standard deviation of return, §3.1), $\sigma_{ij}=\operatorname{Cov}(r_i,r_j)$ for a covariance and $\rho_{ij}=\sigma_{ij}/(\sigma_i\sigma_j)$ for a correlation (§3.2). All are annual figures unless stated.
With weight $w$ in asset 1 and $1-w$ in asset 2,
$$\mu_p=w\mu_1+(1-w)\mu_2,$$
$$\sigma_p^2=w^2\sigma_1^2+(1-w)^2\sigma_2^2+2w(1-w)\rho\,\sigma_1\sigma_2,$$
where $\rho=\rho_{12}$. The expected return is the weighted average of the expected returns; the variance is not the weighted average of the variances — it has a cross term that depends on the correlation.
Derivation. Expectation is linear, so $E[wr_1+(1-w)r_2]=w\mu_1+(1-w)\mu_2$. For the variance use $\operatorname{Var}(aX+bY)=a^2\operatorname{Var}X+b^2\operatorname{Var}Y$ $+\,2ab\operatorname{Cov}(X,Y)$ with $a=w$, $b=1-w$ and $\operatorname{Cov}(r_1,r_2)=\rho\sigma_1\sigma_2$. ∎
For a long-only two-asset portfolio ($0\le w\le1$):
- $\sigma_p\le w\sigma_1+(1-w)\sigma_2$, with equality (for $0\lt w\lt1$) if and only if $\rho=1$. Any correlation below 1 gives a diversification benefit.
- $\sigma_p$ decreases as $\rho$ decreases, for fixed $w$ strictly between 0 and 1.
- If $\rho=-1$, the portfolio with $w=\dfrac{\sigma_2}{\sigma_1+\sigma_2}$ has $\sigma_p=0$: the risks cancel completely.
Why. Rewrite the variance as $$\sigma_p^2=\bigl(w\sigma_1+(1-w)\sigma_2\bigr)^2-2w(1-w)\sigma_1\sigma_2(1-\rho).$$ The subtracted term is $\ge0$ when $0\le w\le1$ and $\rho\le1$, and it is zero for $0\lt w\lt1$ only when $\rho=1$: this proves (1), and since it grows as $\rho$ falls, (2). If $\rho=-1$ the variance is the perfect square $(w\sigma_1-(1-w)\sigma_2)^2$, which vanishes at $w=\sigma_2/(\sigma_1+\sigma_2)$. ∎
| $\rho$ | 1 | 0.5 | 0.2 | 0 | −0.5 | −1 |
|---|---|---|---|---|---|---|
| $\sigma_p$ | 25.00% | 21.79% | 19.62% | 18.03% | 13.23% | 5.00% |
Many Assets: Vectors, Matrices and $w^{\mathsf T}\Sigma w$
With three assets the variance formula has $3$ variance terms and $3$ pairs of cross terms; with 50 stocks it has 50 variances and 1,225 covariances. Writing it out by hand is hopeless — but in matrix form it is a single line, and a spreadsheet or a few lines of Python evaluate it instantly. This is where linear algebra enters finance.
For $n$ assets, the weight vector is $w=(w_1,\dots,w_n)^{\mathsf T}$, the mean vector is $\mu=(\mu_1,\dots,\mu_n)^{\mathsf T}$, and $\mathbf 1=(1,\dots,1)^{\mathsf T}$, so the budget condition is $\mathbf 1^{\mathsf T}w=1$. The covariance matrix $\Sigma$ is the $n\times n$ matrix with entries $$\Sigma_{ij}=\sigma_{ij}=\rho_{ij}\,\sigma_i\sigma_j,$$ so its diagonal holds the variances $\sigma_i^2$ and it is symmetric ($\Sigma_{ij}=\Sigma_{ji}$).
- $\mu_p=w^{\mathsf T}\mu=\sum_i w_i\mu_i$ and $$\sigma_p^2=w^{\mathsf T}\Sigma w=\sum_{i}\sum_{j}w_iw_j\sigma_{ij}.$$
- The covariance of two portfolios $w$ and $v$ is $w^{\mathsf T}\Sigma v$.
- $\Sigma$ is positive semidefinite: $w^{\mathsf T}\Sigma w\ge0$ for every vector $w$. It is positive definite (and invertible) exactly when no combination of the assets is riskless.
Derivation. $r_p=\sum_iw_ir_i$, so $\operatorname{Var}(r_p)=\operatorname{Cov}\bigl(\sum_iw_ir_i,\sum_jw_jr_j\bigr)$, which equals $\sum_i\sum_jw_iw_j\operatorname{Cov}(r_i,r_j)$ by bilinearity of covariance, which is the double sum, i.e. $w^{\mathsf T}\Sigma w$. The same bilinearity gives (2). For (3), $w^{\mathsf T}\Sigma w$ is the variance of the return $\sum w_ir_i$, and a variance cannot be negative; it is zero only if that combination never varies. ∎ For $n=2$, (1) is Theorem 4.1.
- Build $\Sigma$: variances $\sigma_i^2$ on the diagonal, $\rho_{ij}\sigma_i\sigma_j$ off it. Check symmetry.
- Compute the vector $\Sigma w$ (each entry is the covariance of one asset with the whole portfolio).
- Compute $\sigma_p^2=w^{\mathsf T}(\Sigma w)$, the weighted sum of the entries of $\Sigma w$, then $\sigma_p=\sqrt{\sigma_p^2}$.
- In a spreadsheet: weights in a column
w, matrix inS:=SQRT(MMULT(TRANSPOSE(w),MMULT(S,w))). In Python:np.sqrt(w @ S @ w). - Check: with all correlations equal to 1 you must get $\sum w_i\sigma_i$; with one asset at weight 1 you must get its own $\sigma$.
Which asset is responsible for that 10.87%? The weights (50/30/20) do not say: a small weight in a very volatile asset can dominate. The right tool splits $\sigma_p$ into additive pieces.
The risk contribution of asset $i$ is $RC_i=\dfrac{w_i\,(\Sigma w)_i}{\sigma_p}$. The contributions add up to the portfolio volatility, $$\sum_{i=1}^nRC_i=\sigma_p,$$ and $RC_i=w_i\,\dfrac{\partial\sigma_p}{\partial w_i}$: the weight times the marginal risk of asset $i$. The percentage contribution is $RC_i/\sigma_p$.
Derivation. $\sum_iw_i(\Sigma w)_i=w^{\mathsf T}\Sigma w=\sigma_p^2$; divide by $\sigma_p$. For the derivative, $\partial(w^{\mathsf T}\Sigma w)/\partial w_i=2(\Sigma w)_i$ (because $\Sigma$ is symmetric), and $\partial\sigma_p/\partial w_i=\frac{1}{2\sigma_p}\cdot2(\Sigma w)_i$. ∎ This is the calculus behind the "risk budget" reports of fund managers and the risk parity idea of §4.3.
The Minimum-Variance Portfolio and the Efficient Frontier
Among all the mixes we could hold, two questions matter: which is the least risky, and, for each target return, which mix reaches it with the least risk? Both are optimisation problems with a constraint (the weights must add up to 1) — exactly what Lagrange multipliers are for.
The global minimum-variance portfolio (GMV) is the portfolio with the smallest variance among all $w$ with $\mathbf 1^{\mathsf T}w=1$. A portfolio is (mean–variance) efficient if no other portfolio has a higher expected return with the same or lower volatility. Plotted in the $(\sigma,\mu)$ plane, the minimum-variance portfolios for each target return form the minimum-variance frontier; its upper half, from the GMV upwards, is the efficient frontier.
Let $\Sigma$ be positive definite (no riskless combination of the assets). Then the global minimum-variance portfolio and its variance are $$w_{\text{GMV}}=\frac{\Sigma^{-1}\mathbf 1}{\mathbf 1^{\mathsf T}\Sigma^{-1}\mathbf 1},\qquad\sigma_{\text{GMV}}^2=\frac{1}{\mathbf 1^{\mathsf T}\Sigma^{-1}\mathbf 1}.$$ For two assets with covariance $\sigma_{12}$ (and $\sigma_1^2+\sigma_2^2-2\sigma_{12}\gt0$): $$w_1=\frac{\sigma_2^2-\sigma_{12}}{\sigma_1^2+\sigma_2^2-2\sigma_{12}}.$$ Short sales are allowed here; the weights can come out negative.
Derivation (Lagrange multipliers). Minimise $f(w)=\tfrac12w^{\mathsf T}\Sigma w$ subject to $g(w)=\mathbf 1^{\mathsf T}w-1=0$. At the minimum the gradients are parallel: $\nabla f=\lambda\nabla g$, i.e. $\Sigma w=\lambda\mathbf 1$, so $w=\lambda\Sigma^{-1}\mathbf 1$. The constraint gives $\lambda\,\mathbf 1^{\mathsf T}\Sigma^{-1}\mathbf 1=1$, which fixes $\lambda$ and the formula for $w$. Then $\sigma^2=w^{\mathsf T}\Sigma w=\lambda\,w^{\mathsf T}\mathbf 1=\lambda$. Because $f$ is a convex quadratic (Σ positive definite), this stationary point is the global minimum. For two assets, set $\frac{d}{dw}\sigma_p^2=2w\sigma_1^2-2(1-w)\sigma_2^2+2(1-2w)\sigma_{12}=0$ and solve for $w$. ∎ Notice the condition $\Sigma w=\lambda\mathbf 1$: in the GMV portfolio every asset has the same covariance with the portfolio — otherwise shifting money towards the asset with the smaller covariance would lower the risk.
Let $\Sigma$ be positive definite and $\mu$ not a multiple of $\mathbf 1$. Put $$A=\mathbf 1^{\mathsf T}\Sigma^{-1}\mathbf 1,\quad B=\mathbf 1^{\mathsf T}\Sigma^{-1}\mu,\quad C=\mu^{\mathsf T}\Sigma^{-1}\mu,\quad D=AC-B^2\gt0.$$ The least-variance portfolio with expected return $m$ is $$w(m)=\lambda\,\Sigma^{-1}\mathbf 1+\gamma\,\Sigma^{-1}\mu,\quad \lambda=\frac{C-Bm}{D},\ \gamma=\frac{Am-B}{D},$$ and its variance is $$\sigma^2(m)=\frac{Am^2-2Bm+C}{D}.$$ In the $(\sigma,\mu)$ plane this is one branch of a hyperbola with vertex at the GMV portfolio $(1/\sqrt A,\ B/A)$; the efficient frontier is the part with $m\ge B/A$. Every frontier portfolio is a combination of any two others (two-fund separation).
Derivation. Minimise $\tfrac12w^{\mathsf T}\Sigma w$ subject to $\mathbf 1^{\mathsf T}w=1$ and $\mu^{\mathsf T}w=m$. With two multipliers, $\Sigma w=\lambda\mathbf 1+\gamma\mu$, so $w=\lambda\Sigma^{-1}\mathbf 1+\gamma\Sigma^{-1}\mu$. Substituting into the constraints gives the linear system $\lambda A+\gamma B=1$, $\lambda B+\gamma C=m$, whose solution is the $\lambda,\gamma$ above ($D\gt0$ by the Cauchy–Schwarz inequality for the inner product defined by $\Sigma^{-1}$). Then $\sigma^2=w^{\mathsf T}\Sigma w=w^{\mathsf T}(\lambda\mathbf 1+\gamma\mu)=\lambda+\gamma m$, which simplifies to the formula. Since $w(m)$ is affine in $m$, any frontier portfolio is a combination of two frontier portfolios. Completing the square, $\sigma^2-\frac AD\left(m-\frac BA\right)^2=\frac1A$ — a hyperbola. ∎
MINVERSE) $$\Sigma^{-1}\mathbf 1\approx(5.702,\ 385.965,\ 31.579)^{\mathsf T},\quad\Sigma^{-1}\mu\approx(1.7158,\ 25.684,\ 3.1439)^{\mathsf T}.$$ (a) Find $A,B,C,D$, the GMV weights, return and volatility. (b) Find the frontier portfolio with expected return 10% and compare it with the 50/30/20 mix. (c) What happens for a 12% target?
- "Portfolio risk is the weighted average of the risks." Only when every correlation is 1. For Kavya's 60/40 mix the average is 26% but the true volatility 21.1% (Example 1). The return is the weighted average; the risk is not.
- "I own eight mutual funds, so I am diversified." Eight large-cap funds hold largely the same stocks; if their correlations are around 0.95, eight of them are barely less risky than one (Exercise IV.6). Count independent sources of risk, not products.
- "Adding a riskier asset must raise my risk." Not if it is weakly correlated with what you hold: the GMV portfolio of Example 6 adds gold and equity to debt and ends up less volatile than debt alone.
- "Diversification removes all risk." It removes the part that is specific to each company. Market-wide risk remains (Theorem 4.7) — in a broad sell-off almost everything falls together.
- "The correlations I measured in calm years will protect me in a crash." Correlations between stocks tend to rise in market stress (§3.2), exactly when diversification is most wanted. Stress-test with higher correlations (§7.4).
- "The efficient frontier is a forecast." It is computed from estimated inputs. Small changes in expected returns can move the "optimal" weights a lot (§4.3), and the frontier is only as good as its inputs.
- "Negative weights are just numbers." A negative weight means short selling. Retail mutual-fund investors cannot short a fund; the long-only frontier is the relevant one for them.
- "More stocks is always better." After about 20–30 well-spread stocks the extra reduction in risk is small (Example 7), while monitoring and transaction costs keep rising.
Systematic and Unsystematic Risk: How Many Stocks Are Enough?
Why does diversification stop working at some point? Because part of every stock's movement is caused by things that hit the whole market — interest rates, the economy, global risk appetite — and part by things specific to the company — a product approval, a fraud, a lost contract. Averaging over many companies cancels the second part but not the first.
In the single-index model each stock's return is written as $$r_i=\alpha_i+\beta_i\,r_m+\varepsilon_i,$$ where $r_m$ is the market (e.g. a broad index) return, $\beta_i$ the stock's beta (the slope of the market-model regression of §3.2), and the residuals $\varepsilon_i$ have mean 0, are uncorrelated with $r_m$ and (as a simplifying assumption) with each other, with variance $s_i^2$. Then $$\sigma_i^2=\underbrace{\beta_i^2\sigma_m^2}_{\text{systematic}}+\underbrace{s_i^2}_{\text{unsystematic}}.$$ Systematic (market, non-diversifiable) risk is the part explained by the market; unsystematic (specific, idiosyncratic, diversifiable) risk is the rest. The ratio $\beta_i^2\sigma_m^2/\sigma_i^2$ is the $R^2$ of the regression.
- An equal-weighted portfolio of $N$ assets has $$\sigma_p^2=\frac{1}{N}\,\overline{\sigma^2}+\Bigl(1-\frac1N\Bigr)\overline{c},$$ where $\overline{\sigma^2}$ is the average variance and $\overline c$ the average covariance over the $N(N-1)$ ordered pairs $i\ne j$. As $N\to\infty$, $\sigma_p^2\to\overline c$. If every stock has volatility $\sigma$ and every pair correlation $\rho\ge0$, the floor is $\sigma\sqrt\rho$.
- In the single-index model a portfolio with weights $w_i$ has $$\sigma_p^2=\beta_p^2\sigma_m^2+\sum_iw_i^2s_i^2,\qquad\beta_p=\sum_iw_i\beta_i,$$ and with equal weights the specific part is $\overline{s^2}/N\to0$.
Derivation. (1) With $w_i=1/N$, $w^{\mathsf T}\Sigma w=\frac1{N^2}\bigl(\sum_i\sigma_i^2+\sum_{i\ne j}\sigma_{ij}\bigr)$, which is $\frac1{N^2}\bigl(N\overline{\sigma^2}+N(N-1)\overline c\bigr)$. (2) $r_p=\sum w_i\alpha_i+\beta_pr_m+\sum w_i\varepsilon_i$; the three random parts are uncorrelated, so their variances add, and $\operatorname{Var}\sum w_i\varepsilon_i=\sum w_i^2s_i^2$. ∎ The variances shrink like $1/N$; the covariances, which are $N(N-1)$ of the $N^2$ terms, survive. Diversification is the law of large numbers working on the company-specific parts.
| $N$ | 1 | 2 | 5 | 10 | 20 | 30 | 50 | ∞ |
|---|---|---|---|---|---|---|---|---|
| $\sigma_p$ | 30.0% | 23.7% | 19.0% | 17.1% | 16.1% | 15.7% | 15.4% | 15.0% |
Using the illustrative inputs of this section, compare five ways Kavya could hold her ₹6,00,000. As a rough "bad year" marker we use $\mu-2\sigma$: under a normal model a return this low has about a 2.3% chance in any year, and real returns have fatter tails (§3.1), so treat it as a minimum stress, not a worst case.
| holding | $\mu$ | $\sigma$ | expected gain (₹) | "bad year" $\mu-2\sigma$ | bad-year result (₹) |
|---|---|---|---|---|---|
| 100% Alpha Bank | 14.0% | 30.0% | 84,000 | −46.0% | −2,76,000 |
| 60/40 Alpha/Beta (Ex. 1) | 12.4% | 21.1% | 74,400 | −29.8% | −1,78,911 |
| equity index fund | 12.0% | 20.0% | 72,000 | −28.0% | −1,68,000 |
| 50/30/20 E/D/G (Ex. 3) | 9.9% | 10.9% | 59,400 | −11.8% | −71,036 |
| GMV (Ex. 6) | 7.2% | 4.9% | 43,299 | −2.5% | −15,030 |
- Single stock → two stocks. The first step removes a third of the volatility for 1.6 points of expected return, because the two stocks are only weakly correlated (Theorem 4.2).
- Two stocks → index fund. Fifty stocks remove almost all the remaining specific risk; what is left (20%) is market risk, which no equity-only portfolio escapes (Theorem 4.7).
- Equity → equity, debt and gold. To go below the equity floor Kavya must add assets that are driven by different things. The 50/30/20 mix has about half the index fund's volatility; its expected return is lower too.
- The frontier view. The 50/30/20 mix is almost on the efficient frontier (Example 6: 10.84% volatility for 10% expected return). The GMV is the least risky point but also the lowest return; anything below it on the curve is inefficient.
The mathematics, not a recommendation. Every efficient row trades expected return for lower risk; none is "best" without knowing Kavya's goals, time horizon and ability to sit through a bad year (§4.2 and §8.1). What the mathematics does say firmly is that the first row is not efficient: concentrated holdings carry risk that is not compensated by extra expected return.
Exercises
All inputs are illustrative annual figures. Unless stated, the three-asset data are those of Example 3: $\mu=(12\%,7\%,9\%)$, $\sigma=(20\%,5\%,15\%)$, $\rho_{ED}=0.2$, $\rho_{EG}=0$, $\rho_{DG}=0.1$. Round percentages to two decimals.
- 1 Find the expected return and volatility of a portfolio with 40% in A.
- 2 Repeat part 1 with $\rho=0$ and with $\rho=-0.5$.
- 3 Find $\operatorname{Cov}(r_A,r_B)$.
- 4 What weight in A gives an expected return of 12%? What is the volatility of that portfolio?
- 5 Without computing, explain why the volatility in part 1 must be below $0.4\times25\%+0.6\times10\%$.
- 6 ₹5,00,000 is invested in the part 1 portfolio. Find the expected rupee gain over a year and the range "expected ± one standard deviation" in rupees.
- 7 If instead $\rho=-1$, find the weights that make the portfolio riskless and its return.
- 1 Find the minimum-variance weights, volatility and expected return for $\rho=0.3$.
- 2 Show that the minimum-variance portfolio is 100% B exactly when $\rho=\sigma_B/\sigma_A$. What is that correlation here?
- 3 For $\rho=0.6$, find the minimum-variance weights. What does the sign of $w_A$ mean, and is the result usable by a long-only investor?
- 4 Two assets have the same volatility $\sigma$ and correlation $\rho\gt-1$. Show that the minimum-variance portfolio is 50/50 and that $\sigma_{\min}=\sigma\sqrt{(1+\rho)/2}$. Evaluate for $\sigma=20\%$, $\rho=0.3$.
- 5 For uncorrelated assets, use a Lagrange multiplier to minimise $w_1^2\sigma_1^2+w_2^2\sigma_2^2$ subject to $w_1+w_2=1$, and show $w_1=\dfrac{1/\sigma_1^2}{1/\sigma_1^2+1/\sigma_2^2}$ ("inverse-variance weights"). Evaluate for A and B.
- 6 For the three-asset GMV weights $(1.347\%,\ 91.192\%,\ 7.461\%)$ of Example 6, compute $\Sigma w$ and explain what you see.
- 1 Find the expected return and volatility of the mix 70/20/10 (equity/debt/gold).
- 2 Repeat for 30/50/20.
- 3 Find the covariance and the correlation between the portfolios of parts 1 and 2.
- 4 Find the percentage risk contributions of the three assets in both portfolios.
- 5 Could $\begin{pmatrix}0.04&0.05\\0.05&0.04\end{pmatrix}$ be a covariance matrix? Give two reasons.
- 6 Explain in one or two sentences why every covariance matrix satisfies $w^{\mathsf T}\Sigma w\ge0$, and what $w^{\mathsf T}\Sigma w=0$ for some $w\ne0$ would mean financially.
- 1 Stocks each have volatility 35% and pairwise correlation 0.2. Find the volatility of an equal-weighted portfolio of 1, 10 and 30 stocks, and the floor.
- 2 In part 1, what fraction of the diversifiable variance has been removed with 10 stocks? What fraction of the total variance of one stock remains?
- 3 A stock has beta 0.8 and volatility 28%; the market's volatility is 15%. Split its variance into systematic and specific parts and find the regression $R^2$.
- 4 A portfolio holds 25 stocks equally, average beta 1.1, each with specific volatility 25% and uncorrelated residuals; market volatility 15%. Find its volatility.
- 5 Thirty bank stocks each have volatility 30% and pairwise correlation 0.7. Find the volatility of an equal-weighted portfolio and the floor. Why is this not a well-diversified portfolio?
- 6 Eight large-cap funds each have volatility 18% and pairwise correlation 0.95. Find the volatility of an equal split among them and comment.
E1. NISM-style Holding a larger number of stocks from different sectors mainly reduces (A) systematic risk (B) unsystematic risk (C) interest-rate risk of the whole market (D) inflation risk.
E2. NISM-style Two funds have volatilities 20% and 30% and correlation 0.5. A 50:50 portfolio of them has volatility closest to (A) 25.0% (B) 21.8% (C) 18.0% (D) 12.5%.
E3. NISM-style Which statement about two assets with correlation $-1$ is correct? (A) Combining them always increases risk (B) Some combination of them has zero risk (C) Their expected returns must be equal (D) Diversification has no effect.
E4. Case study An investor holds ₹10 lakh as 60% equity, 30% debt and 10% gold, with the three-asset inputs of Example 3. (a) Find the expected return and volatility. (b) Find the percentage risk contributions. (c) Using Example 6, find a frontier portfolio with at least the same expected return and lower risk, or explain why none is close.
E5. Case study A portfolio manager runs a concentrated 15-stock portfolio. Assume each stock has volatility 32% and pairwise correlation 0.3. (a) Find the portfolio volatility. (b) Find the floor. (c) An index fund has volatility 18%. How much of the concentrated portfolio's volatility is the price of concentration rather than market exposure?
E6. Interview Two assets each have volatility 20%. For which correlations can some portfolio of them (long or short) have volatility below 10%? What portfolio achieves the minimum?
Suggested plan — 2 lectures × 60 min.
- Two assets and correlation. Start from the 📍 situation: ask the class whether adding a lower-return stock can make a portfolio better (most say no). Derive Theorem 4.1 from $\operatorname{Var}(aX+bY)$ and do Example 1. Project Interactive 1 and press Play: let students watch the curve bend from a straight line ($\rho=1$) to a V that touches zero ($\rho=-1$). Do Example 2 and discuss why $\rho=-1$ would be an arbitrage. Then matrices: Definition 4.2, the strategy box and Example 3 on the board; Example 4 (risk contributions) as a surprise — "50% equity is 87% of the risk".
- Optimisation and the limit of diversification. Derive Theorem 4.5 with a Lagrange multiplier and interpret $\Sigma w=\lambda\mathbf 1$. Example 5, then Example 6 with Interactive 2: sweep the target and note where weights turn negative; move the equity–gold correlation. Finish with Theorem 4.7, Example 7 (and the chapter-intro teaser widget) and the 📈 Market case.
Class activity. Give each group a different correlation matrix for equity, debt and gold (one "calm years", one "crisis" with all correlations raised by 0.3). Each group computes the 50/30/20 volatility and the risk contributions in a spreadsheet, then compares results: how much does the crisis matrix raise the risk?
Misconceptions to watch for. Portfolio σ = weighted average of σ's; more funds = more diversification; diversification removes all risk; negative weights are harmless; the frontier is a forecast; correlations are stable.
Suggested homework. I(1, 2, 4, 6), II(1, 2, 4, 5), III(1, 3, 4, 5), IV(1, 3, 5, 6); Practice corner E2, E4, E6.
Project. Estimate a covariance matrix for three asset classes from real monthly data and draw your own efficient frontier, with and without short sales.
Why it matters. Every frontier you see in a brochure was computed from some data over some period. Building one yourself shows how much the answer depends on those choices — the main lesson of §4.3.
What to do.
- Download at least five years of month-end values for three series — for example a broad equity index (from the index provider's website), the NAV of a debt index fund and of a gold ETF or gold fund (from the fund house's website or AMFI's NAV history). Put dates in column A and the three series in B–D.
- Monthly returns in F–H:
=B3/B2-1, copied across and down. Means=AVERAGE(F3:F62); annualise with*12. Volatilities=STDEV.S(F3:F62)*SQRT(12). - Covariance matrix: in a 3×3 block,
=COVARIANCE.S(F$3:F$62,$F3:$F62)-style formulas (or=COVARIANCE.S(x,y)*12cell by cell) — check that it is symmetric and that its diagonal equals the squared volatilities. - Portfolio volatility for weights in a column
w:=SQRT(MMULT(TRANSPOSE(w),MMULT(S,w)))(enter as an array formula in older spreadsheets). Check it against Example 3's method by hand for one mix. - GMV:
=MMULT(MINVERSE(S),{1;1;1}), then divide by the sum of its entries (Theorem 4.5). Frontier: compute $A,B,C,D$ and tabulate $\sigma(m)$ for $m$ from the GMV return to the highest mean; plot $\sigma$ against $m$. - Long-only frontier: use Solver — minimise the volatility cell by changing
w, subject toSUM(w)=1,w>=0andMMULT(TRANSPOSE(w),mu)=m; repeat for several $m$ and plot both frontiers on one chart.
Python version: R = prices.pct_change().dropna(); mu = 12*R.mean(); S = 12*R.cov(); Si = np.linalg.inv(S); w_gmv = Si@np.ones(3)/(np.ones(3)@Si@np.ones(3)); for the long-only frontier use scipy.optimize.minimize with bounds (0,1).
Hand in: your $\mu$, volatilities and correlation matrix; the GMV weights; one chart with both frontiers and your own current (or a sample) allocation marked on it; two sentences on how the answer changes if you use only the first half of the data.
Going further. Recompute the matrix on a rolling three-year window and plot the equity–gold correlation through time. Does it change sign? What happens to the GMV weights when it does?
🧪 Self-check quiz
Q1. The volatility of a long-only two-asset portfolio equals the weighted average of the two volatilities exactly when the correlation is
Q2. In the matrix formula $\sigma_p^2=w^{\mathsf T}\Sigma w$, the off-diagonal entries of $\Sigma$ are
Q3. Two assets have equal volatility and correlation 0.4. The minimum-variance portfolio is
Q4. Stocks each have volatility 30% and pairwise correlation 0.25. As the number of stocks in an equal-weighted portfolio grows without limit, the portfolio volatility tends to
Q5. At the global minimum-variance portfolio found with a Lagrange multiplier, $\Sigma w=\lambda\mathbf 1$. This means