Chapter 3 · The Statistics of Returns

3.2 Correlation, Beta and Risk-Adjusted Performance

Two return series at once: covariance and correlation, the regression line of a stock on the index (beta, alpha, $R^2$), market vs specific risk, and the Sharpe, Sortino, Treynor and information ratios — used correctly, with their noise, and with correlations that rise in a crisis.

📍 The situation

Meera, 31, a software tester in Hyderabad, has been investing for three years. Her brother tells her that "Beta Pharma Ltd" (a hypothetical share) "has a beta of only 0.6, so it is a low-risk stock". She owns five bank shares and thought she was diversified, yet on one bad day in our simulated market all five fell between 3% and 5% together. And she is choosing between two flexi-cap funds: a ranking website shows Fund P with a five-year return of 18% and a Sharpe ratio of 0.52, and Fund Q with 15% and 0.61. Which number should she believe, and what do they mean?

All three questions involve two return series at once — a stock and the index, two stocks, a fund and its benchmark. This section gives the tools for that. Covariance and correlation measure whether two returns move together. The regression line of a stock's returns on the index's returns has a slope called beta and an intercept called alpha, and it splits the stock's risk into a market part and a specific part ($R^2$ says how much is which). Finally, risk-adjusted ratios — Sharpe, Sortino, Treynor and the information ratio — put return and risk on one scale so funds can be compared fairly, and the same statistics tell us how much of a difference between two funds is simply noise. The mathematics is the least-squares line from school statistics, found by setting two partial derivatives to zero, and the Cauchy–Schwarz inequality.

⚠️ About the numbers on this page

All returns, betas and ratios on this page come from small hypothetical data sets or are simulated in your browser with a seeded random-number generator; the funds and companies are hypothetical. They are not the statistics of any real share, fund or index, and nothing here is a recommendation. The risk-free rate is an illustrative figure (6–6.5% a year) standing in for a short-term government-security yield; for real comparisons use the current T-bill or overnight rate, and the convention the fund factsheet states.

Covariance and Correlation

Look at a scatter plot with the index's monthly return on the horizontal axis and a bank share's on the vertical. If the points rise from left to right, the two tend to move together. Covariance turns this into a number by asking whether the deviations from the means have the same sign: when both are above average (or both below), the product of the deviations is positive.

📘 Definition 3.6 — Covariance and correlation

For paired returns $(x_1,y_1),\dots,(x_n,y_n)$ with means $\bar x,\bar y$ and volatilities $s_x,s_y$ (Definition 3.1):

  • the covariance is $s_{xy}=\dfrac1{n-1}\displaystyle\sum_{t=1}^n(x_t-\bar x)(y_t-\bar y)$;
  • the correlation is $\rho_{xy}=\dfrac{s_{xy}}{s_x\,s_y}$.

Covariance has units ($\%^2$) and depends on the scale of the data; correlation is a pure number. Spreadsheets: COVARIANCE.S and CORREL. Always compute them from returns, never from price levels.

📐 Theorem 3.6 (Rule) — Properties of correlation, and the variance of a sum
  1. $-1\le\rho_{xy}\le1$, and $\rho_{xy}=\pm1$ exactly when the points lie on a straight line $y=a+bx$ (with $b$ of the same sign as $\rho$).
  2. Correlation is unchanged by changing units or adding constants: replacing $x$ by $a+bx$ with $b\gt0$ (percent to decimal, adding a risk-free rate) leaves $\rho$ the same, while the covariance is multiplied by $b$.
  3. For a portfolio with weights $w_1,w_2$: $$\operatorname{Var}(w_1x+w_2y)=w_1^2s_x^2+w_2^2s_y^2+2w_1w_2\,\rho_{xy}\,s_xs_y .$$ The lower the correlation, the lower the portfolio's risk — the start of diversification (§4.1).

Why (part 1). Write $u_t=x_t-\bar x$, $v_t=y_t-\bar y$. For every real $\lambda$, $\sum(v_t-\lambda u_t)^2\ge0$, i.e. $\lambda^2\sum u_t^2-2\lambda\sum u_tv_t+\sum v_t^2\ge0$. A quadratic in $\lambda$ that is never negative has discriminant $\le0$: $\left(\sum u_tv_t\right)^2\le\sum u_t^2\sum v_t^2$ (the Cauchy–Schwarz inequality). Dividing by $(n-1)^2s_x^2s_y^2$ gives $\rho^2\le1$. Equality means some $\lambda$ makes every $v_t-\lambda u_t=0$: the points lie on a line. Part 2 follows because the deviations of $a+bx$ are $b$ times those of $x$, and part 3 by expanding $\sum\big(w_1u_t+w_2v_t\big)^2$. ∎

✏️ Example 1 — A bank share and the index over six months
Monthly returns (%) over six months of a broad market index $x$ and of the hypothetical share Alpha Bank Ltd $y$ are: $x=2,\,-1,\,3,\,-4,\,1,\,5$ and $y=5,\,-4,\,0,\,-5,\,4,\,6$. Find the covariance and the correlation.
Means $\bar x=6/6=1\%$, $\bar y=6/6=1\%$. Deviations: $u=1,\,-2,\,2,\,-5,\,0,\,4$; $\ v=4,\,-5,\,-1,\,-6,\,3,\,5$.
Sums $\sum u_tv_t=4+10-2+30+0+20=62$, $\ \sum u_t^2=50$, $\ \sum v_t^2=112$.
Covariance $s_{xy}=62/5=12.4\ (\%^2)$; $s_x=\sqrt{50/5}\approx3.16\%$, $s_y=\sqrt{112/5}\approx4.73\%$.
Correlation $\rho=\dfrac{62}{\sqrt{50\times112}}=\dfrac{62}{74.83}\approx0.83$: a strong positive relation, as expected for a bank share and the index. (Six months is far too few for a reliable estimate — Interactive 1 shows how much such estimates vary.)
✏️ Example 2 — How much does correlation matter for a 50:50 portfolio?
Two assets have annual volatilities 20% and 15%. Find the volatility of a portfolio with half in each when the correlation is (a) 0.9, (b) 0.3, (c) −0.2.
Formula $\sigma_p^2=0.25\times20^2+0.25\times15^2+2\times0.25\,\rho\times20\times15=156.25+150\rho\ (\%^2)$.
Values (a) $\rho=0.9$: $291.25$, $\sigma_p\approx17.07\%$. (b) $\rho=0.3$: $201.25$, $\sigma_p\approx14.19\%$. (c) $\rho=-0.2$: $126.25$, $\sigma_p\approx11.24\%$.
Interpret With correlation 0.9 the portfolio is barely less risky than the average of the two (17.5%); with negative correlation it is less risky than either asset alone. Five bank shares are highly correlated with one another, which is why Meera's "diversified" holding fell together. Chapter 4 builds on exactly this formula.

Beta, Alpha and the Market-Model Regression

Correlation says whether a share moves with the index; it does not say by how much. For that we fit a straight line through the scatter plot — the line that best predicts the share's return from the index's return.

📘 Definition 3.7 — The market model: beta, alpha and specific risk

The market model (single-index model) writes the return of asset $i$ in period $t$ as $$r_{i,t}-r_f=\alpha_i+\beta_i\,(r_{m,t}-r_f)+\varepsilon_{i,t},$$ where $r_m$ is the market index return, $r_f$ the risk-free rate for the period, and the residual $\varepsilon_{i,t}$ has mean 0 and is uncorrelated with $r_m$. Here

  • beta $\beta_i$ is the sensitivity to the market: if the index's excess return is 1 percentage point higher, the asset's is expected to be $\beta_i$ points higher;
  • alpha $\alpha_i$ is the average excess return not explained by market exposure (Jensen's alpha when the data are annualised);
  • $\varepsilon_{i,t}$ is the specific (idiosyncratic, unsystematic) part of the return; its standard deviation $\sigma_{\varepsilon}$ is the specific risk.

For monthly data $r_f$ is small, and the model is often fitted to raw returns, $r_i=a+\beta r_m+\varepsilon$; the slope is then almost the same.

📐 Theorem 3.7 (Formula) — Least-squares beta and alpha, $R^2$ and the split of risk

For data $(x_t,y_t)$ (index and asset returns, or excess returns), the line $y=\hat a+\hat\beta x$ that minimises the sum of squared residuals $\sum(y_t-a-\beta x_t)^2$ is $$\hat\beta=\frac{s_{xy}}{s_x^2}=\rho_{xy}\frac{s_y}{s_x},\qquad \hat a=\bar y-\hat\beta\,\bar x .$$ Its coefficient of determination is $R^2=\rho_{xy}^2$, and the asset's variance splits exactly into a market part and a specific part: $$s_y^2=\underbrace{\hat\beta^2s_x^2}_{\text{systematic}}+\underbrace{s_e^2}_{\text{specific}},\qquad R^2=\frac{\hat\beta^2s_x^2}{s_y^2},$$ where $s_e^2=\frac1{n-1}\sum e_t^2$ is the variance of the residuals $e_t=y_t-\hat a-\hat\beta x_t$.

Derivation (the calculus link). Let $S(a,\beta)=\sum(y_t-a-\beta x_t)^2$. At the minimum both partial derivatives vanish: $\partial S/\partial a=-2\sum(y_t-a-\beta x_t)=0$ gives $\bar y=a+\beta\bar x$, so the line passes through the point of means. Substituting $a=\bar y-\beta\bar x$, $S=\sum(v_t-\beta u_t)^2$ with $u,v$ the deviations, and $dS/d\beta=-2\sum u_t(v_t-\beta u_t)=0$ gives $\beta=\sum u_tv_t/\sum u_t^2=s_{xy}/s_x^2$. (S is a parabola opening upwards in $\beta$, so this is a minimum.) Then $\sum e_t^2=\sum(v_t-\hat\beta u_t)^2$ $=\sum v_t^2-\hat\beta\sum u_tv_t$ $=\sum v_t^2-\hat\beta^2\sum u_t^2$, which is the variance split after dividing by $n-1$; and $\hat\beta^2s_x^2/s_y^2=s_{xy}^2/(s_x^2s_y^2)=\rho^2$. ∎

✏️ Example 3 — Beta, alpha and $R^2$ of Alpha Bank
Using the six months of Example 1, find the regression line of Alpha Bank's return on the index return, its $R^2$, and the systematic and specific parts of its variance. If the index falls 3% next month, what does the line predict?
Slope $\hat\beta=\sum u_tv_t/\sum u_t^2=62/50=1.24$.
Intercept $\hat a=\bar y-\hat\beta\bar x=1-1.24\times1=-0.24\%$ a month. The line is $\hat y=-0.24+1.24x$.
Split $R^2=0.8285^2\approx0.69$. Sums of squares: total $112$ $=$ explained $1.24^2\times50=76.88$ $+$ residual $35.12$. So about 69% of the share's variance is market risk and 31% is specific; the residual volatility is $\sqrt{35.12/5}\approx2.65\%$ a month.
Predict $\hat y=-0.24+1.24\times(-3)\approx-3.96\%$ — a prediction of the average response; the actual month could easily be a few percent either side (residual volatility 2.65%). With six data points the estimated beta itself is very uncertain.
✏️ Example 4 — What a beta does and does not say
A hypothetical share has annual volatility 30% and beta 1.2 against an index whose volatility is 15%. (a) Split its variance into systematic and specific parts and find $R^2$ and the correlation. (b) Meera's brother calls Beta Pharma (beta 0.6) "low risk". Its volatility is 35%. What is its specific risk, and is he right?
(a) Systematic: $(1.2\times15\%)^2=18\%^2=324\ (\%^2)$. Total: $30^2=900$. Specific: $900-324=576$, i.e. $\sigma_\varepsilon=24\%$. $R^2=324/900=0.36$ and $\rho=\sqrt{0.36}=0.6$.
(b) Systematic $(0.6\times15)^2=81$; total $35^2=1225$; specific $1144$, $\sigma_\varepsilon\approx33.8\%$; $R^2\approx0.066$.
Verdict Wrong. A low beta means low market risk, not low risk: Beta Pharma's price is driven almost entirely by company-specific news (drug approvals, results), and on its own it is riskier than the bank share. Beta measures what a share adds to a diversified portfolio, where specific risks largely cancel; for a concentrated holding, total volatility matters.
Interactive 1 — Scatter, Regression Line and Beta: Signal, Noise and Crises

Each dot is one simulated month: the index's return across, a share's return up. The share is built as $r=\beta\,r_m+\varepsilon$ with the true beta and specific volatility you set. The red line is the least-squares fit; the dashed line is the true relation. Press Play to draw a new sample every second and watch the estimated beta jump around — the tally collects the estimates. Switch to crisis: the index becomes far more volatile while specific risk stays the same, and the correlation rises. (Simulated data.)

1.2020%
estimated beta (true)—
estimated alpha / month—
correlation (true)—
R² — market share of risk—
beta estimates so far—

Risk-Adjusted Performance: Sharpe, Sortino, Treynor and Information Ratios

Fund P returned 18% and Fund Q 15%. But P might simply have taken more risk — in a rising market, a riskier fund usually returns more, and in a falling one it loses more. A fair comparison asks how much extra return each fund delivered per unit of the risk it took. Different ratios use different notions of "risk", and each answers a different question.

📘 Definition 3.8 — Sharpe and Sortino ratios

Let $\bar R$ be a fund's average return over a period, $r_f$ the risk-free rate for the same period, and $\sigma$ the volatility of its returns.

  • The Sharpe ratio is $\ S=\dfrac{\bar R-r_f}{\sigma}$: excess return per unit of total volatility.
  • With a target return $\tau$ (often $r_f$ or $0$), the downside deviation is $\ \text{DD}_\tau=\sqrt{\frac1n\sum_t\min(R_t-\tau,0)^2}$, and the Sortino ratio is $\ \dfrac{\bar R-\tau}{\text{DD}_\tau}$: excess return per unit of downside volatility only.

Ratios are usually quoted annualised. Conventions (divisor, target, data frequency) differ between providers, so compare ratios only when they are computed the same way.

📘 Definition 3.9 — Treynor ratio, Jensen's alpha, tracking error and information ratio
  • Treynor ratio $\ T=\dfrac{\bar R-r_f}{\beta}$: excess return per unit of market risk (beta from Theorem 3.7).
  • Jensen's alpha $\ \alpha=\bar R-\big[r_f+\beta(\bar R_m-r_f)\big]$: the return above what the fund's beta alone would have earned (the intercept of the market model, annualised).
  • For a fund with benchmark $B$, the active return in period $t$ is $R_t-B_t$; the tracking error TE is the volatility of the active returns, and the information ratio is $\ \text{IR}=\dfrac{\overline{R-B}}{\text{TE}}$: average active return per unit of active risk.
📐 Theorem 3.8 (Rule) — Annualising the Sharpe ratio, and why leverage cannot change it
  1. If period returns are independent and there are $m$ periods a year, $$S_{\text{ann}}\approx S_{\text{period}}\sqrt m ,$$ because the excess return scales with $m$ and the volatility with $\sqrt m$ (Theorem 3.1). The same applies to the Sortino and information ratios.
  2. Putting a fraction $w\gt0$ of money in the fund and $1-w$ in the risk-free asset (borrowing if $w\gt1$) gives a portfolio with excess return $w(\bar R-r_f)$ and volatility $w\sigma$ — the same Sharpe ratio. Sharpe measures the quality of the risk–return trade-off, not its size; the Treynor ratio is likewise unchanged, since beta becomes $w\beta$.

Why (part 2). The portfolio return is $wR+(1-w)r_f$, so its excess return is $w(R-r_f)$; the risk-free part has no variance, so the volatility is $|w|\sigma=w\sigma$. The ratio is $\frac{w(\bar R-r_f)}{w\sigma}=S$. ∎  This is why a high return alone proves nothing: any fund's return can be scaled up with leverage. It also gives the geometry of Interactive 2 — in the risk–return plane, a line from the risk-free point through a fund has slope equal to its Sharpe ratio (the capital allocation line of §4.2).

✏️ Example 5 — Which portfolio did better?
Over the same five years (hypothetical figures, annualised) Portfolio X returned 15% with volatility 18% and beta 1.1; Portfolio Y returned 12% with volatility 11% and beta 0.7; the market index returned 13% with volatility 15%. The illustrative risk-free rate is 6.5%. Compute each one's Sharpe ratio, Treynor ratio and Jensen's alpha.
Sharpe X: $8.5/18\approx0.47$; Y: $5.5/11=0.50$; index: $6.5/15\approx0.43$.
Treynor X: $8.5/1.1\approx7.73\%$; Y: $5.5/0.7\approx7.86\%$; index: $6.5\%$ (beta 1).
Alpha Market excess return $13-6.5=6.5\%$. X: $15-(6.5+1.1\times6.5)=15-13.65=1.35\%$. Y: $12-(6.5+0.7\times6.5)=12-11.05=0.95\%$.
Read it Both beat the index on every measure. Y has the better trade-off per unit of risk (Sharpe, Treynor); X has the larger alpha, i.e. more return beyond its market exposure. Which matters depends on the use (see the 🧭 box below). And X's higher raw return is mostly its higher beta in a rising market.
✏️ Example 6 — Sortino from ten months of returns
A fund's monthly returns (%) are $4,\,-2,\,3,\,-5,\,1,\,6,\,-1,\,2,\,0,\,4$, and the risk-free rate is 0.5% a month. Using the target $\tau=r_f$, compute the monthly Sharpe and Sortino ratios and annualise both.
Mean, σ $\bar R=12/10=1.2\%$; sample volatility $s\approx3.29\%$. Sharpe $=(1.2-0.5)/3.29\approx0.213$ a month, $\times\sqrt{12}\approx0.74$ a year.
Downside Returns below 0.5%: $-2,-5,-1,0$, with shortfalls $-2.5,-5.5,-1.5,-0.5$. $\text{DD}=\sqrt{(6.25+30.25+2.25+0.25)/10}=\sqrt{3.9}\approx1.97\%$ (all ten months count in the divisor).
Sortino $(1.2-0.5)/1.97\approx0.354$ a month, $\times\sqrt{12}\approx1.23$ a year.
Interpret The Sortino ratio is higher because the upside months (+4, +6) no longer count as "risk". It rewards funds whose volatility is mostly on the upside — but with only four bad months the downside deviation is itself very imprecise.
✏️ Example 7 — Tracking error and information ratio of an active fund
A large-cap fund's yearly returns minus its benchmark's were $+3,\,-1,\,+2,\,+4,\,-2$ percentage points over five years. Find its average active return, tracking error and information ratio. How does it compare with an index fund whose active returns are about $-0.3\%$ every year with a tracking error of $0.1\%$?
Mean $(3-1+2+4-2)/5=1.2\%$ a year.
TE Deviations $1.8,-2.2,0.8,2.8,-3.2$; squares add to $26.8$; $\text{TE}=\sqrt{26.8/4}\approx2.59\%$.
IR $1.2/2.59\approx0.46$.
Compare The index fund's IR is $-0.3/0.1=-3$: it reliably trails its benchmark by its costs, which is exactly what it is meant to do — for an index fund the tracking difference (here −0.3%) and the tracking error are the quality measures, not the IR (§2.2). For the active fund, an IR of 0.46 over five years is encouraging but, as Theorem 3.9 shows, not statistically convincing.
🧭 Which ratio answers which question?
  1. "This fund will be most of my money." Total risk matters → Sharpe (or Sortino if you care mainly about losses and the return distribution is asymmetric).
  2. "This fund will be one small part of a well-diversified portfolio." Its specific risk will mostly cancel; only its market risk remains → Treynor and Jensen's alpha.
  3. "Is the manager adding value over the benchmark I could buy cheaply in an index fund?" → information ratio (and tracking error), against the right benchmark, after costs.
  4. Always: same period, same data frequency, same risk-free rate, same benchmark, and a long enough history — then ask whether the difference is bigger than the noise (Theorem 3.9).

Comparing Funds Correctly: Noise, Causation and Crises

Ratios computed from a few years of data are estimates, and they carry the same sampling error as the mean return (Theorem 3.2) — because the mean return sits in their numerator.

📐 Theorem 3.9 (Rule) — How precise is an estimated Sharpe ratio?

If a fund's returns are independent and roughly normal, an annualised Sharpe ratio $\hat S$ estimated from $T$ years of data has approximate standard error $$\operatorname{SE}(\hat S)\approx\sqrt{\frac{1+\hat S^2/2}{T}} .$$ For typical values ($\hat S$ between 0 and 1) this is close to $1/\sqrt T$: about $0.45$ for 5 years and $0.32$ for 10 years. Differences between two funds' Sharpe ratios much smaller than this cannot be distinguished from luck.

Where it comes from. The leading term is Theorem 3.2: the numerator is a mean return with standard error $\sigma/\sqrt T$ (annualised), and dividing by $\sigma$ gives $1/\sqrt T$. The extra $\hat S^2/2$ comes from the uncertainty in the estimated volatility in the denominator (a large-sample result for normal data, quoted without proof). Fat tails and serial correlation make the true error larger. ∎

📐 Theorem 3.10 (Formula) — Correlations in a one-factor market rise with market volatility

Suppose two assets follow the market model with betas $\beta_i,\beta_j$ and specific risks $\sigma_{\varepsilon_i},\sigma_{\varepsilon_j}$, with residuals uncorrelated with each other and with the market. Then $\sigma_i^2=\beta_i^2\sigma_m^2+\sigma_{\varepsilon_i}^2$ and $$\rho_{ij}=\frac{\beta_i\beta_j\sigma_m^2}{\sigma_i\sigma_j},\qquad \rho_{im}=\frac{\beta_i\sigma_m}{\sqrt{\beta_i^2\sigma_m^2+\sigma_{\varepsilon_i}^2}} .$$ If the market's volatility $\sigma_m$ rises while specific risks stay the same, every correlation $\rho_{im}$ and (for betas of the same sign) every $\rho_{ij}$ increases towards 1.

Why. $\operatorname{Cov}(r_i,r_j)=\operatorname{Cov}(\beta_ir_m+\varepsilon_i,\beta_jr_m+\varepsilon_j)=\beta_i\beta_j\sigma_m^2$, since all the other covariances are zero; divide by $\sigma_i\sigma_j$. Writing $\rho_{im}=1/\sqrt{1+\sigma_{\varepsilon_i}^2/(\beta_i^2\sigma_m^2)}$ shows it increases with $\sigma_m$, and $\rho_{ij}=\rho_{im}\rho_{jm}$. ∎  This is one simple mechanism behind the observation that correlations rise in market crises: when one big common shock dominates the news, every share moves with it, and diversification across shares helps least exactly when it is wanted most. (Real crises add other effects — forced selling, liquidity — that a fixed-beta model ignores.)

✏️ Example 8 — Five bank shares in calm and in crisis
Model each of five bank shares as having beta 1 and specific volatility 20% a year, with independent residuals. Compare a calm market (index volatility 15%) with a crisis (40%): find the correlation between two banks, the volatility of one bank, and the volatility of an equal-weighted portfolio of all five.
Calm $\sigma_i^2=15^2+20^2=625$, $\sigma_i=25\%$; $\rho_{ij}=225/625=0.36$. Portfolio: market part $15^2=225$, specific part $20^2/5=80$ (five independent residuals averaged), total $305$, $\sigma_p\approx17.5\%$ — diversification cut risk by 30%.
Crisis $\sigma_i^2=40^2+20^2=2000$, $\sigma_i\approx44.7\%$; $\rho_{ij}=1600/2000=0.80$. Portfolio: $1600+80=1680$, $\sigma_p\approx41.0\%$ — only 8% less than a single bank.
Lesson The banks' correlation rose from 0.36 to 0.80 although nothing about the banks changed. Holding several shares of one sector diversifies only specific risk; in a market-wide fall they behave almost like one share. Diversifying across asset classes with genuinely different drivers is the subject of Chapter 4.

Correlation is not causation. A high correlation between two series says they move together in the data, not that one drives the other. Two shares may both follow interest rates; a "leading indicator" may be correlated with the index in one sample by chance; and with hundreds of candidate indicators, some will show impressive correlations purely by luck (data snooping, §7.3). Worse, the correlation of two price levels that both trend upward is high even if their returns are unrelated — always correlate returns.

Interactive 2 — Fund League Table: Ranking Skill or Ranking Luck?

Three hypothetical funds are simulated with known true properties: A — high beta (β 1.4, no alpha), B — quality (β 0.75, true alpha 2% a year) and C — index fund (β 1, alpha −0.3% from costs). Dots show each fund's sample return and volatility; the line from the risk-free point (purple) through a dot has slope equal to its Sharpe ratio; hollow circles mark the true values. Choose a measure: the readouts rank the funds on this sample. Press Play to draw new histories and see how often the fund that is truly best on that measure actually comes first. (Simulated data; illustrative risk-free rate.)

6.5%
A — high beta—
B — quality—
C — index fund—
truly best on this measure—
truly best ranked 1st—
⚠️ Common mistakes and myths
  1. "Low beta means low risk." Beta measures market risk only. A low-beta share can be very volatile because of company-specific risk (Example 4).
  2. "Five shares, so I am diversified." Five shares in one sector are highly correlated; they share most of their risk (Example 8).
  3. Correlating prices instead of returns. Any two upward-trending price series look highly correlated. Use returns.
  4. "Correlation proves the indicator drives the market." Correlation is not causation, and a correlation found by searching many indicators is often luck.
  5. "Correlations are fixed numbers." They change with market conditions and tend to rise in crises (Theorem 3.10), just when diversification is needed most. Stress-test with higher correlations.
  6. Comparing raw returns across funds with different risk. In a rising market a higher-beta fund usually returns more; that is not skill. Use risk-adjusted measures.
  7. Comparing ratios computed differently. Different periods, data frequencies, risk-free rates or benchmarks make Sharpe ratios incomparable.
  8. Treating a small Sharpe difference as meaningful. With 5 years of data the standard error of a Sharpe ratio is about 0.45 (Theorem 3.9); 0.52 vs 0.61 is noise.
  9. "Alpha is the fund's return minus the index's." That is the active return. Jensen's alpha adjusts for beta: a fund with beta 1.3 that beat a rising index by 2% may have negative alpha.
  10. Using Sharpe for an option-selling or other strongly skewed strategy. Volatility misses rare large losses (§3.1); look at Sortino, skewness and maximum drawdown too.
📈 Market case — Meera chooses between Fund P and Fund Q, the right way

Meera downloads five years of monthly returns for both hypothetical flexi-cap funds and their common benchmark index, and computes (all annualised; illustrative risk-free rate 6.5%):

returnvolatilitybetadownside dev.tracking error
Fund P18%22%1.2014%8.0%
Fund Q15%14%0.859%4.5%
benchmark14%16%1——
  1. Sharpe. P: $11.5/22\approx0.52$; Q: $8.5/14\approx0.61$; benchmark $7.5/16\approx0.47$.
  2. Sortino (target $r_f$). P: $11.5/14\approx0.82$; Q: $8.5/9\approx0.94$.
  3. Treynor. P: $11.5/1.2\approx9.58\%$; Q: $8.5/0.85=10.0\%$; benchmark $7.5\%$.
  4. Jensen's alpha. P: $18-(6.5+1.2\times7.5)=2.5\%$; Q: $15-(6.5+0.85\times7.5)=2.125\%$.
  5. Information ratio. P: $(18-14)/8=0.50$; Q: $(15-14)/4.5\approx0.22$.
  6. Is any of it significant? Theorem 3.9: $\operatorname{SE}\approx\sqrt{(1+0.52^2/2)/5}\approx0.48$ for P and $\approx0.49$ for Q. The Sharpe gap of 0.09 is a fifth of one standard error. The funds also share most of their risk (same market), but even so five years cannot separate them.

What she learns. P's higher return came largely from higher beta in a rising market; per unit of total or market risk Q did slightly better, while P earned slightly more alpha and a higher IR. No measure separates them convincingly. The mathematics tells her what she cannot conclude: that either manager is better. So she turns to what she can know: expense ratios (a certain, compounding cost, §2.2), whether each fund's beta and drawdowns suit how much of her money it will hold, how consistently it has followed its stated style, and — if she cannot find a convincing reason to pay for active management — a low-cost index fund. Interactive 2 shows how often the truly better fund fails to top a five-year league table.

Exercises

All data are hypothetical or simulated; risk-free rates are illustrative. Volatility, covariance and tracking error use the divisor $n-1$, downside deviation the divisor $n$; give ratios to two decimals.

I
Covariance and correlation.
  • 1 Monthly returns (%) of an index and a share are $x=1,-2,3,0,3$ and $y=2,-1,1,-1,4$. Compute their covariance and correlation.
  • 2 Two assets have annual volatilities 15% and 20% and covariance $0.018$ (in decimal units). Find their correlation.
  • 3 A 60:40 portfolio holds equity (volatility 18%) and debt (volatility 6%) with correlation 0.1. Find its volatility.
  • 4 Two shares' prices both rise steadily over ten years, but their monthly returns are unrelated. What correlation would you expect from (a) the prices, (b) the returns? Which is meaningful?
  • 5 Show that if $\rho=1$, a 50:50 portfolio's volatility is exactly the average of the two volatilities.
  • 6 Returns are converted from percent to decimals and a constant 0.5% is subtracted from each. What happens to the covariance and to the correlation?
1 Both means are 1; deviations $u=0,-3,2,-1,2$ and $v=1,-2,0,-2,3$; $\sum uv=14$, $\sum u^2=\sum v^2=18$. Covariance $14/4=3.5\ (\%^2)$; correlation $14/18\approx0.78$.
2 $\rho=0.018/(0.15\times0.20)=0.6$.
3 $\sigma_p^2=0.36\times324+0.16\times36+2\times0.6\times0.4\times0.1\times18\times6=116.64+5.76+5.184=127.58$, so $\sigma_p\approx11.30\%$.
4 (a) Close to $+1$, because both price series are dominated by the same upward trend. (b) Near $0$. Only (b) is meaningful: it measures whether the shares move together month by month, which is what matters for risk.
5 $\sigma_p^2=\tfrac14\sigma_1^2+\tfrac14\sigma_2^2+\tfrac12\sigma_1\sigma_2=\left(\tfrac{\sigma_1+\sigma_2}{2}\right)^2$. Perfectly correlated assets give no diversification.
6 Subtracting a constant changes nothing; dividing both series by 100 divides the covariance by $10{,}000$. The correlation is unchanged (Theorem 3.6, part 2).
II
Beta, alpha and $R^2$.
  • 1 The covariance of a share with the index is 0.0270 and the index's variance is 0.0225 (annual, decimals). Find the share's beta.
  • 2 A share has correlation 0.7 with the index, volatility 28%, and the index volatility is 16%. Find its beta, $R^2$ and specific volatility.
  • 3 A regression on monthly returns gives $\hat a=0.1\%$ and $\hat\beta=0.85$. What return does the line predict in a month when the index falls 4%?
  • 4 Roughly what annual alpha does a monthly alpha of 0.1% correspond to?
  • 5 For the data of I.1, find the regression line of $y$ on $x$.
  • 6 A share has beta 0. Is it risk-free? Explain.
  • 7 A portfolio holds 50%, 30% and 20% in shares with betas 1.3, 0.9 and 0.6. Find the portfolio beta (beta is linear in the weights).
1 $\beta=0.0270/0.0225=1.2$.
2 $\beta=0.7\times28/16=1.225$; $R^2=0.49$; specific variance $=(1-R^2)\sigma^2$, so $\sigma_\varepsilon=28\%\sqrt{0.51}\approx20.0\%$.
3 $0.1+0.85\times(-4)=-3.3\%$ (on average).
4 About $12\times0.1\%=1.2\%$ a year (compounded, $1.001^{12}-1\approx1.2\%$).
5 $\hat\beta=14/18\approx0.78$, $\hat a=1-0.78\times1\approx0.22\%$: $\hat y\approx0.22+0.78x$.
6 No. Beta 0 means its returns are uncorrelated with the index, not that they do not vary: all its risk is specific. In a large diversified portfolio that risk largely cancels, but held alone it can be very risky.
7 $0.5\times1.3+0.3\times0.9+0.2\times0.6=0.65+0.27+0.12=1.04$.
III
Risk-adjusted ratios (illustrative risk-free rate 6.5% a year unless stated).
  • 1 A fund returned 14% with volatility 16%. Find its Sharpe ratio.
  • 2 A strategy's monthly Sharpe ratio is 0.2. Annualise it.
  • 3 Monthly returns (%) are $3,-2,4,-1,0,2$. With target $\tau=0$, find the downside deviation and the monthly Sortino ratio.
  • 4 A fund returned 13% with beta 0.8. Find its Treynor ratio.
  • 5 A fund returned 16% with beta 1.2 while the market returned 13%. Find Jensen's alpha.
  • 6 A fund's average active return is 1.5% a year and its tracking error 4%. Find its information ratio.
  • 7 A fund has Sharpe ratio 0.5 and volatility 12%. An investor puts half her money in it and half in a deposit earning the risk-free rate. Find her portfolio's expected excess return, volatility and Sharpe ratio.
1 $(14-6.5)/16\approx0.47$.
2 $0.2\sqrt{12}\approx0.69$.
3 Mean $=6/6=1\%$; shortfalls below 0: $-2,-1$; $\text{DD}=\sqrt{(4+1)/6}\approx0.913\%$; Sortino $=1/0.913\approx1.10$ a month.
4 $(13-6.5)/0.8\approx8.13\%$.
5 $16-[6.5+1.2\times(13-6.5)]=16-14.3=1.7\%$.
6 $1.5/4=0.375$.
7 The fund's excess return is $0.5\times12\%=6\%$. Half in the fund: excess return $3\%$, volatility $6\%$, Sharpe $3/6=0.5$ — unchanged (Theorem 3.8).
IV
Comparing correctly: noise, crises and causation.
  • 1 A fund's 3-year Sharpe ratio is 0.8. Find the approximate standard error and a 95% interval. Can we be confident its true Sharpe ratio is positive?
  • 2 How many years of data are needed for the standard error of a Sharpe ratio near 0.5 to be 0.25?
  • 3 A share has beta 1 and specific volatility 25%. Find its correlation with the index when the index volatility is 12% and when it is 36%.
  • 4 A newspaper notes that a share index rose in years when a cricket team won a series. Explain why this is no basis for investing.
  • 5 A large-cap fund has $R^2=0.98$ against its benchmark, a tracking error of 1.5%, and charges a much higher expense ratio than an index fund. What does this suggest, and which ratio would you check?
1 $\operatorname{SE}\approx\sqrt{(1+0.32)/3}\approx0.66$; interval $0.8\pm1.3$, about $-0.5$ to $2.1$. No — three years cannot rule out a true Sharpe ratio of zero.
2 $(1+0.125)/T=0.0625\Rightarrow T=18$ years.
3 $\rho=\dfrac{\sigma_m}{\sqrt{\sigma_m^2+625}}$: at 12%, $\sqrt{144/769}\approx0.43$; at 36%, $\sqrt{1296/1921}\approx0.82$.
4 There is no plausible causal link, the sample of years is small, and among the thousands of things one could correlate with an index some will match by chance. A correlation found by searching is likely to vanish in future data.
5 It behaves almost exactly like its benchmark (a "closet index" fund) while charging active fees. Check its information ratio after costs: with a small tracking error and a fee gap of a percentage point or more, the IR is likely to be negative.
🎯 Practice corner NISM-styleCase studyInterview

E1. NISM-style The Sharpe ratio of a mutual fund scheme measures excess return over the risk-free rate per unit of (A) beta (B) standard deviation (C) downside deviation (D) tracking error.

Reason Sharpe divides by total volatility. (A) is Treynor, (C) Sortino, (D) belongs to the information ratio.
Answer (B).

E2. NISM-style A stock has a beta of 1.2 with respect to the market index. If the index rises by 5%, the expected change in the stock, ignoring alpha, is (A) +5% (B) +6% (C) +4.17% (D) +1.2%.

Compute $1.2\times5\%=6\%$. (C) divides instead of multiplying.
Answer (B).

E3. NISM-style Which measure is most suitable for judging how consistently an active fund manager has outperformed the scheme's benchmark? (A) Standard deviation (B) Beta (C) Information ratio (D) $R^2$.

Reason The information ratio divides the average active return by its volatility (tracking error): high when outperformance is large and steady. Standard deviation and beta measure risk; $R^2$ measures how closely the fund follows the benchmark, not whether it beats it.
Answer (C).

E4. Case study Six monthly returns (%) of a benchmark index and of Fund F are: benchmark $2,-3,4,1,-2,4$; Fund F $3,-4,4,2,-3,4$. (a) Find Fund F's beta, monthly alpha (raw-return regression) and $R^2$. (b) What does the line predict for a month in which the benchmark falls 5%? (c) Find the monthly active returns, the annualised tracking error and the information ratio. (d) Summarise the fund in one sentence.

(a) Both means are 1%. Deviations $u=1,-4,3,0,-3,3$ and $v=2,-5,3,1,-4,3$: $\sum uv=52$, $\sum u^2=44$, $\sum v^2=64$. $\hat\beta=52/44\approx1.18$; $\hat a=1-1.18\times1\approx-0.18\%$ a month; $R^2=52^2/(44\times64)\approx0.96$.
(b) $-0.18+1.18\times(-5)\approx-6.09\%$.
(c) Active returns $1,-1,0,1,-1,0$: mean $0$, monthly TE $=\sqrt{4/5}\approx0.894\%$, annualised $0.894\%\times\sqrt{12}\approx3.10\%$; IR $=0$.
(d) Over these months Fund F matched the benchmark's average return while taking more market risk (beta 1.18) — a negative alpha — so on this (tiny) sample it added no value; six months prove nothing either way.

E5. Case study Three hypothetical portfolios over the same period (annualised; illustrative risk-free rate 6%; market return 14%, market volatility 16%): K returned 16% with volatility 24% and beta 1.3; L returned 12% with volatility 12% and beta 0.6; M (an index fund) returned 14% with volatility 16% and beta 1.0. (a) Compute each one's Sharpe ratio, Treynor ratio and Jensen's alpha. (b) Which would you prefer as your only investment, and which as a small addition to an already diversified portfolio? (c) What is wrong with choosing K because it "returned the most"?

(a) Sharpe: K $10/24\approx0.42$, L $6/12=0.50$, M $8/16=0.50$. Treynor: K $10/1.3\approx7.69\%$, L $6/0.6=10\%$, M $8\%$. Alpha: K $16-(6+1.3\times8)=-0.4\%$, L $12-(6+0.6\times8)=1.2\%$, M $0$.
(b) Only investment: total risk matters, so Sharpe — L and M tie at 0.50, both ahead of K. As a small addition to a diversified portfolio: Treynor and alpha — L is clearly best (positive alpha, highest Treynor).
(c) K's higher return is more than explained by its beta of 1.3 in a market that beat the risk-free rate by 8%; its alpha is negative. Leveraging the index fund M to beta 1.3 would have given $6+1.3\times8=16.4\%$ with less specific risk.

E6. Interview Shares A and B each have correlation 0.5 with the market index. What is the lowest possible correlation between A and B? (Hint: a correlation matrix must be positive semidefinite.)

Idea Write each share's standardised return as $0.5\,M+\sqrt{1-0.25}\,U$ with $M$ the standardised market return and $U$ a part uncorrelated with $M$. Then $\rho_{AB}=0.25+0.75\,\operatorname{corr}(U_A,U_B)$.
Range Since $\operatorname{corr}(U_A,U_B)\in[-1,1]$, $\rho_{AB}\in[0.25-0.75,\,0.25+0.75]=[-0.5,\,1]$. (Equivalently, the determinant of the $3\times3$ correlation matrix must be $\ge0$.)
Answer $-0.5$. A good answer adds that "both are correlated with the market" does not force them to be correlated with each other; in the one-factor model of Theorem 3.10, with uncorrelated residuals, $\rho_{AB}=0.25$.
🧑‍🏫 For the teacher

Suggested plan — 2 lectures × 60 min.

  1. Moving together. Open with the 📍 situation: ask whether "beta 0.6 = low risk" is true and why five bank shares fell together. Covariance and correlation (Definition 3.6) with Example 1 on the board; Cauchy–Schwarz for $|\rho|\le1$ and the two-asset variance (Theorem 3.6), Example 2. The regression line by calculus (Definition 3.7, Theorem 3.7), Examples 3–4. Project Interactive 1: press Play at 24 months and let the class watch beta estimates jump; then 120 months; then switch to crisis and ask why the correlation rose although the share did not change (Theorem 3.10, Example 8).
  2. Risk-adjusted performance. Sharpe and Sortino (Definition 3.8), Treynor, alpha and IR (Definition 3.9), leverage invariance (Theorem 3.8); Examples 5–7 and the 🧭 box on which ratio answers which question. Then the noise in rankings (Theorem 3.9): project Interactive 2 with 5 years and Play for a minute — how often does the truly best fund top the table? — then 30 years. Close with the 📈 Market case and Practice corner E5.

Class activity. "Fund league": each student picks fund A, B or C in Interactive 2 before a new 5-year history is drawn; whoever's fund ranks first by Sharpe scores a point. After ten rounds compare the scores with the true ranking — and discuss what this means for fund advertisements that quote last year's rank.

Misconceptions to watch for. Low beta = low risk; correlating prices; active return confused with alpha; ignoring the noise in Sharpe ratios; believing correlations are stable; reading causation into correlation.

Suggested homework. I(1, 3, 4, 6), II(2, 5, 6, 7), III(3, 5, 7), IV(1, 3, 5); Practice corner E2, E4, E6.

🛠️ Do it yourselfSpreadsheet / Python mini-project · about 60 minutes

Project. Estimate the beta, alpha and $R^2$ of two mutual-fund schemes against their benchmark from published NAV and index data, compute their Sharpe, Sortino, Treynor and information ratios, and check whether the difference between them exceeds the noise.

Why it matters. Factsheets and rating sites report these ratios with different conventions and periods. Computing them yourself shows exactly what each one measures and how fragile a ranking can be.

What to do.

  1. Download five years of month-end NAVs for two schemes in the same category (AMC or AMFI website) and the month-end values of their benchmark's total-return index (index provider's website). Put dates in A, fund 1 in B, fund 2 in C, the benchmark in D.
  2. Monthly returns in F, G, H: =B3/B2-1 and so on. Risk-free rate per month in a cell rf, e.g. =(1+annual_rate)^(1/12)-1 with an annual T-bill yield you look up.
  3. For each fund: =SLOPE(G:G,H:H) (beta), =INTERCEPT(G:G,H:H) (monthly alpha, raw returns), =RSQ(G:G,H:H), =CORREL(G:G,H:H). Draw the scatter plot with a linear trendline and "display equation".
  4. Sharpe (annualised): =(AVERAGE(F:F)-rf)/STDEV.S(F:F)*SQRT(12). Downside deviation: =SQRT(SUMPRODUCT((F3:F62<rf)*(F3:F62-rf)^2)/COUNT(F3:F62)); Sortino likewise. Treynor: =12*(AVERAGE(F:F)-rf)/beta.
  5. Active returns =F3-H3 in a new column; tracking error =STDEV.S(...)*SQRT(12); information ratio =12*AVERAGE(...)/TE.
  6. Standard error of each Sharpe ratio: =SQRT((1+S^2/2)/5). Is the difference between the two funds larger than one standard error?

Python version: import numpy as np; b, a = np.polyfit(rm, rf1, 1); r2 = np.corrcoef(rm, rf1)[0,1]**2; sharpe = (rf1.mean()-rfm)/rf1.std(ddof=1)*12**0.5.

Hand in: the two scatter plots with fitted lines, a table of all ratios side by side with the factsheet's figures, and a paragraph: which fund would you call "better", on what evidence, and how confident can you be?

Going further. Compute a rolling 36-month beta and correlation (SLOPE/CORREL over a moving window) and chart them: do they change over time, and do correlations jump in months when the benchmark fell sharply?

Instructor note: students' betas and Sharpe ratios will differ from factsheets because of the period, data frequency, risk-free rate and whether the price index or total-return index was used; finding which convention reproduces the factsheet is a good exercise in itself. Regressing on raw rather than excess returns changes the intercept noticeably but the slope only slightly. Expect Sharpe ratio differences between funds of one category to be well inside one standard error.

🧪 Self-check quiz

0/5 correct

Q1. A share has correlation 0.6 with the index, volatility 30%, and the index volatility is 15%. Its beta is

Q2. You put 40% of your money in a fund with Sharpe ratio 0.6 and 60% in a deposit earning the risk-free rate. Your portfolio's Sharpe ratio is

Q3. In the market model, if the index becomes much more volatile while each share's specific risk stays the same, correlations between shares

Q4. The Sortino ratio differs from the Sharpe ratio because its denominator

Q5. A share's regression on the index has $R^2=0.25$. This means