Chapter 3 · The Statistics of Returns

3.1 Return, Volatility, Fat Tails and Drawdowns

How much does a fund really swing? Daily returns and volatility, the square-root-of-time rule, histograms, skewness and fat tails, maximum drawdown and recovery maths, and why the compound return is the average minus about $\sigma^2/2$.

📍 The situation

Rohan, 26, works in Bengaluru and runs two SIPs: one in "Fund A — Large Cap Index, Direct Growth" and one in "Fund B — Small Cap, Direct Growth" (both hypothetical). One Tuesday Fund B's NAV falls 4.2% in a single day and his office group chat fills with opinions. One friend says: "Relax — the factsheet says Fund B's volatility is 27%, so the most it can lose in a year is 27%." Another says: "Fund B has averaged 17% a year, Fund A only 13%, so B will make you far richer — it doubles every four years." A third forwards a chart showing that a fund like B once took years to get back to an old high.

Each claim is a statement about the statistics of returns, and each is partly or wholly wrong. To judge them Rohan needs four tools, which are the four parts of this section. How much do returns swing? — volatility, and how to convert a daily figure into a yearly one with the square-root-of-time rule. How often do big days happen? — histograms, skewness, kurtosis and the "fat tails" that make a 4% day far more common than the bell curve suggests. How far can it fall, and how long to recover? — maximum drawdown and the recovery arithmetic in which −50% needs +100%. And why does the average return overstate what you actually earn? — volatility drag, the gap of about $\sigma^2/2$ between the average return and the compound return. The mathematics is school statistics (mean, standard deviation, the normal curve) plus one inequality and one Taylor expansion.

⚠️ About the numbers on this page

Every price, NAV and return on this page is either a small hypothetical data set chosen to make the arithmetic clear or simulated in your browser by a seeded random-number generator. None of it is the history of a real index, fund or share, and nothing here is a forecast or a recommendation. The funds and companies are hypothetical; the 252-trading-days-a-year figure is a convention. When you apply these tools to a real fund, use its published NAV history (see 🛠️ Do it yourself) and remember that past volatility and drawdowns describe the past only.

Returns and Volatility: from Days to Years

A price series is converted into a return series before anything else is done with it, because returns are comparable across shares of different prices and across time. Rohan's Fund A NAV moving from ₹100 to ₹101 and a share moving from ₹2,000 to ₹2,020 are the same event: a 1% return.

📘 Definition 3.1 — Period returns, mean return and volatility

Let $P_0,P_1,\dots,P_n$ be closing prices (or NAVs) at the ends of equal periods — trading days, weeks or months — with dividends reinvested. The period return is $$r_t=\frac{P_t}{P_{t-1}}-1,\qquad t=1,\dots,n.$$ The mean return is $\bar r=\frac1n\sum_{t=1}^n r_t$, and the volatility is the sample standard deviation of the returns, $$s=\sqrt{\frac{1}{n-1}\sum_{t=1}^n\left(r_t-\bar r\right)^2}.$$ Volatility is quoted per period ("daily volatility 1.1%") or, more often, annualised (Theorem 3.1). Its square $s^2$ is the variance.

Finance theory often uses the log return $\log(P_t/P_{t-1})$ instead (§1.3 Measuring Returns); for daily data the two differ by a few hundredths of a percent and everything on this page works for either.

The divisor $n-1$ (not $n$) makes $s^2$ an unbiased estimate of the true variance — spreadsheets call this STDEV.S. With hundreds of daily returns the difference is negligible.

✏️ Example 1 — Ten days of a bank share
A hypothetical share, Alpha Bank Ltd, closes on 11 consecutive trading days at ₹1,000, 1,012, 1,005, 1,021, 1,018, 998, 1,004, 1,026, 1,019, 1,031 and 1,040. Find the ten daily returns, their mean, the daily volatility and the annualised volatility (use 252 trading days a year). What would "annualising the mean" by multiplying by 252 give, and should anyone believe it?
Returns $r_1=1012/1000-1=1.200\%$, then $-0.692\%$, $1.592\%$, $-0.294\%$, $-1.965\%$, $0.601\%$, $2.191\%$, $-0.682\%$, $1.178\%$, $0.873\%$.
Mean The sum is $4.003\%$, so $\bar r\approx0.400\%$ a day.
Volatility The deviations $r_t-\bar r$ are $0.80,\,-1.09,\,1.19,\,-0.69,\,-2.36,\,0.20,\,1.79,\,-1.08,\,0.78,\,0.47$ (in %); their squares add to $14.57$ (in $\%^2$). So $s^2=14.57/9\approx1.619$ and $s\approx1.273\%$ a day.
Annualise $\sigma_{\text{ann}}=1.273\%\times\sqrt{252}\approx1.273\times15.875\approx20.2\%$ a year.
The mean? $252\times0.400\%\approx101\%$ a year — absurd. Ten good days say almost nothing about the long-run mean: by Theorem 3.2 the standard error of the daily mean is $1.273\%/\sqrt{10}\approx0.40\%$, as large as the mean itself. Ten days are enough for a rough volatility, nowhere near enough for an expected return.
📐 Theorem 3.1 (Formula) — Annualising: the square-root-of-time rule

Suppose the returns of successive periods are independent with the same mean $\mu$ and standard deviation $\sigma$, and there are $m$ periods in a year ($m\approx252$ trading days, $52$ weeks or $12$ months). Then, to first order (ignoring compounding within the year), $$\mu_{\text{ann}}\approx m\,\mu,\qquad \sigma_{\text{ann}}=\sigma\sqrt m .$$ More generally, over a horizon of $h$ periods the mean scales like $h$ and the volatility like $\sqrt h$. For daily data, $\sigma_{\text{ann}}=\sigma_{\text{day}}\sqrt{252}\approx15.87\,\sigma_{\text{day}}$; conversely $\sigma_{\text{day}}=\sigma_{\text{ann}}/\sqrt{252}$.

Why $\sqrt{\ }$? Use log returns, which add over time: the log return of the year is the sum $X_1+\cdots+X_m$ of the period log returns. Means always add, so the mean of the sum is $m\mu$. For independent variables the variances add too, $\operatorname{Var}(X_1+\cdots+X_m)=m\sigma^2$, so the standard deviation is $\sigma\sqrt m$. (For simple returns the same holds to first order, since the year's simple return is close to the sum of the period returns.) ∎  The key word is independent: if good days tend to follow good days (momentum) the yearly volatility is larger than $\sigma\sqrt m$; if markets tend to reverse (mean reversion) it is smaller. The 252 is a convention for the number of trading days in a year, not a law.

✏️ Example 2 — Translating volatility between days, weeks, months and rupees
(a) Fund A's daily volatility is 1.1%; Fund B's monthly volatility is 5%. Annualise both. (b) A factsheet reports an annualised volatility of 16%. What is the daily and the weekly volatility? (c) For a ₹10,00,000 portfolio with 16% annual volatility, how big is a "one-sigma" day in rupees, and if daily returns were normal, on roughly how many days a year would the portfolio move by more than one sigma, and by more than two?
(a) $1.1\%\times\sqrt{252}\approx17.5\%$; $\ 5\%\times\sqrt{12}\approx17.3\%$. The two funds are about equally volatile, though the raw numbers look very different.
(b) Daily $16\%/\sqrt{252}\approx1.01\%$; weekly $16\%/\sqrt{52}\approx2.22\%$. A handy rule: divide an annual volatility by 16 to get the daily one.
(c) One sigma $\approx1.008\%$ of ₹10 lakh $\approx$ ₹10,079. For a normal distribution $P(|Z|\gt1)\approx31.7\%$ and $P(|Z|\gt2)\approx4.6\%$, so about $0.317\times252\approx80$ days a year beyond ±₹10,000 and about $0.046\times252\approx11.5$ days beyond ±₹20,000.
Check Real return series usually have more very large days than this normal count, and slightly fewer moderate ones — the fat tails of the next subsection.

How well can we know the mean and the volatility? Example 1 hinted at something important: from the same data, volatility is estimated far more precisely than the mean. This is why risk models trust volatility estimates while good analysts are humble about expected returns.

📐 Theorem 3.2 (Rule) — Standard errors of the mean return and of volatility

Let $r_1,\dots,r_n$ be independent returns with mean $\mu$ and standard deviation $\sigma$.

  1. The sample mean $\bar r$ has standard error $\operatorname{SE}(\bar r)=\sigma/\sqrt n$. If the data cover $T$ years, the annualised mean has standard error $$\operatorname{SE}(\text{annual mean})\approx\frac{\sigma_{\text{ann}}}{\sqrt T},$$ whatever the sampling frequency — daily data do not pin the mean down better than yearly data over the same $T$ years.
  2. For roughly normal returns the sample volatility has standard error $\operatorname{SE}(s)\approx\sigma/\sqrt{2(n-1)}$, which does shrink as the sampling frequency rises.

An approximate 95% interval for a quantity is its estimate $\pm1.96$ standard errors.

Why (part 1). $\operatorname{Var}(\bar r)=\frac1{n^2}\operatorname{Var}(r_1+\cdots+r_n)=\frac1{n^2}\,n\sigma^2=\frac{\sigma^2}n$ by independence. With $m$ periods a year and $T$ years, $n=mT$; the annualised mean $m\bar r$ has standard error $m\sigma/\sqrt{mT}=\sigma\sqrt m/\sqrt T=\sigma_{\text{ann}}/\sqrt T$ — the $m$'s cancel. Part 2 is a standard large-sample result for normal data, which we quote without proof. ∎

✏️ Example 3 — How sure is a ten-year average?
A fund's ten years of history show an average return of 14% a year and an annualised volatility of 18%. (a) Give an approximate 95% interval for its true long-run mean return. (b) The volatility was estimated from 2,520 daily returns. Give an approximate 95% interval for the true volatility. (c) How many years of data would make the standard error of the mean 2%?
(a) $\operatorname{SE}=18\%/\sqrt{10}\approx5.69\%$, so the interval is $14\%\pm1.96\times5.69\%\approx14\%\pm11.2\%$: roughly 3% to 25% a year. Ten years of data cannot tell a mediocre fund from an excellent one by its average alone.
(b) $\operatorname{SE}(s)\approx18\%/\sqrt{2\times2519}\approx0.25\%$, so about $18\%\pm0.5\%$ (if returns were normal; fat tails widen this somewhat, but it stays far narrower than (a)).
(c) $18/\sqrt T=2$ gives $T=81$ years.
Lesson A fund's past average return is a very noisy estimate of its future expected return; its past volatility is a far better guide to its future volatility (though volatility itself changes over time). Treat "it averaged 17%" with caution — §3.2 applies the same idea to Sharpe ratios.

The Shape of Returns: Histograms, Skewness and Fat Tails

Mean and volatility describe the centre and the spread of the returns. A histogram — the number of days whose return falls in each small interval — shows the whole shape. The benchmark shape is the normal (bell) curve, under which about 68% of days lie within one standard deviation of the mean, 95% within two and 99.7% within three. Two numbers measure how a real histogram departs from it.

📘 Definition 3.2 — Skewness and kurtosis

For returns $r_1,\dots,r_n$ with mean $\bar r$, let $m_k=\frac1n\sum_{t=1}^n(r_t-\bar r)^k$ be the $k$-th central moment ($m_2$ is the variance with divisor $n$). Then

  • skewness $=\dfrac{m_3}{m_2^{3/2}}$ — negative when the left tail is longer (occasional large falls, many small gains), positive when the right tail is longer;
  • kurtosis $=\dfrac{m_4}{m_2^{2}}$, and excess kurtosis $=$ kurtosis $-\,3$.

Both are unit-free. A normal distribution has skewness $0$ and kurtosis $3$ (excess kurtosis $0$). Spreadsheet functions SKEW and KURT use small-sample corrections, and KURT already reports the excess kurtosis; for long series the corrections are negligible.

📘 Definition 3.3 — Fat tails

A return distribution has fat tails (is leptokurtic) if its excess kurtosis is positive: compared with a normal distribution of the same mean and volatility, very large moves (beyond about 3σ) are much more frequent, and the centre is usually taller and narrower, so moderate moves are a little rarer. Daily returns of shares, indices and funds are typically fat-tailed, and share-index returns are often negatively skewed.

Because kurtosis uses fourth powers, a single extreme day can dominate it — which is exactly what it is designed to detect, and also why it is unstable in short samples.

✏️ Example 4 — One crash month in a year of a small-cap fund
Ten monthly returns (in %) of a hypothetical small-cap fund are $3,\,2,\,-1,\,4,\,2,\,-12,\,5,\,3,\,1,\,3$. Compute the mean, the moments $m_2,m_3,m_4$, the skewness and the excess kurtosis. How much of $m_4$ comes from the single month of −12%?
Mean The sum is $10$, so $\bar r=1\%$ and the deviations are $2,\,1,\,-2,\,3,\,1,\,-13,\,4,\,2,\,0,\,2$.
Moments Sums of squares, cubes and fourth powers of the deviations: $212$, $-2088$ and $28964$. Dividing by $n=10$: $m_2=21.2$, $m_3=-208.8$, $m_4=2896.4$ (in $\%^2,\%^3,\%^4$).
Shape $\sqrt{m_2}\approx4.60\%$; skewness $=-208.8/21.2^{3/2}\approx-208.8/97.6\approx-2.14$; kurtosis $=2896.4/21.2^2\approx6.44$, excess kurtosis $\approx3.44$. Strongly negatively skewed and fat-tailed.
Who is responsible? The −12% month contributes $13^4=28561$ of the $28964$ — about $98.6\%$ of $m_4$ — and more than all of $m_3$ (the other months' cubes add to $+109$). Remove it and the series looks almost normal. (Spreadsheet SKEW and KURT give about $-2.54$ and $7.05$ because of their small-sample corrections.)
📐 Theorem 3.3 (Rule) — How often are large moves? The normal benchmark and Chebyshev's bound

(a) Any distribution (Chebyshev). If a return $X$ has mean $\mu$ and finite standard deviation $\sigma$, then for every $k\gt0$ $$P\big(|X-\mu|\ge k\sigma\big)\le\frac1{k^2}.$$ (b) Normal distribution. If $X$ is normal, the two-sided probabilities $P(|X-\mu|\gt k\sigma)$ and the expected number of such days in 10 years ($2{,}520$ trading days) are:

$k$normal probabilitydays in 10 years (normal)Chebyshev boundfat-tailed model*
131.7%800100%580 days
24.55%11525%120 days
30.27%6.811.1%33 days
40.0063%0.166.25%12 days
50.000057%0.00144%5.3 days

*A common fat-tailed model: Student's $t$ distribution with 4 degrees of freedom, rescaled to the same volatility. It is an illustration, not a fitted model of any real market.

Why Chebyshev holds. Let $A$ be the event $|X-\mu|\ge k\sigma$. Then $$\sigma^2=E\big[(X-\mu)^2\big]\ge E\big[(X-\mu)^2\,\mathbf 1_A\big]\ge k^2\sigma^2\,P(A),$$ because the squared deviation is non-negative everywhere and at least $k^2\sigma^2$ on $A$. Divide by $k^2\sigma^2$. ∎  Chebyshev is a guaranteed ceiling for any distribution; real markets sit between the normal row and that ceiling. The table's last column shows the practical point: under fat tails a 4σ day is not a once-in-sixty-years event but something to expect about once a year.

Interactive 1 — Histogram of Simulated Daily Returns: Bell Curve or Fat Tails?

Ten years (2,520 days) of daily returns are simulated from the model you choose, all with the same volatility. The blue bars are the histogram; the green curve is the normal curve with the sample's own mean and volatility; bars outside ±3σ turn red. Switch the vertical scale to log to see the tails: there the normal curve is a parabola that plunges, while fat-tailed data form a straight-sided "tent". Press Play to watch the histogram build up day by day; 🎲 draws a new sample. (Simulated data, not a real index.)

16%
days simulated—
daily vol · annualised—
skewness—
excess kurtosis—
days beyond ±3σ (normal expects)—
worst day—
✏️ Example 5 — How rare is a 5% fall?
In a simulated market, an index fund with daily volatility 1% (about 16% a year) falls 5% in one day. (a) If daily returns were normal with mean 0, how often would a fall of 5% or more happen? (b) Under the fat-tailed model of Theorem 3.3 (Student $t$ with 4 d.f., same volatility)? (c) What is the most that any distribution with this volatility could allow for $P(|X|\ge5\%)$?
(a) A 5% fall is $z=-5$. One-sided $P(Z\le-5)\approx2.87\times10^{-7}$. With 252 days a year that is one such day every $1/(252\times2.87\times10^{-7})\approx13{,}800$ years.
(b) A unit-variance $t_4$ variable is $T/\sqrt2$ with $T\sim t_4$, so $P(X\le-5)=P(T\le-5\sqrt2)=P(T\le-7.07)\approx1.06\times10^{-3}$: about one day in 950, i.e. roughly once every 3.8 years.
(c) Chebyshev: $P(|X|\ge5\sigma)\le1/25=4\%$ — about 10 days a year at most, for both tails together. No real market comes near this ceiling, but it shows the normal answer is not a law of nature.
Moral The same 5% day is "impossible" under one model and "every few years" under another with the same volatility. Risk numbers built on the normal curve (such as parametric VaR, §7.4) understate the danger of rare large losses; the volatility alone does not tell you how fat the tails are.

Drawdowns: How Far It Falls and How Long It Takes to Recover

Volatility treats a +2% day and a −2% day alike and says nothing about order. Investors, however, experience risk as the fall from the highest value they have seen — "my ₹5 lakh became ₹3.5 lakh". That is the drawdown, and it is what makes people sell at the bottom.

📘 Definition 3.4 — Drawdown, maximum drawdown and recovery

For a value (NAV) series $V_0,V_1,\dots,V_n$ define the running peak $M_t=\max_{s\le t}V_s$ and the drawdown $$D_t=\frac{V_t}{M_t}-1\ \le 0 .$$ The maximum drawdown is $\text{MDD}=-\min_t D_t$, the largest percentage fall from a peak to a later trough (reported as a positive number, "a 24% drawdown"). A stretch of time with $D_t\lt0$ is a period under water; the recovery time of a drawdown is the time from its trough until the old peak is regained.

📐 Theorem 3.4 (Formula) — Recovery arithmetic
  1. After a fall of $d$ ($0\lt d\lt1$), the gain needed to get back to the peak is $$g=\frac{1}{1-d}-1=\frac{d}{1-d},$$ which is larger than $d$ and tends to infinity as $d\to1$: −10% needs +11.1%, −30% needs +42.9%, −50% needs +100%, −75% needs +300%.
  2. Growing at a constant $c$ a year from the trough, the recovery takes $\displaystyle T=\frac{\log\frac{1}{1-d}}{\log(1+c)}$ years.
  3. A gain and a loss of the same percentage $x$, in either order, leave $(1+x)(1-x)=1-x^2\lt1$: you end below where you started.

Why. After the fall the value is $(1-d)M$; to return to $M$ it must be multiplied by $1/(1-d)$, a gain of $1/(1-d)-1=d/(1-d)$. Since $1-d\lt1$, $g=d/(1-d)\gt d$. Part 2 solves $(1+c)^T=1/(1-d)$ by logarithms, exactly as for doubling times (§1.1). Part 3 multiplies growth factors, which is how returns combine. ∎

✏️ Example 6 — Reading a drawdown off a NAV series
Month-end NAVs (₹) of a hypothetical equity fund over one year are: 50.0, 53.2, 55.0, 51.7, 46.2, 41.8, 44.0, 48.9, 53.6, 56.1, 54.3, 58.0 (months 0 to 11). Find the running peak and the drawdown each month, the maximum drawdown, the gain that was needed to recover, and the time under water.
Peaks The running peak is 50.0, 53.2, then 55.0 from month 2 to month 8, 56.1 in months 9–10 and 58.0 in month 11.
Drawdowns Months 3–8: $51.7/55-1=-6.0\%$, $-16.0\%$, $-24.0\%$, $-20.0\%$, $-11.1\%$, $-2.5\%$; month 10: $54.3/56.1-1\approx-3.2\%$; all other months $0$.
Maximum MDD $=24.0\%$, from the peak of 55.0 (month 2) to the trough of 41.8 (month 5). Recovering needed $55/41.8-1\approx31.6\%$ — check: $0.24/0.76\approx0.316$ (Theorem 3.4).
Under water From month 3 to month 8 (six month-ends below the old peak); the trough-to-recovery time was 4 months (month 5 to month 9). Over the year the fund rose $58/50-1=16\%$ — the same 16% an investor who bought in month 2 and panicked in month 5 turned into a 24% loss.
✏️ Example 7 — How long to get back?
A portfolio falls 20%, 35% or 50% from its peak. For each, find the gain needed to recover and the recovery time if it then grows steadily at 10% or at 12% a year.
Gains $0.2/0.8=25\%$; $\ 0.35/0.65\approx53.8\%$; $\ 0.5/0.5=100\%$.
At 10% $T=\log(1/0.8)/\log1.1\approx2.34$ years; $\log(1/0.65)/\log1.1\approx4.52$ years; $\log2/\log1.1\approx7.27$ years.
At 12% about $1.97$, $3.80$ and $6.12$ years.
Interpret Recovery time grows much faster than the size of the fall: going from a 20% to a 50% drawdown triples the recovery time. And this assumes smooth growth afterwards — real recoveries are themselves volatile, so an investor should be prepared for longer. Avoiding deep drawdowns (through diversification and asset allocation, Chapter 4) is worth more than the arithmetic of averages suggests.

Volatility Drag: Why the Average Return Overstates Growth

In §1.3 (Measuring Returns) we saw that the arithmetic mean of yearly returns is not the CAGR: a +50% year followed by a −50% year averages 0% but loses 25%. The gap is not an accident of that example; it is driven by volatility, and its size can be estimated.

📘 Definition 3.5 — Volatility drag

For period returns $r_1,\dots,r_n$ (each $\gt-1$) let $A=\frac1n\sum r_t$ be the arithmetic mean and $G=\left[\prod(1+r_t)\right]^{1/n}-1$ the geometric mean (the compound growth rate per period, i.e. the CAGR for yearly returns). The volatility drag is $A-G$: the part of the average return that is lost to the ups and downs.

📐 Theorem 3.5 (Formula) — Geometric mean ≈ arithmetic mean − σ²/2

With $A$, $G$ as in Definition 3.5 and $\sigma^2=\frac1n\sum(r_t-A)^2$:

  1. $G\le A$, with equality only when all the returns are equal;
  2. when the returns are small (a few percent per period, or moderate yearly returns), $$G\approx A-\frac{\sigma^2}{2}.$$ So doubling the volatility quadruples the drag. At $A=12\%$ a year, volatility $15\%$ costs about $1.1\%$ a year, $25\%$ costs about $3.1\%$, and $40\%$ costs $8\%$.

Why. (1) is the AM–GM inequality applied to the growth factors $1+r_t\gt0$: their geometric mean $1+G$ is at most their arithmetic mean $1+A$ — the result of §1.3 that the average return is never below the CAGR. (2) Take logarithms and use the Taylor expansion $\log(1+r)\approx r-\frac{r^2}{2}$ (the calculus link): $$\log(1+G)=\frac1n\sum\log(1+r_t)\approx A-\frac{1}{2n}\sum r_t^2=A-\frac{\sigma^2+A^2}{2},$$ using $\frac1n\sum r_t^2=\sigma^2+A^2$. Exponentiating, $1+G\approx1+L+\frac{L^2}{2}$ with $L=A-\frac{\sigma^2+A^2}{2}$, and keeping terms up to second order, $G\approx A-\frac{\sigma^2+A^2}{2}+\frac{A^2}{2}=A-\frac{\sigma^2}{2}$. ∎  §3.3 meets the same $\sigma^2/2$ as the drift correction of geometric Brownian motion.

✏️ Example 8 — Same average, different wealth
Fund S returns exactly 12% every year. Fund V alternates +40% and −16%. (a) Find each fund's arithmetic mean, volatility (divisor $n$) and CAGR. (b) Compare the CAGR of V with the estimate $A-\sigma^2/2$. (c) What does ₹1,00,000 become in each fund after 10 years?
(a) S: $A=12\%$, $\sigma=0$, CAGR $12\%$. V: $A=(40-16)/2=12\%$, $\sigma=28\%$ (each return is 28 points from the mean); over two years the factor is $1.40\times0.84=1.176$, so CAGR $=\sqrt{1.176}-1\approx8.44\%$.
(b) $A-\sigma^2/2=12\%-0.28^2/2=12\%-3.92\%=8.08\%$, close to the true $8.44\%$ (the approximation is cruder for returns as large as 40%).
(c) S: $1{,}00{,}000\times1.12^{10}\approx$ ₹3,10,585. V: $1{,}00{,}000\times1.176^5\approx$ ₹2,24,924 — about ₹85,700 less, with the same average return.
Moral When a factsheet or an advertisement quotes an "average annual return", ask for the CAGR. Lower volatility is worth real money even at the same average — one reason diversification (Chapter 4) raises long-run growth.
Interactive 2 — A Simulated Fund: Drawdowns Under Water and the Volatility Drag

Twenty years of monthly returns are simulated with the average return and volatility you choose. Top panel: the fund's value (₹1 lakh at the start, blue), its running peak (green) and the path the same average return would give with no volatility (dashed). Bottom panel: the drawdown — the red area is time "under water". Keep the shocks fixed and raise the volatility: the path ends lower and dives deeper. Press Play to draw the path month by month; 🎲 draws new shocks. (Simulated data.)

12%22%
CAGR with no volatility—
CAGR predicted (Thm 3.5)—
CAGR in this simulation—
maximum drawdown—
gain needed to recover it—
longest time under water—
⚠️ Common mistakes and myths
  1. "Volatility 27% means I can lose at most 27% in a year." Volatility is a standard deviation, not a limit. Even under the normal model a two-sigma bad year is $17\%-2\times27\%=-37\%$, roughly a 1-in-40 event, and drawdowns within a year can be deeper still.
  2. Annualising volatility by multiplying by 252 (or 12). Volatility scales with $\sqrt{252}$ (or $\sqrt{12}$); only the mean scales with 252. A 1% daily volatility is about 16% a year, not 252%.
  3. Treating the $\sqrt{\text{time}}$ rule as exact. It needs independent returns; momentum, mean reversion and changing volatility all break it.
  4. "The normal curve says this crash should happen once in 10,000 years." Then the normal curve is the wrong model. Real return series have fat tails; a 4σ or 5σ day is rare, not impossible (Example 5).
  5. Trusting skewness and kurtosis from a short sample. One extreme month can create — or its absence can hide — almost all of the kurtosis (Example 4).
  6. "−30% then +30% gets me back." $0.7\times1.3=0.91$. A fall of $d$ needs a gain of $d/(1-d)$: −50% needs +100%.
  7. "Average return 17% means my money grows at 17%." The compound growth is about $A-\sigma^2/2$; with 27% volatility that is about 13.4%, not 17% (Theorem 3.5).
  8. "Its volatility was low last year, so it is safe." Volatility clusters — calm periods are followed by turbulent ones — and a short calm sample says little about the next crisis.
  9. "This fund has never had a big drawdown." Maximum drawdown depends on the period examined; a young fund may simply never have been tested by a bad market.
  10. "Past volatility or past average return is a forecast." They are descriptions of the past. The mean in particular is estimated very imprecisely (Theorem 3.2).
📈 Market case — Rohan's two funds through a simulated decade

A risk report for Rohan's two funds, computed from ten years of simulated monthly data, gives: Fund A (large-cap index) average return $A=13\%$ a year, volatility $16\%$; Fund B (small cap) $A=17\%$, volatility $27\%$. He works through his friends' claims.

  1. Daily swings. $16\%/\sqrt{252}\approx1.01\%$ and $27\%/\sqrt{252}\approx1.70\%$ a day. Fund B's −4.2% day was $4.2/1.70\approx2.5$ of its daily sigmas: under the normal model about $0.7\%$ of days (one-sided), so one or two days a year, and more often with fat tails. Unpleasant, but not a sign that anything is broken.
  2. "At most 27% a year?" No. A one-in-forty bad year under the normal model is $17\%-2\times27\%=-37\%$ for B and $13\%-2\times16\%=-19\%$ for A, and peak-to-trough drawdowns can exceed one calendar year's loss.
  3. Growth, not average. By Theorem 3.5 the compound rates are about $13-0.16^2/2\times100=11.7\%$ for A and $17-0.27^2/2\times100\approx13.4\%$ for B. ₹10 lakh would grow over 10 years to about ₹30.3 lakh and ₹35.0 lakh. B's advantage is 1.7 points a year, not 4 — and "doubles every four years" (17%) is really about every $72/13.4\approx5.4$ years.
  4. Is B's higher average even real? Standard errors over 10 years (Theorem 3.2): $16/\sqrt{10}\approx5.1\%$ for A and $27/\sqrt{10}\approx8.5\%$ for B. The 4-point gap between the averages is well inside the noise; ten years cannot prove B's expected return is higher.
  5. The drawdown test. Suppose B falls 45% from a peak — a fall well within what a fund this volatile can suffer. It then needs $0.45/0.55\approx81.8\%$ to recover, which at B's estimated 13.4% compound rate takes $\log(1/0.55)/\log1.134\approx4.8$ years. Rohan asks himself whether he would keep his SIP running through that.

What he learns. The mathematics does not say which fund to hold; it turns vague worries into numbers. B's extra expected growth (if any) is smaller than its average suggests, uncertain, and comes with much deeper falls. How much of each to hold depends on his horizon and his ability to sit through a drawdown — a portfolio question for Chapter 4 and a planning question for §8.1. Interactive 2 lets him see what a 27% volatility path can look like.

Exercises

All data are hypothetical or simulated. Use 252 trading days, 52 weeks or 12 months a year; volatility uses the divisor $n-1$ unless a part says otherwise; give percentages to two decimals.

I
Returns, volatility and the square-root-of-time rule.
  • 1 A share closes at ₹200, 204, 201, 207 and 205 on five days. Find the four daily returns, their mean and the daily volatility.
  • 2 Annualise a daily volatility of 0.9% and a monthly volatility of 6%.
  • 3 A fund's annualised volatility is 22%. Find its daily and weekly volatility.
  • 4 A ₹5,00,000 portfolio has annual volatility 15%. How large, in rupees, is a one-sigma day? A two-sigma day?
  • 5 Using the data of part 1, someone annualises the mean daily return by multiplying by 252. Explain, using a standard error, why the result is meaningless.
  • 6 A fund's 5-year average return is 15% a year with volatility 20%. Give an approximate 95% interval for its true mean return.
  • 7 How many years of data are needed for the standard error of the mean annual return to be 2% when the volatility is 18%? Does switching from monthly to daily data help?
1 Returns $2.00\%$, $-1.47\%$, $2.99\%$, $-0.97\%$; mean $\approx0.64\%$; volatility $s\approx2.19\%$ a day.
2 $0.9\%\sqrt{252}\approx14.29\%$; $\ 6\%\sqrt{12}\approx20.78\%$.
3 Daily $22\%/\sqrt{252}\approx1.39\%$; weekly $22\%/\sqrt{52}\approx3.05\%$.
4 Daily volatility $15\%/\sqrt{252}\approx0.945\%$, so one sigma $\approx$ ₹4,725 and two sigma $\approx$ ₹9,449.
5 $252\times0.64\%\approx160\%$ a year. But $\operatorname{SE}(\bar r)=2.19\%/\sqrt4\approx1.09\%$ a day, larger than the mean itself: four days carry essentially no information about the expected return.
6 $\operatorname{SE}=20\%/\sqrt5\approx8.94\%$; interval $15\%\pm17.5\%$, roughly $-2.5\%$ to $32.5\%$.
7 $18/\sqrt T=2\Rightarrow T=81$ years. No: by Theorem 3.2 the standard error of the annual mean is $\sigma_{\text{ann}}/\sqrt T$ whatever the frequency; only a longer history helps. (Daily data do sharpen the volatility estimate.)
II
Histograms, skewness, kurtosis and tails.
  • 1 Five monthly returns are $1,\,3,\,-1,\,2,\,-5$ (%). Compute $m_2$, the skewness and the kurtosis (divisor $n$).
  • 2 Nine months return exactly 0% and the tenth returns 10%. Compute the skewness and kurtosis. What does this show about kurtosis?
  • 3 If daily returns are normal with mean 0 and volatility 1.2%, on how many days a year should the move exceed ±2.4%?
  • 4 In the same model, how often should a fall of 3.6% or more occur?
  • 5 For any return distribution, at most what fraction of days can lie 4 or more standard deviations from the mean? Prove it.
  • 6 A simulated 10-year daily series shows 19 days beyond ±3σ. Compare with the normal expectation and say what it suggests.
  • 7 A strategy earns a small profit on most days and occasionally loses heavily. What sign of skewness would you expect, and why is "a high average return and low volatility" not enough to judge it?
1 Mean $0$; $m_2=(1+9+1+4+25)/5=8$, $m_3=(1+27-1+8-125)/5=-18$, $m_4=(1+81+1+16+625)/5=144.8$. Skewness $-18/8^{3/2}\approx-0.80$; kurtosis $144.8/64\approx2.26$ (excess $-0.74$). Short samples can even show kurtosis below 3.
2 Mean $1$; deviations nine times $-1$ and once $9$: $m_2=9$, $m_3=(-9+729)/10=72$, $m_4=(9+6561)/10=657$. Skewness $72/27\approx2.67$, kurtosis $657/81\approx8.11$. A single outlier creates a very large kurtosis.
3 $2.4\%=2\sigma$: $P(|Z|\gt2)\approx4.55\%$, so about $0.0455\times252\approx11.5$ days a year.
4 $P(Z\le-3)\approx0.00135$; $0.00135\times252\approx0.34$ a year, i.e. about once every 2.9 years.
5 At most $1/4^2=6.25\%$, by Chebyshev: $\sigma^2\ge E[(X-\mu)^2\mathbf 1_{|X-\mu|\ge4\sigma}]\ge16\sigma^2P(|X-\mu|\ge4\sigma)$.
6 The normal model expects $0.0027\times2520\approx6.8$ days; 19 is nearly three times as many. It points to fat tails (positive excess kurtosis) — or to volatility that changes over time, which produces fat tails in the pooled data.
7 Negative skewness: many small gains, a long left tail. Volatility weights a rare large loss only through its square, spread over many quiet days, so such strategies (for example selling far out-of-the-money options, §6.2) can show a high average and low volatility right up to the loss. Look at skewness, worst days and maximum drawdown as well.
III
Drawdowns and recovery.
  • 1 NAVs are 100, 110, 99, 105, 88, 96, 112. Find the running peak, each drawdown and the maximum drawdown.
  • 2 What gain is needed to recover from a 30% fall? From a 60% fall?
  • 3 After a 40% fall, how many years does recovery take at a steady 11% a year?
  • 4 A fund falls 25% and then rises 25%. What is the net result?
  • 5 Funds X and Y both gained 20% over three years. X's maximum drawdown was 10%, Y's 45%. Which would most investors have found easier to hold, and why does the end-to-end return hide the difference?
  • 6 Show that $g(d)=d/(1-d)$ is increasing and convex on $0\lt d\lt1$ by computing $g'(d)$ and $g''(d)$. What does $g''\gt0$ mean for an investor?
1 Peaks 100, 110, 110, 110, 110, 110, 112; drawdowns $0$, $0$, $-10\%$, $-4.55\%$, $-20\%$, $-12.73\%$, $0$. MDD $=20\%$ (110 to 88).
2 $0.3/0.7\approx42.86\%$; $\ 0.6/0.4=150\%$.
3 $\log(1/0.6)/\log1.11\approx4.89$ years.
4 $0.75\times1.25=0.9375$: a net loss of $6.25\%=0.25^2$ (Theorem 3.4, part 3).
5 X: a 10% dip is easy to sit through. Y: an investor saw almost half the money vanish and needed an 82% rebound; many would have sold near the bottom and locked in the loss. The start and end values say nothing about the path in between — that is exactly what drawdown measures.
6 $g(d)=\frac{1}{1-d}-1$, so $g'(d)=\frac{1}{(1-d)^2}\gt0$ and $g''(d)=\frac{2}{(1-d)^3}\gt0$. So each extra percentage point of loss requires a larger extra gain than the one before — deep losses are disproportionately hard to repair.
IV
Volatility drag.
  • 1 Yearly returns are +30%, −20%, +30%, −20%. Find the arithmetic mean and the CAGR.
  • 2 For the returns of part 1 compute $\sigma$ (divisor $n$) and the estimate $A-\sigma^2/2$. Compare with the exact CAGR.
  • 3 Estimate the compound growth rate of a fund with average return 14% and volatility 25%.
  • 4 A fund's average return is 8%. At roughly what volatility does its compound growth rate fall to zero?
  • 5 An index rises 10% one day and falls back exactly to its starting level the next. A product that delivers twice the index's daily return does what over the two days? Explain using volatility drag.
  • 6 Monthly returns have mean 1% and volatility 5%. Estimate the monthly geometric mean and the annual CAGR, and compare with the CAGR if every month returned exactly 1%.
1 $A=5\%$; CAGR $=(1.3\times0.8)^{1/2}-1=\sqrt{1.04}-1\approx1.98\%$.
2 $\sigma=25\%$; $5\%-0.25^2/2=5\%-3.125\%=1.875\%$, close to the exact $1.98\%$.
3 $14\%-0.25^2/2=14\%-3.125\%\approx10.9\%$ a year.
4 $0.08=\sigma^2/2\Rightarrow\sigma=\sqrt{0.16}=40\%$.
5 The index goes $+10\%$ then $-9.09\%$ (since $1.1\times(1/1.1)=1$). Twice the daily returns: $+20\%$ then $-18.18\%$, so $1.2\times0.8182\approx0.9818$ — a loss of about $1.8\%$ while the index is flat. Doubling the daily returns quadruples $\sigma^2/2$, so the drag grows faster than the doubled average.
6 Monthly $G\approx1\%-0.05^2/2=0.875\%$; CAGR $\approx1.00875^{12}-1\approx11.0\%$, against $1.01^{12}-1\approx12.7\%$ with no volatility.
🎯 Practice corner NISM-styleCase studyInterview

E1. NISM-style A fund's daily standard deviation of returns is 1.2%. Assuming 252 trading days and independent returns, its annualised standard deviation is closest to (A) 302% (B) 19.0% (C) 14.4% (D) 1.2%.

Compute $1.2\%\times\sqrt{252}\approx1.2\times15.87\approx19.0\%$. (A) multiplies by 252; (C) multiplies by 12, treating the figure as monthly; (D) forgets to annualise.
Answer (B).

E2. NISM-style An investor's portfolio falls by 40%. The return needed to recover the original value is (A) 40% (B) 60% (C) 66.7% (D) 140%.

Compute $0.4/0.6\approx66.7\%$: ₹100 becomes ₹60, and ₹60 must grow by ₹40, i.e. by two-thirds.
Answer (C).

E3. NISM-style A return distribution that has more extreme observations than a normal distribution with the same standard deviation is best described as having (A) a high mean (B) negative skewness (C) positive excess kurtosis (D) low volatility.

Reason Kurtosis measures the weight of the tails; positive excess kurtosis means fat tails. Skewness measures asymmetry (a distribution can have fat tails on both sides with zero skew); the mean and the volatility are fixed by the comparison.
Answer (C).

E4. Case study A hypothetical share closes on seven consecutive Fridays at ₹500, 520, 494, 470, 488, 515 and 530. (a) Find the six weekly returns, their mean and the weekly volatility. (b) Annualise the volatility (52 weeks). (c) Find the maximum drawdown and the gain that was needed to recover from it. (d) Is the annualised mean $52\times\bar r$ a sensible estimate of the share's expected return?

(a) Returns $4.00\%$, $-5.00\%$, $-4.86\%$, $3.83\%$, $5.53\%$, $2.91\%$; mean $\approx1.07\%$; volatility $s\approx4.72\%$ a week.
(b) $4.72\%\times\sqrt{52}\approx34.1\%$ a year — a very volatile share.
(c) Peak ₹520, trough ₹470: MDD $=1-470/520\approx9.62\%$. Recovery needed $520/470-1\approx10.6\%$; it was regained at ₹530.
(d) $52\times1.07\%\approx56\%$ a year, with a standard error of about $34.1\%/\sqrt{6/52}\approx100\%$ (Theorem 3.2 with $T=6/52$ years). Meaningless: six weeks say something about volatility, nothing about the expected return.

E5. Case study ₹10,00,000 is invested for five years. Fund X returns +25%, −15%, +30%, −20%, +20% in successive years; Fund Y returns 8% every year. (a) Find each fund's arithmetic mean return and CAGR. (b) Compute the final values. (c) Compare X's CAGR with $A-\sigma^2/2$ (divisor $n$). (d) Find X's maximum drawdown on the year-end values.

(a) X: $A=(25-15+30-20+20)/5=8\%$; growth factor $1.25\times0.85\times1.30\times0.80\times1.20=1.326$, so CAGR $=1.326^{1/5}-1\approx5.81\%$. Y: $A=$ CAGR $=8\%$.
(b) X: ₹13,26,000. Y: $10{,}00{,}000\times1.08^5\approx$ ₹14,69,328 — ₹1,43,328 more, with the same average return.
(c) $\sigma\approx21.12\%$, so $8\%-0.2112^2/2\approx8\%-2.23\%=5.77\%$, very close to the true $5.81\%$.
(d) Year-end values: ₹12,50,000; 10,62,500; 13,81,250; 11,05,000; 13,26,000. The deepest fall is from ₹13,81,250 to ₹11,05,000: MDD $=20\%$ (the earlier fall was 15%).

E6. Interview A stock rises 10% on one day and falls 10% the next, alternating for a whole year of 252 trading days. Without a calculator, roughly what happens to ₹100? Its average daily return is zero — why is the result not ₹100?

Idea Each two-day pair multiplies by $1.1\times0.9=0.99$; there are 126 pairs, so the factor is $0.99^{126}$.
Estimate $\log0.99\approx-0.01$, so $0.99^{126}\approx e^{-1.26}\approx0.28$: about ₹28. (Exactly $0.2819$.)
Why Volatility drag: with daily $\sigma=10\%$ the geometric mean is about $0-\sigma^2/2=-0.5\%$ a day, and $252\times0.5\%=126\%$ of log growth is lost. A good answer adds that no real stock has 10% daily volatility for a year, but leveraged daily products in volatile, sideways markets suffer the same effect on a smaller scale.
🧑‍🏫 For the teacher

Suggested plan — 2 lectures × 60 min.

  1. Volatility and tails. Open with the 📍 situation and take a vote on the three friends' claims. Returns and volatility (Definition 3.1) with Example 1 worked on the board; the $\sqrt{\text{time}}$ rule (Theorem 3.1) as "variances add"; Example 2. Then the honesty point of Theorem 3.2 and Example 3: the mean is hard to know. Project Interactive 1: start with the normal model, switch to fat-tailed, then switch the scale to log and ask the class what the straight-sided tent means. Skewness and kurtosis (Definitions 3.2–3.3) with Example 4; Chebyshev and the tail table (Theorem 3.3), Example 5.
  2. Drawdowns and drag. Drawdowns (Definition 3.4) with Example 6 drawn as a running-peak picture; the recovery arithmetic (Theorem 3.4) and Example 7. Volatility drag (Theorem 3.5) from AM–GM and $\log(1+r)\approx r-r^2/2$, Example 8. Project Interactive 2: hold the shocks fixed and move volatility from 10% to 40% while the class predicts the CAGR with $A-\sigma^2/2$. Finish with the 📈 Market case and Practice corner E6.

Class activity. Give each group a different 🎲 sample in Interactive 1 (fat-tailed model) and ask them to record the excess kurtosis and the number of ±3σ days; collect the results on the board to show how much these statistics vary between samples of the same model — a live lesson in sampling error.

Misconceptions to watch for. Volatility as a maximum loss; annualising by 252 instead of $\sqrt{252}$; trusting a short-sample mean; believing the normal curve's "once in 10,000 years"; symmetric thinking about gains and losses; quoting the average return as the growth rate.

Suggested homework. I(1, 3, 6, 7), II(1, 2, 5, 6), III(1, 3, 5, 6), IV(1, 2, 5); Practice corner E1, E4, E5.

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

Project. Build a "risk report" for one mutual fund from its published daily NAV history: daily returns, annualised volatility, a histogram with the normal curve, skewness, excess kurtosis, a count of 3σ days, the drawdown chart and the volatility drag.

Why it matters. A factsheet gives one standard deviation. Your own sheet shows the whole shape — the tails and the drawdowns that the single number hides — and lets you check the factsheet.

What to do.

  1. Download several years of daily NAVs for any one scheme from the AMC's website or AMFI's NAV history page. Put dates in column A and NAVs in column B (oldest first, rows 2 onwards).
  2. Returns in C3 (copy down): =B3/B2-1. Statistics: =AVERAGE(C:C), =STDEV.S(C:C), annualised volatility =STDEV.S(C:C)*SQRT(252), =SKEW(C:C), =KURT(C:C) (already excess kurtosis).
  3. 3σ days: =COUNTIF(C:C,">"&(m+3*s))+COUNTIF(C:C,"<"&(m-3*s)) with m, s the cells holding the mean and volatility; compare with =0.0027*COUNT(C:C).
  4. Histogram: bins from −5% to +5% in steps of 0.25% in column E, counts with =FREQUENCY(C:C,E2:E42); normal expectation in G: =COUNT(C:C)*0.0025*NORM.DIST(E2,m,s,FALSE). Chart both; then repeat with a logarithmic vertical axis.
  5. Drawdown: running peak in D2 =B2, D3 =MAX(D2,B3); drawdown in F3 =B3/D3-1; maximum drawdown =-MIN(F:F). Chart column F as an area chart.
  6. Drag: from month-end NAVs compute monthly returns, their mean $A$ and volatility, the actual CAGR =(last/first)^(1/years)-1, and compare with $(1+A-\sigma^2/2)^{12}-1$.

Python version: r = nav.pct_change().dropna(); r.std()*252**0.5; r.skew(); r.kurt(); dd = nav/nav.cummax()-1; dd.min() (pandas).

Hand in: the histogram on both scales, the drawdown chart, a table of your statistics next to the factsheet's standard deviation, and three sentences on what the tails and the maximum drawdown tell an investor that the volatility does not.

Going further. Compute a rolling one-year volatility (=STDEV.S over the previous 252 returns, times SQRT(252)) and chart it: is volatility constant? Then standardise each return by the rolling volatility at the time and recompute the kurtosis — how much of the fat tail was just changing volatility?

Instructor note: factsheets differ in convention (daily vs monthly data, period length, divisor), so students' volatilities need not match the factsheet exactly; ask them to find out which convention reproduces it. Excel's SKEW/KURT include small-sample corrections and KURT subtracts 3 — a frequent source of "my kurtosis is 3 lower than yours". Standardising by rolling volatility usually removes a good part, but not all, of the excess kurtosis.

🧪 Self-check quiz

0/5 correct

Q1. A fund's daily volatility is 1%. With 252 trading days and independent returns, its annual volatility is about

Q2. After a 50% fall, the gain needed to get back to the previous peak is

Q3. A return series with large positive excess kurtosis has

Q4. A fund has an average annual return of 10% and an annual volatility of 20%. Its long-run compound growth rate is approximately

Q5. You want to estimate a fund's expected annual return more precisely. Which helps?