Chapter 3 · The Statistics of Returns

3.3 Random Walks and Monte Carlo Simulation of Prices

Nobody can tell you where the Nifty will be next year — but mathematics can tell you the odds. Random walks and the efficient-market idea, lognormal prices, geometric Brownian motion and its σ²/2 correction, and Monte Carlo simulation of price paths, target and loss probabilities, and the range of outcomes of a SIP.

📍 The situation

Meera, 34, a physiotherapist in Kochi, has ₹10,00,000 in "Fund N — Nifty 50 Index, Direct Growth" (a hypothetical fund) and a ₹10,000 monthly SIP in the same fund. Her investing app shows a cheerful line: "At 12% a year your SIP becomes ₹47.6 lakh in 15 years." Meera wants something more honest. What is the chance that her ₹10 lakh is worth less than ₹10 lakh a year from now? Ten years from now? What range of values is realistic, rather than one number? Is ₹47.6 lakh a likely outcome, a typical outcome or an average one — and how often would she end up below it?

Her cousin, who trades, has a different question: "The index has closed higher seven days in a row — the trend is clearly up. Isn't that proof prices are predictable?" Both questions are about randomness in prices. In this section we build the standard model of it — the random walk, its multiplicative version for prices (the lognormal model and geometric Brownian motion), and the computer experiment that answers questions no formula can, Monte Carlo simulation. This is where the probability of the course enters in earnest, and we will be just as careful about what the model cannot tell us.

⚠️ About the numbers on this page

The drift and volatility inputs (for example "expected return 12%, volatility 18% a year" for a broad equity index) are illustrative assumptions for our simulated market, not estimates of any real index's history and not forecasts. Every price path, chart and table is simulated in this page or computed from the model. Real markets have fat tails, jumps and changing volatility (§3.1), so model probabilities are a starting point, not a guarantee. People and funds are hypothetical; nothing here is investment advice.

Random Walks and the Efficient-Market Idea

Look at a chart of daily closing prices and you see wiggles that seem to have a mind of their own. The simplest mathematical model of such a wiggle is a walk in which every step is a fresh, independent random draw. Throughout this section $\log$ is the natural logarithm.

📘 Definition 3.10 — Random walk

Let $X_1,X_2,\dots$ be independent, identically distributed random variables (the steps) with mean $m$ and variance $s^2$. The sequence $$S_n=S_0+X_1+X_2+\cdots+X_n\qquad(n=0,1,2,\dots)$$ is a random walk started at $S_0$. It has drift $m$ per step; if $m=0$ it is driftless. In the simple symmetric random walk each step is $+1$ or $-1$ with probability $\tfrac12$ each.

The key property is that the future steps $X_{n+1},X_{n+2},\dots$ are independent of the path so far: knowing how the walk got to $S_n$ tells you nothing extra about where it goes next.

📐 Theorem 3.11 — Mean and spread of a random walk

For the random walk of Definition 3.10, $$E[S_n-S_0]=nm,\qquad \operatorname{Var}(S_n-S_0)=ns^2,\qquad \operatorname{SD}(S_n-S_0)=s\sqrt n.$$ If the steps are normal, $S_n-S_0\sim N(nm,\,ns^2)$ exactly; for any steps with finite variance, the central limit theorem makes $S_n-S_0$ approximately normal for large $n$.

Why. Expectation is additive, so $E[X_1+\cdots+X_n]=nm$. For variance we need independence: the variance of a sum of independent variables is the sum of their variances, so $\operatorname{Var}=s^2+\cdots+s^2=ns^2$. A sum of independent normals is normal. ∎ The drift grows like $n$ but the spread grows only like $\sqrt n$ — this is exactly the √time rule for annualising volatility (§3.1), and it explains why noise dominates short horizons and drift dominates long ones.

✏️ Example 1 — A coin-toss market
In a toy market a share moves ₹1 up or ₹1 down each day with probability ½ each, independently. After 100 days, find the expected change, its standard deviation, and the probability that the price has moved by ₹20 or more in either direction. How long does it take for the typical move to double?
Moments Each step has $m=0$ and $s^2=E[X^2]=1$. By Theorem 3.11 the change $S_{100}-S_0$ has mean ₹0 and SD $\sqrt{100}=$ ₹10.
Exact If $B$ is the number of up-days, $B\sim\text{Binomial}(100,\tfrac12)$ and the change is $2B-100$. A move of ₹20 or more means $B\le40$ or $B\ge60$: $$P=2\,P(B\le40)\approx0.0569.$$
Normal check The change takes only even values, so "$|S_{100}-S_0|\ge20$" is "$\gt18$"; with the midpoint 19, $2\Phi(-1.9)\approx0.0574$ — close to the exact value. (Without this continuity correction, $2\Phi(-2)\approx0.0455$.)
Doubling The SD is $\sqrt n$, so to double it from ₹10 to ₹20 you need $n=400$ days — four times as long. Distance travelled by a random walk grows like the square root of time.

Why should real prices behave like this? The classical answer is an economic argument, not a law of nature.

📘 Definition 3.11 — The efficient-market hypothesis (EMH)

A market is informationally efficient with respect to a set of information if prices already reflect that information, so that it cannot be used to earn returns above a fair reward for risk, after costs, in a consistent way. The three usual forms:

  • weak form — the information is the history of prices and volumes (so chart patterns alone should not give a consistent edge);
  • semi-strong form — all public information (results, news, announcements);
  • strong form — all information, including private information (insider trading is illegal precisely because this form fails).

The weak form suggests modelling changes in (log) prices as unpredictable from past changes — a random walk, possibly with a positive drift that rewards investors for bearing risk.

The argument: if everyone knew that a share would rise tomorrow, they would buy today, and the price would rise today. Predictable moves get traded away; what is left is news, and news is by definition unpredictable. How efficient real markets are is debated — researchers document anomalies, and costs, taxes and limits to arbitrage matter — so treat the EMH as a benchmark: any strategy claiming to beat it must show its edge survives costs and careful statistics (§7.3). Note what the EMH does not say: it does not say prices are "correct", that crashes cannot happen, or that investors cannot earn money — a positive drift is the reward for risk.

✏️ Example 2 — Drift is invisible from day to day
In our simulated market, an index's daily log returns are independent with mean 0.045% and standard deviation 1.1% (illustrative). Using the convention of 252 trading days a year, find the mean and SD of the one-year log return, the probability of an up day and of an up year (assume normal returns), and the number of trading days after which the expected drift equals one standard deviation of the noise.
One year Log returns add (§1.3), so the yearly log return is a 252-step random walk: mean $252\times0.045\%=11.34\%$, SD $1.1\%\times\sqrt{252}\approx17.46\%$.
Up day $P(R\gt0)=\Phi(0.045/1.1)=\Phi(0.041)\approx0.516$ — barely better than a coin.
Up year $\Phi(11.34/17.46)=\Phi(0.649)\approx0.742$.
Drift = noise After $n$ days the drift is $nm$ and the noise SD is $s\sqrt n$. They are equal when $\sqrt n=s/m$, i.e. $$n=\left(\frac{1.1}{0.045}\right)^2\approx598$$ trading days, about 2.4 years. On a daily or weekly chart you are looking almost entirely at noise; the drift reveals itself only over years.

Log Returns and Lognormal Prices

An additive random walk is a poor model for a price: it can wander below zero, and it treats a ₹50 move the same whether the share costs ₹500 or ₹5,000. Investors think in percentages. So we let the log price do the random walk: with daily (or monthly, or yearly) log returns $R_k=\log(S_k/S_{k-1})$ independent and identically distributed, $$\log S_n=\log S_0+R_1+\cdots+R_n,\qquad S_n=S_0\,e^{R_1+\cdots+R_n}.$$ The price is now always positive, and its moves are proportional to its level. By Theorem 3.11, $\log(S_n/S_0)$ is normal (exactly if the $R_k$ are normal, approximately otherwise). A variable whose logarithm is normal has a name.

📘 Definition 3.12 — Lognormal distribution

A positive random variable $V$ is lognormal with parameters $a$ and $b^2$ if $\log V\sim N(a,b^2)$. Equivalently $V=e^{a+bZ}$ with $Z\sim N(0,1)$. The parameters are the mean and variance of $\log V$, not of $V$.

📐 Theorem 3.12 (Formula) — Facts about a lognormal variable

If $V=e^{a+bZ}$ with $Z\sim N(0,1)$ and $b\gt0$, then

  1. $P(V\le x)=\Phi\!\left(\dfrac{\log x-a}{b}\right)$ for $x\gt0$, where $\Phi$ is the standard normal distribution function;
  2. the median is $e^{a}$, the mean is $E[V]=e^{a+b^2/2}$, and the mode (most likely value) is $e^{a-b^2}$, so mode $\lt$ median $\lt$ mean;
  3. the central interval with probability $1-2\alpha$ is $\left[e^{a-zb},\,e^{a+zb}\right]$ with $z=\Phi^{-1}(1-\alpha)$ (for 90%, $z\approx1.645$; for 95%, $z\approx1.96$).

Derivation. (1) and (3): $V\le x$ means $a+bZ\le\log x$, i.e. $Z\le(\log x-a)/b$, because $\log$ is increasing; the median is the case $Z\le0$. For the mean, complete the square in the normal integral: $$E[e^{bZ}]=\int_{-\infty}^{\infty}e^{bz}\frac{e^{-z^2/2}}{\sqrt{2\pi}}\,dz=e^{b^2/2}\int_{-\infty}^{\infty}\frac{e^{-(z-b)^2/2}}{\sqrt{2\pi}}\,dz=e^{b^2/2},$$ since the last integrand is the $N(b,1)$ density. Hence $E[V]=e^{a}E[e^{bZ}]=e^{a+b^2/2}$. The mode comes from differentiating the density of $V$; we will not need it. ∎ The lognormal is skewed to the right: a long tail of big winners pulls the mean above the median.

✏️ Example 3 — One year of a bank share
"Alpha Bank Ltd" (hypothetical) trades at ₹800. Suppose its one-year log return is normal with mean 8% and SD 30%. Find the median, mean and most likely price in a year, the probability that it ends below ₹800 and below ₹600, and the central 90% range. Is the range symmetric around ₹800?
Set-up $S_1=800\,e^{a+bZ}$ with $a=0.08$, $b=0.30$, so $\log S_1\sim N(\log800+0.08,\ 0.09)$.
Centre Median $800e^{0.08}\approx$ ₹866.63; mean $800e^{0.08+0.045}\approx$ ₹906.52; mode $800e^{0.08-0.09}\approx$ ₹792.04.
Losses $P(S_1\lt800)=\Phi(-0.08/0.30)$ $=\Phi(-0.267)\approx0.395$. $$P(S_1\lt600)=\Phi\!\left(\frac{\log0.75-0.08}{0.30}\right)=\Phi(-1.226)\approx0.110.$$
90% range $800e^{0.08\pm1.645\times0.30}$, i.e. from about ₹529 to about ₹1,420.
Shape Not symmetric: the range reaches ₹271 below ₹800 but ₹620 above it. A ±49% move in log terms is a fall of 39% but a rise of 64% in rupees. Even with a positive expected return, the chance of being below today's price after a year is almost 40%.

Geometric Brownian Motion and the σ²/2 Correction

A random walk moves in steps of one day. Make the steps shorter and shorter — a step of size $\pm\sqrt{\Delta t}$ every $\Delta t$ years — and, by the central limit theorem, the walk settles down to a continuous random curve whose increments are normal. That limit is the basic building block of continuous-time finance.

📘 Definition 3.13 — Brownian motion (Wiener process)

A standard Brownian motion $W(t)$, $t\ge0$, is a random function with $W(0)=0$, continuous paths, independent increments (changes over non-overlapping time intervals are independent) and normal increments: $W(t)-W(s)\sim N(0,\,t-s)$ for $s\lt t$. In particular $W(t)\sim N(0,t)$, so $W(t)=\sqrt t\,Z$ in distribution with $Z\sim N(0,1)$.

📘 Definition 3.14 — Geometric Brownian motion (GBM)

A price follows geometric Brownian motion with drift $\mu$ and volatility $\sigma\gt0$ (both per year) if $$S(t)=S_0\exp\!\Big(\big(\mu-\tfrac12\sigma^2\big)t+\sigma W(t)\Big),\qquad t\ge0.$$ We write $\nu=\mu-\tfrac12\sigma^2$ for the log drift (the median growth rate). In the shorthand of stochastic calculus this is the model $\dfrac{dS}{S}=\mu\,dt+\sigma\,dW$: over a short time the percentage change has mean $\mu\,dt$ and SD $\sigma\sqrt{dt}$.

GBM is the model behind the Black–Scholes formula (§6.3) and most textbook Monte Carlo. Here $\mu$ is the expected continuously compounded growth rate, $E[S(t)]=S_0e^{\mu t}$; the expected simple return over one year is $e^{\mu}-1$ (12.75% for $\mu=12\%$).

📐 Theorem 3.13 (Formula) — The distribution of a GBM price

Under GBM, for each $t\gt0$:

  1. $\log\dfrac{S(t)}{S_0}\sim N\!\left(\nu t,\ \sigma^2t\right)$ with $\nu=\mu-\tfrac12\sigma^2$ — so $S(t)$ is lognormal, and log returns over disjoint periods are independent;
  2. $\text{median}\,S(t)=S_0e^{\nu t}$ and $E[S(t)]=S_0e^{\mu t}$; their ratio is $e^{\sigma^2t/2}\gt1$;
  3. the typical (median) path grows at the continuously compounded rate $\nu=\mu-\tfrac12\sigma^2$, which is also the long-run growth rate of almost every path: $\tfrac1t\log\big(S(t)/S_0\big)\to\nu$ as $t\to\infty$.

Derivation. $\log(S(t)/S_0)=\nu t+\sigma W(t)$, and $\sigma W(t)\sim N(0,\sigma^2t)$, which gives (1); independence of increments of $W$ gives independence of log returns. Theorem 3.12 with $a=\nu t$, $b^2=\sigma^2t$ gives median $S_0e^{\nu t}$ and mean $S_0e^{\nu t+\sigma^2t/2}=S_0e^{\mu t}$. For (3), $\tfrac1t\log(S(t)/S_0)=\nu+\sigma W(t)/t$, and $W(t)/t$ has SD $1/\sqrt t\to0$ (in fact it tends to 0 for almost every path). ∎

Where the σ²/2 comes from. The average over many possible futures grows at $\mu$, but the typical future grows at $\mu-\sigma^2/2$. The gap is the volatility drag you met in §3.1 — the arithmetic mean return exceeds the geometric mean by about half the variance — now written exactly. A few very lucky paths, far out in the right tail, carry the mean; most paths do worse than the mean. (The $-\tfrac12\sigma^2$ is what Itô's lemma produces when you take the logarithm of $dS/S=\mu\,dt+\sigma\,dW$; we only need its consequence.)

📐 Theorem 3.14 (Rule) — Target and loss probabilities, and confidence bands

Under GBM with $\nu=\mu-\tfrac12\sigma^2$, for a horizon $t\gt0$ and a level $K\gt0$:

$$P\big(S(t)\ge K\big)=\Phi(d),\qquad d=\frac{\log(S_0/K)+\nu t}{\sigma\sqrt t}.$$

In particular the probability of a loss is $$P\big(S(t)\lt S_0\big)=\Phi\!\left(-\frac{\nu\sqrt t}{\sigma}\right),$$ which, when $\nu\gt0$, falls as the horizon grows but is never zero.

The central $(1-2\alpha)$ confidence band at time $t$ is $S_0\exp\!\big(\nu t\pm z\,\sigma\sqrt t\big)$ with $z=\Phi^{-1}(1-\alpha)$. In rupees the band widens with $t$; the band for the annualised growth rate, $\nu\pm z\sigma/\sqrt t$, narrows.

Why. $S(t)\ge K$ means $\nu t+\sigma\sqrt t\,Z\ge\log(K/S_0)$, i.e. $Z\ge-d$, and $P(Z\ge-d)=\Phi(d)$. Put $K=S_0$ for the loss probability; the band is Theorem 3.12(3). Dividing $\log(S(t)/S_0)$ by $t$ gives the annualised rate $\nu+\sigma Z/\sqrt t$. ∎

✏️ Example 4 — Meera's ₹10 lakh: chance of a loss, and the honest range
Model Meera's index fund as GBM with $\mu=12\%$ and $\sigma=18\%$ a year (illustrative inputs). For her ₹10,00,000, find (a) the probability of being below ₹10 lakh after 1, 5, 10 and 20 years; (b) the median and mean value after 10 years; (c) the central 90% range after 1 year and after 10 years.
Log drift $\nu=0.12-\tfrac12(0.18)^2=0.12-0.0162=0.1038$, i.e. 10.38% a year; and $\nu/\sigma\approx0.5767$.
(a) Loss $P=\Phi(-0.5767\sqrt t)$: 1 year $\Phi(-0.577)\approx28.2\%$; 5 years $\Phi(-1.290)\approx9.9\%$; 10 years $\Phi(-1.824)\approx3.4\%$; 20 years $\Phi(-2.579)\approx0.50\%$.
(b) 10 years Median $10e^{1.038}\approx$ ₹28.24 lakh; mean $10e^{1.2}\approx$ ₹33.20 lakh. The mean is $e^{0.162}\approx1.18$ times the median.
(c) Bands 1 year: $10e^{0.1038\pm1.645\times0.18}$, from ₹8.25 lakh to ₹14.92 lakh. 10 years: $\sigma\sqrt{10}\approx0.569$, so $10e^{1.038\pm0.936}$, from about ₹11.07 lakh to ₹72.01 lakh.
Read it Time lowers the probability of a loss, but the rupee range explodes: after 10 years the 90% band runs from barely above the starting amount to more than seven times it. "Long-term equity is safe" is not what the model says; it says "losses become less likely, outcomes become more spread out".
✏️ Example 5 — How likely is doubling, and how long to be fairly sure?
With the inputs of Example 4, what is the probability that Meera's ₹10 lakh is worth at least ₹20 lakh after 7 years? After how many years is that probability 50%, 75% and 90%?
7 years $K/S_0=2$: $$d=\frac{0.1038\times7-\log2}{0.18\sqrt7}=\frac{0.7266-0.6931}{0.4762}\approx0.070,$$ so $P=\Phi(0.070)\approx52.8\%$ — a coin toss.
50% $\Phi(d)=\tfrac12$ when $d=0$: $t=\log2/\nu=0.6931/0.1038\approx6.7$ years — the "rule of 72" doubling time at the median rate.
75%, 90% Solve $\dfrac{\nu t-\log2}{\sigma\sqrt t}=z$ with $z=0.674$ and $1.282$. Writing $x=\sqrt t$, this is the quadratic $\nu x^2-z\sigma x-\log2=0$. For 75%: $$x=\frac{0.1214+\sqrt{0.1214^2+4(0.1038)(0.6931)}}{2(0.1038)}\approx3.234,$$ so $t\approx10.5$ years. For 90%: $t\approx15.4$ years.
Lesson Going from "50–50" to "fairly sure" more than doubles the horizon. Goal planning that uses a single assumed rate silently aims at roughly the 50% (or worse — see Example 8) level.
Interactive 1 — Price Paths Under Geometric Brownian Motion

Each thin line is one simulated path of ₹100 invested (monthly steps of the exact GBM update; blue while above ₹100, red while below). The green band is the model's central 90% range and the green line its median $100e^{\nu t}$; the purple dashed line is the mean $100e^{\mu t}$; the amber line is the target. Readouts compare the simulated paths with Theorems 3.13–3.14 at the time shown. Play grows the paths, then draws a fresh set. Raise σ and watch the median fall away from the mean.

12.0%18%
time shown—
log drift ν = μ − σ²/2—
P(below ₹100): sim / theory—
P(at or above target): sim / theory—
median: sim / theory—
mean: sim / theory—
✏️ Example 6 — Same expected return, very different typical outcome
Two hypothetical index funds have the same drift $\mu=12\%$ but different volatilities: Fund P has $\sigma=15\%$, Fund Q has $\sigma=35\%$. Prove or disprove: "Because their expected returns are equal, a typical investor ends up with about the same amount after 20 years in either." Then show that if $\sigma\gt\sqrt{2\mu}$, the typical outcome shrinks even though the mean grows.
Means Both have $E[S(20)]/S_0=e^{0.12\times20}=e^{2.4}\approx11.02$.
Medians P: $\nu=0.12-0.01125=0.10875$, median multiple $e^{2.175}\approx8.80$. Q: $\nu=0.12-0.06125=0.05875$, median multiple $e^{1.175}\approx3.24$. The typical Fund Q investor ends with about 37% of what the typical Fund P investor has. The statement is false.
Losses 20-year loss probability: P $\Phi(-0.10875\sqrt{20}/0.15)=\Phi(-3.24)\approx0.06\%$; Q $\Phi(-0.05875\sqrt{20}/0.35)=\Phi(-0.75)\approx22.6\%$.
General claim The median $S_0e^{(\mu-\sigma^2/2)t}$ grows if and only if $\nu\gt0$, i.e. $\sigma^2\lt2\mu$. If $\sigma\gt\sqrt{2\mu}$ ($\approx49\%$ for $\mu=12\%$), then $\nu\lt0$ and the median tends to 0, while the mean $S_0e^{\mu t}$ still grows. Example: $\sigma=60\%$ gives $\nu=-6\%$; after 20 years ₹100 has median ₹30 but mean ₹1,102, and 67% of paths are below ₹100. ∎
Takeaway Volatility is not just "a bumpier ride to the same place". With returns compounding, it lowers the growth of the typical path — the σ²/2 correction. (This is one reason highly leveraged or very volatile products can disappoint even when their average return looks attractive.)

Monte Carlo Simulation: Paths, Targets and SIP Outcomes

For a lump sum under GBM, Theorem 3.14 answers everything. But many real questions have no tidy formula. Will a stop-loss be hit at some point during the year? What will a SIP be worth — a sum of many lognormal pieces, which is not lognormal? What if returns have fat tails? The universal method is to let the computer live through thousands of possible futures and count.

📘 Definition 3.15 — Monte Carlo estimate

To estimate $\theta=E[f(\text{path})]$ — for example a probability $p=P(\text{event})$, where $f$ is 1 on the event and 0 otherwise — simulate $N$ independent paths from the model and take the average $$\hat\theta_N=\frac1N\sum_{j=1}^{N}f(\text{path}_j).$$ For a probability, $\hat p_N$ is simply the fraction of simulated paths on which the event happens. Percentiles of the simulated outcomes (10th, 50th, 90th…) estimate the outcome's percentiles; plotted over time they form a fan chart.

📐 Theorem 3.15 (Rule) — How accurate is a Monte Carlo estimate?

If $f$ has finite variance $v$, then $E[\hat\theta_N]=\theta$ (no bias), $\hat\theta_N\to\theta$ as $N\to\infty$ (law of large numbers), and the Monte Carlo standard error is $\sqrt{v/N}$. For a probability, $$\text{SE}(\hat p_N)=\sqrt{\frac{p(1-p)}{N}},$$ and an approximate 95% confidence interval is $\hat p_N\pm1.96\,\text{SE}$. The error shrinks like $1/\sqrt N$: four times as many paths halve the error.

Why. The $f(\text{path}_j)$ are independent with mean $\theta$ and variance $v$, so their average has mean $\theta$ and variance $v/N$ (exactly as in Theorem 3.11); the central limit theorem makes it approximately normal. For an indicator, $v=p(1-p)$. ∎ Crucially, the standard error measures only simulation noise. It says nothing about whether the model and its inputs are right.

🧭 How to simulate GBM price paths
  1. Choose the step $\Delta t$ (1/252 for days, 1/12 for months) and the number of steps $n=T/\Delta t$.
  2. Make standard normals. From uniform random numbers $u_1,u_2$ in $(0,1)$, the Box–Muller formulas $$z_1=\sqrt{-2\log u_1}\,\cos(2\pi u_2),\quad z_2=\sqrt{-2\log u_1}\,\sin(2\pi u_2)$$ give two independent $N(0,1)$ values. (In a spreadsheet, NORM.S.INV(RAND()); in Python, rng.standard_normal() after rng = np.random.default_rng(seed).) Fix the seed so results can be reproduced.
  3. Step exactly in logs: $$S_{k+1}=S_k\exp\!\big(\nu\Delta t+\sigma\sqrt{\Delta t}\,z_k\big),\quad \nu=\mu-\tfrac12\sigma^2.$$ This has exactly the GBM distribution for any step size; the naive step $S_k(1+\mu\Delta t+\sigma\sqrt{\Delta t}z_k)$ is only approximate and can even go negative.
  4. Record what you need on each path (end value, minimum, SIP corpus…), repeat for $N$ paths, and compute fractions and percentiles.
  5. Report the Monte Carlo error (Theorem 3.15) and check the code against a case with a known answer (the lump-sum formulas of Theorem 3.14) before trusting it on a new question.
✏️ Example 7 — A stop-loss: touching ₹8 lakh versus ending below it
Meera says: "I'll sell if my ₹10 lakh ever drops to ₹8 lakh this year." With $\mu=12\%$, $\sigma=18\%$, compare (a) the probability of ending the year at or below ₹8 lakh with (b) the probability of touching ₹8 lakh at some point during the year. Use the classical formula for continuous monitoring and Monte Carlo with monthly, weekly and daily checks.
(a) End $b=\log0.8\approx-0.2231$: $$P=\Phi\!\left(\frac{-0.2231-0.1038}{0.18}\right)=\Phi(-1.816)\approx3.5\%.$$
(b) Formula For $X(t)=\nu t+\sigma W(t)$ and a level $b\lt0$, the reflection principle gives $$P\Big(\min_{s\le t}X(s)\le b\Big)=\Phi\!\left(\tfrac{b-\nu t}{\sigma\sqrt t}\right)+e^{2\nu b/\sigma^2}\,\Phi\!\left(\tfrac{b+\nu t}{\sigma\sqrt t}\right).$$ Here $e^{2\nu b/\sigma^2}=e^{-1.430}\approx0.239$ and $\Phi(-0.663)\approx0.254$, so $P\approx0.0347+0.0607\approx9.5\%$.
Monte Carlo Simulating 4,00,000 paths and checking the price only at the end of each period gives about 6.5% (monthly checks), 7.9% (weekly) and 8.8% (daily) — rising towards the continuous value as checks become more frequent. (SE ≈ 0.04 percentage points.)
Lesson A path that touches ₹8 lakh often recovers by year-end, so the touch probability is nearly three times the end probability. Questions about the path — stop-losses, drawdowns, margin calls — need path simulation, and the answer depends on how often you look.
✏️ Example 8 — Meera's SIP: one number versus the whole range
Meera invests ₹10,000 at the start of every month for 15 years. Model the fund as GBM with $\mu=12\%$, $\sigma=18\%$ (monthly log returns independent $N(\nu/12,\sigma^2/12)$). Compare (a) the app's "12% calculator" figure, (b) the exact expected corpus, and (c) the Monte Carlo distribution of the corpus. How often does the SIP end below the app's figure, and below the money invested?
(a) Calculator With $i=1.12^{1/12}-1$ (§1.2): $$FV=10{,}000\,\frac{(1+i)^{180}-1}{i}(1+i)\approx\text{₹}\,47.59\text{ lakh},$$ from ₹18,00,000 invested.
(b) Mean The instalment paid $k$ months before the end grows by a factor with mean $e^{\mu k/12}$ (Theorem 3.13), so $$E[FV]=10{,}000\sum_{k=1}^{180}e^{0.01k}\approx\text{₹}\,50.75\text{ lakh}.$$ This is the annuity-due formula at the mean rate $e^{0.12}-1=12.75\%$.
(c) Simulation With 10,00,000 simulated paths (exact monthly steps): median ≈ ₹44.4 lakh; 10th percentile ≈ ₹24.7 lakh; 90th percentile ≈ ₹84.0 lakh (5th–95th: ₹21.1–₹101.4 lakh). About 56% of paths end below the calculator's ₹47.6 lakh, 61% below the mean, 33% below twice the money invested and about 2.2% below the ₹18 lakh invested.
Read it The calculator figure is not "what you will get": it sits above the median, because a constant-rate calculation tracks the mean, and the mean is pulled up by a few very good paths. A fair statement is: "in this model, a typical outcome is about ₹44 lakh; one in ten outcomes is below about ₹25 lakh and one in ten above about ₹84 lakh."
Interactive 2 — Monte Carlo SIP Fan Chart: Goal Probability and Its Error

A SIP at the start of every month, with fund returns simulated from GBM (2,000 seeded paths, exact monthly steps). The blue fan shows the 10th–90th (light) and 25th–75th (dark) percentiles of the corpus year by year, with the median as a solid line; the amber dashed line is the expected (mean) corpus — what a constant-rate calculator at the mean rate shows; grey is the money put in; green is your goal. Play adds paths from 20 to 2,000: watch the goal probability settle and its Monte Carlo error shrink like $1/\sqrt N$.

₹10,00015
12.0%18%
paths used—
money invested—
mean corpus (exact)—
median corpus (sim)—
10th – 90th percentile—
P(below money invested)—
P(goal reached) ± SE—
📈 Market case — A ₹40 lakh education goal: which SIP gives which chance?

Kavya and Arjun want ₹40,00,000 in 12 years for their daughter's higher studies — already stated in future rupees, after allowing for education inflation (§1.1). They will invest at the start of each month in a broad equity index fund, which we model as GBM with $\mu=12\%$, $\sigma=18\%$ (illustrative). We compare a single-number plan with a Monte Carlo plan.

  1. The single-number plan. A calculator at "12% a year" ($i=1.12^{1/12}-1$, 144 months) asks for $A=40{,}00{,}000\big/\big[\tfrac{(1+i)^{144}-1}{i}(1+i)\big]\approx$ ₹12,983 a month.
  2. Its real chance. Simulating 10,00,000 paths of 144 monthly returns, that SIP reaches ₹40 lakh on only about 45% of paths. Its median corpus is about ₹38.0 lakh, with a 10th–90th percentile range of about ₹22.8–₹66.1 lakh.
  3. One simulation, every SIP amount. On each path the corpus is proportional to the SIP amount $A$ (every instalment is scaled by $A$). So if $F$ is the simulated corpus of ₹1 a month, $P(\text{goal})=P(AF\ge G)$, and the SIP that succeeds with probability $q$ is $G$ divided by the $(1-q)$-quantile of $F$.
  4. The menu.
chance of ₹40 lakhSIP (₹/month)invested (₹)
≈ 45% (the "12%" plan)12,98318,69,552
50%13,65119,65,744
75%17,98925,90,416
80%19,22427,68,256
90%22,81532,85,360
  1. A "conservative rate" shortcut. Planning at an assumed 8% a year gives ₹16,844 a month, which succeeds on about 70% of simulated paths. A lower assumed rate is a crude way of buying probability; the simulation tells you how much you bought.
  2. Other levers. Starting earlier, a yearly step-up (§1.2), or moving the corpus gradually into debt funds in the last few years (which narrows the fan near the goal date) all change the picture; each can be tested in the same simulation.

The mathematics, not a recommendation. Monte Carlo turns "how much should we invest?" into "how much for which chance?", which is the honest form of the question. But every number above inherits the model's assumptions: normal monthly returns, a constant and known $\mu$ and $\sigma$, no costs or taxes. Real markets have fatter tails, and nobody knows the true drift (see "How precisely can the drift be estimated?" below). A family that wants a high chance of success is really choosing among the 75%–90% rows, not the 45% one — and any such plan needs revisiting every year as actual returns arrive.

The limits of the model. A Monte Carlo answer is only as good as the distribution you feed it. Three limitations matter most. (i) Shape: real daily returns have fat tails and volatility that comes in clusters (§3.1); a normal model makes extreme days far too rare. (ii) Dependence: correlations between assets rise in crises (§3.2), which a model with fixed correlations misses. (iii) Parameters: the drift $\mu$ — the input that matters most for long horizons — is the one we know least, as the next paragraph shows.

How precisely can the drift be estimated? The standard-error rule of §3.1 says that the mean return estimated from $T$ years of data has standard error $\sigma/\sqrt T$, whatever the sampling frequency. Under GBM this is exact and easy to see: the natural estimate of the log drift is $\hat\nu=\tfrac1T\log\big(S(T)/S(0)\big)$, because the sum of the daily (or monthly) log returns telescopes to $\log(S(T)/S(0))$ however often you sample; by Theorem 3.13 it is $N(\nu T,\sigma^2T)$, so $$\hat\nu\sim N\!\left(\nu,\ \frac{\sigma^2}{T}\right).$$ Only the length of history helps. With $\sigma=18\%$ and 20 years of data, $\operatorname{SD}(\hat\nu)=18\%/\sqrt{20}\approx4.0\%$: a 95% interval for the drift is about ±7.9 percentage points. That uncertainty dwarfs the Monte Carlo error of a few thousand paths. (Volatility, by contrast, is estimated better with more frequent data.)

⚠️ Common mistakes and myths
  1. "A random walk means prices cannot rise over time." A random walk can have a positive drift; the model says changes are unpredictable, not that the average change is zero.
  2. "Seven up days in a row prove a trend." In a year of 252 fair coin-flip days, a streak of 7 or more up days happens about 63% of the time (Exercise I, part 5). Patterns appear in pure noise.
  3. "Expected return is what a typical investor gets." Under GBM the median grows at $\mu-\sigma^2/2$, not $\mu$; for very volatile assets the median can even shrink while the mean grows (Example 6).
  4. "In the long run equities cannot lose." The loss probability falls with the horizon but never reaches zero (3.4% at 10 years in Example 4, before fat tails), and the rupee range of outcomes widens.
  5. "The SIP calculator's figure is my answer." It is roughly the mean; in Example 8 about 56% of simulated outcomes fell below it.
  6. "The worst of 1,000 simulated paths is the worst case." It is the worst case of the model. Normal-return models understate crashes; stress tests (§7.4) are needed as well.
  7. "More simulations make the answer right." More paths remove simulation noise only (Theorem 3.15). They cannot fix wrong inputs: the drift itself is uncertain by several percentage points (the standard-error rule of §3.1).
  8. Forgetting the σ²/2 in code. Using $\mu$ as the log drift ($S_{k+1}=S_ke^{\mu\Delta t+\sigma\sqrt{\Delta t}z}$) makes the mean grow at $\mu+\sigma^2/2$ and overstates every outcome.
  9. Using the naive step $S(1+\mu\Delta t+\sigma\sqrt{\Delta t}z)$ with large steps. It is only approximately right and can produce negative prices; step exactly in logs.
  10. "Efficient markets mean prices are always right, so I can't lose." The EMH is about predictability, not correctness or safety; efficient prices still fall when bad news arrives.

Exercises

Unless stated otherwise, prices follow GBM with drift $\mu$ and volatility $\sigma$ per year (illustrative inputs), $\nu=\mu-\tfrac12\sigma^2$, and $\log$ is the natural logarithm. Use 252 trading days a year. Give probabilities to three decimal places or as percentages to one decimal place.

I
Random walks and the efficient-market idea.
  • 1 A toy share moves ₹2 up or ₹2 down each day with probability ½ each, independently. Find the mean and SD of its change over 400 days.
  • 2 A walk steps $+1$ with probability 0.55 and $-1$ with probability 0.45. Find the mean and SD of $S_{100}-S_0$.
  • 3 Daily log returns are independent with mean 0.05% and SD 1.2%. Find the mean and SD of the one-year log return and the probability of an up day (normal model).
  • 4 For part 3, after how many trading days does the expected drift equal one SD of the accumulated noise? Express it in years.
  • 5 In a fair random walk over 252 days, find the probability that 7 specified consecutive days are all up days, and the expected number of 7-day windows in the year that are all up. A short recursion (see 🛠️ Do it yourself) shows that the probability of at least one streak of 7 or more up days is about 63%. What does this say about the cousin's "seven up days" argument?
  • 6 Under the random-walk model, a share has risen five days running. What is the probability that it rises tomorrow if the probability of an up day is 0.52? What kind of evidence would contradict the weak-form EMH?
  • 7 Explain in two sentences why weak-form efficiency does not mean that a long-term index investor cannot earn a positive return.
1 Step mean 0, variance 4. Mean ₹0; SD $2\sqrt{400}=$ ₹40.
2 $m=0.55-0.45=0.1$, $s^2=E[X^2]-m^2=1-0.01=0.99$. Mean $100\times0.1=10$; SD $\sqrt{99}\approx9.95$.
3 Mean $252\times0.05\%=12.6\%$; SD $1.2\%\times\sqrt{252}\approx19.05\%$; $P(\text{up day})=\Phi(0.05/1.2)=\Phi(0.0417)\approx0.517$.
4 $n=(s/m)^2=(1.2/0.05)^2=576$ days $\approx2.3$ years.
5 Specified 7 days: $(1/2)^7=1/128\approx0.0078$. There are $252-7+1=246$ windows, so the expected number of all-up windows is $246/128\approx1.92$. A 7-day streak is therefore a normal feature of pure noise (about a 63% chance in a year); on its own it is no evidence of predictability.
6 Independence of steps means the past five days are irrelevant: the probability is still 0.52. Evidence against the weak form would be a rule based only on past prices that earns more than a fair reward for risk, after costs, consistently and out of sample (with proper statistics — §7.3).
7 The weak form says changes cannot be predicted from past prices; it allows a positive average change (drift) as the reward for bearing equity risk. An index investor collects that drift without trying to predict the next move.
II
Lognormal prices and GBM.
  • 1 A share at ₹500 has one-year log return $N(0.09,\,0.25^2)$. Find the median and mean price in a year.
  • 2 For part 1, find the probability that the price is below ₹450 in a year.
  • 3 For part 1, find the central 95% range of the price in a year.
  • 4 With $\mu=10\%$, $\sigma=20\%$, find $\nu$ and the median and mean value of ₹1,00,000 after 10 years.
  • 5 For part 4, find the probability of a loss after 1 year and after 10 years.
  • 6 For part 4, find the probability that ₹1 lakh has at least doubled after 10 years.
  • 7 With $\mu=10\%$, for which volatilities does the median value grow over time?
  • 8 Show that under GBM the ratio mean ÷ median after $t$ years is $e^{\sigma^2t/2}$, and evaluate it for $\sigma=20\%$, $t=10$.
1 Median $500e^{0.09}\approx$ ₹547.09; mean $500e^{0.09+0.03125}\approx$ ₹564.45.
2 $\Phi\!\left(\dfrac{\log0.9-0.09}{0.25}\right)=\Phi(-0.781)\approx21.7\%$.
3 $500e^{0.09\pm1.96\times0.25}$: from about ₹335.16 to about ₹893.02.
4 $\nu=0.10-0.02=0.08$. Median $1{,}00{,}000\,e^{0.8}\approx$ ₹2,22,554; mean $1{,}00{,}000\,e^{1}\approx$ ₹2,71,828.
5 $\nu/\sigma=0.4$: 1 year $\Phi(-0.4)\approx34.5\%$; 10 years $\Phi(-0.4\sqrt{10})=\Phi(-1.265)\approx10.3\%$.
6 $d=\dfrac{0.8-\log2}{0.2\sqrt{10}}=\dfrac{0.1069}{0.6325}\approx0.169$, so $P=\Phi(0.169)\approx56.7\%$.
7 The median $S_0e^{\nu t}$ grows iff $\nu\gt0$, i.e. $\sigma^2\lt2\mu=0.2$: $\sigma\lt\sqrt{0.2}\approx44.7\%$.
8 $\dfrac{S_0e^{\mu t}}{S_0e^{(\mu-\sigma^2/2)t}}=e^{\sigma^2t/2}$. For $\sigma=0.2$, $t=10$: $e^{0.2}\approx1.221$ — the mean is 22% above the median.
III
Monte Carlo methods.
  • 1 In 5,000 simulated paths, 1,150 end below a target. Give the estimate of the shortfall probability, its standard error and an approximate 95% confidence interval.
  • 2 A probability near 0.3 is to be estimated with standard error at most 0.5 percentage points. How many paths are needed?
  • 3 Write the exact one-month GBM update for $\mu=12\%$, $\sigma=18\%$, and apply it to a price of ₹100 with $z=-1.5$.
  • 4 Use Box–Muller with $u_1=0.3$, $u_2=0.8$ to produce two standard normal values.
  • 5 For one annual step with $\mu=12\%$, $\sigma=18\%$ and $z=-2$, compare the naive step $1+\mu+\sigma z$ with the exact factor $e^{\nu+\sigma z}$. For which $z$ would the naive annual step give a negative price?
  • 6 A planner simulates 200 twenty-year paths ($\mu=12\%$, $\sigma=18\%$) and reports "not one path lost money, so a loss is impossible". Give the model's loss probability, the probability of seeing no losing path among 200, and a rough 95% upper bound for the loss probability based on "0 out of 200" alone.
1 $\hat p=1150/5000=0.23$; $\text{SE}=\sqrt{0.23\times0.77/5000}\approx0.0060$; 95% CI $0.23\pm0.0117$, about 21.8% to 24.2%.
2 We need $\sqrt{0.3\times0.7/N}\le0.005$, i.e. $N\ge0.21/0.000025$ $=8{,}400$ paths.
3 $\nu=0.1038$: $S_{k+1}=S_k\exp\big(0.1038/12+0.18\sqrt{1/12}\,z\big)$. With $z=-1.5$: exponent $0.00865-0.07794=-0.06929$, so ₹100 becomes about ₹93.31.
4 $\sqrt{-2\log0.3}\approx1.5518$ and $2\pi u_2\approx5.0265$: $z_1=1.5518\cos(5.0265)\approx0.480$, $z_2=1.5518\sin(5.0265)\approx-1.476$.
5 Naive: $1+0.12-0.36=0.76$. Exact: $e^{0.1038-0.36}=e^{-0.2562}\approx0.774$. The naive factor is negative when $1+0.12+0.18z\lt0$, i.e. $z\lt-6.22$ — very rare, but with fat-tailed inputs or high volatility it happens; the exact log step is always positive.
6 Model: $\Phi(-0.1038\sqrt{20}/0.18)=\Phi(-2.579)\approx0.50\%$. Seeing no loss in 200 paths: $(1-0.00496)^{200}\approx0.37$ — quite likely. "Rule of three": with 0 events in $N$ trials, an approximate 95% upper bound is $3/N=1.5\%$. A small simulation cannot show that a rare event is impossible (and the normal model already understates crashes).
IV
Lump sums, SIP outcomes, confidence bands and limitations. Use $\mu=12\%$, $\sigma=18\%$ unless stated otherwise.
  • 1 ₹5,00,000 is invested for 15 years. Find the median, the mean and the central 80% range (10th to 90th percentile) of its value.
  • 2 For part 1, find the probability that the investment is worth less than ₹5 lakh after 15 years.
  • 3 Find the exact expected corpus of a SIP of ₹10,000 at the start of each month for 10 years, and the amount invested.
  • 4 A planner's simulation of a SIP reports: 10th percentile ₹24.7 lakh, median ₹44.4 lakh, 90th percentile ₹84.0 lakh, mean ₹50.7 lakh. Write three sentences explaining these numbers to a client, and say which one should not be presented as "what you will get".
  • 5 With $\sigma=20\%$ and 25 years of history, find the standard deviation of the estimated drift and the half-width of an approximate 95% interval for it.
  • 6 Give two reasons why a fan chart from a normal-return GBM simulation may understate the real risk of a SIP.
  • 7 Over 15 years, a ₹18 lakh lump sum has about a 1.3% chance of ending below ₹18 lakh, but the ₹10,000-a-month SIP of Example 8 (also ₹18 lakh in total) has about 2.2%. Explain the difference.
1 $\nu t=1.557$, $\sigma\sqrt t=0.18\sqrt{15}\approx0.6971$. Median $5e^{1.557}\approx$ ₹23.72 lakh; mean $5e^{1.8}\approx$ ₹30.25 lakh; 80% range $5e^{1.557\pm1.2816\times0.6971}$: about ₹9.71 lakh to ₹57.97 lakh.
2 $\Phi(-0.1038\sqrt{15}/0.18)=\Phi(-2.233)\approx1.3\%$.
3 $$E[FV]=10{,}000\sum_{k=1}^{120}e^{0.01k}$$ $$=10{,}000\,\frac{e^{0.01}(e^{1.2}-1)}{e^{0.01}-1}\approx\text{₹}\,23{,}31{,}737;$$ invested ₹12,00,000.
4 For example: "In our model, half of the simulated futures end above ₹44 lakh and half below. One in ten ends below about ₹25 lakh, and one in ten above about ₹84 lakh. The average, ₹51 lakh, is pulled up by a few very good futures, so most outcomes are below it." The mean (and any constant-rate calculator figure) should not be presented as "what you will get".
5 $\operatorname{SD}(\hat\nu)=0.20/\sqrt{25}=4\%$; half-width $1.96\times4\%\approx7.8$ percentage points.
6 Any two of: real returns have fat tails and volatility clustering, so big falls are more frequent than the normal model says; $\mu$ and $\sigma$ are uncertain and may change (regimes); costs, taxes and behaviour (stopping a SIP after a fall) are ignored; correlations and crashes across markets are not modelled.
7 In a SIP, later instalments are invested for only a few years, and short horizons have a much higher loss probability (28% for 1 year); also a fall near the end hits the whole accumulated corpus. The lump sum is exposed for the full 15 years, where the loss probability is low.
🎯 Practice corner NISM-styleCase studyInterview

E1. NISM-style According to the weak form of the efficient-market hypothesis, (A) share prices always equal their intrinsic value (B) analysis of past prices and volumes alone should not give a consistent excess return after costs (C) insider information cannot help an investor (D) share prices cannot fall sharply.

Reason The weak form concerns the information in past prices and volumes (Definition 3.11). (A) and (D) are not claims of any form of the EMH; (C) would be the strong form.
Answer (B).

E2. NISM-style If daily returns are independent with a standard deviation of 1%, then, using 252 trading days, the annualised volatility is closest to (A) 2.52% (B) 15.9% (C) 25.2% (D) 252%.

Idea Variances of independent returns add, so the SD scales with $\sqrt{\text{time}}$ (Theorem 3.11): $1\%\times\sqrt{252}\approx15.87\%$.
Distractors (C) and (D) scale the SD by 25.2 or 252 instead of $\sqrt{252}\approx15.87$; (A) divides instead.
Answer (B).

E3. NISM-style An index's annual log returns are independent and normal with mean 8% and standard deviation 20%. The probability that an investment in it shows a loss after 4 years is closest to (A) 5% (B) 21% (C) 34% (D) 50%.

Compute The 4-year log return is $N(0.32,\,0.4^2)$, so $P(\text{loss})=\Phi(-0.32/0.4)=\Phi(-0.8)\approx0.212$.
Distractors (C) is the one-year value $\Phi(-0.4)$; (D) ignores the drift.
Answer (B).

E4. Case study Rahul has ₹25,00,000 and wants ₹50,00,000 in 8 years. He models a diversified equity fund as GBM with $\mu=11\%$, $\sigma=16\%$ (illustrative). (a) Find the probability of reaching the goal. (b) Find the median and mean value at 8 years and the central 90% range. (c) Find the probability of a loss. (d) How many years would give an 80% chance of reaching ₹50 lakh?

(a) $\nu=0.11-0.0128=0.0972$. $$d=\frac{0.7776-\log2}{0.16\sqrt8}=\frac{0.0845}{0.4525}\approx0.187,$$ so $P\approx\Phi(0.187)\approx57.4\%$.
(b) Median $25e^{0.7776}\approx$ ₹54.41 lakh; mean $25e^{0.88}\approx$ ₹60.27 lakh; 90% range $25e^{0.7776\pm1.645\times0.4525}$: about ₹25.84 lakh to ₹114.5 lakh.
(c) $\Phi(-0.0972\sqrt8/0.16)=\Phi(-1.718)\approx4.3\%$.
(d) Solve $\dfrac{\nu t-\log2}{\sigma\sqrt t}=0.8416$: with $x=\sqrt t$, $0.0972x^2-0.1347x-0.6931=0$, so $x\approx3.452$ and $t\approx11.9$ years.
Comment Even though the median (₹54 lakh) exceeds the goal, the plan fails about 43% of the time in the model — and more often if returns have fat tails or the drift is lower than assumed.

E5. Case study A planning tool runs 2,000 simulations of a client's retirement SIP and reports "1,240 of 2,000 scenarios reach the goal". (a) Estimate the success probability with its standard error and a 95% interval. (b) How many simulations would make the 95% half-width 1 percentage point? (c) The tool's drift input was estimated from 15 years of data with $\sigma=16\%$. Compare the uncertainty in the drift with the simulation error, and say which matters more.

(a) $\hat p=0.62$; $\text{SE}=\sqrt{0.62\times0.38/2000}\approx0.0109$; 95% interval about 59.9% to 64.1%.
(b) We need $1.96\sqrt{0.62\times0.38/N}\le0.01$, i.e. $$N\ge\frac{1.96^2\times0.2356}{0.0001}\approx9{,}051,$$ so about 9,100 simulations.
(c) By the standard-error rule of §3.1, $\operatorname{SD}(\hat\nu)=0.16/\sqrt{15}\approx4.1$ percentage points: the true drift could plausibly be 8 points higher or lower than the input. Re-running the tool with the drift 4 points lower would change the success probability far more than ±2 points of simulation error.
Conclusion Model and parameter uncertainty matter far more than the number of simulations. A good report shows results for several drift and volatility inputs, not one probability to two decimals.

E6. Interview A game: each round your whole stake is multiplied by 1.5 (heads) or 0.6 (tails), with a fair coin. The expected multiplier per round is 1.05, so "on average you gain 5% a round". You play 100 rounds starting with ₹1. What are the expected final stake, the median final stake, and the probability of ending with more than ₹1? Would you play?

Mean Rounds are independent, so $E[\text{stake}]=1.05^{100}\approx$ ₹131.5.
Median With $k$ heads the stake is $1.5^k0.6^{100-k}$. The median number of heads is 50: $(1.5\times0.6)^{50}=0.9^{50}\approx$ ₹0.0052. Per round, the typical growth factor is the geometric mean $\sqrt{0.9}\approx0.949$ — a 5.1% loss per round.
Probability The stake exceeds ₹1 iff $k\log1.5+(100-k)\log0.6\gt0$, i.e. $k\gt100\log(1/0.6)/\log2.5\approx55.75$, so $k\ge56$: $P(\text{Bin}(100,\tfrac12)\ge56)\approx13.6\%$.
Interview answer This is the σ²/2 effect in its starkest form: the mean grows because of a few astronomically lucky sequences, while 86% of players lose and the typical player is wiped out. Whether to play depends on whether you can play many independent copies with small stakes (betting only a fraction of your wealth each round changes the answer — the Kelly idea of §7.2), not on the expected value alone.
🧑‍🏫 For the teacher

Suggested plan — 2 lectures × 60 min.

  1. Random walks and lognormal prices. Open with the cousin's "seven up days" claim: have the class toss coins seven at a time and count streaks, then reveal the 63% figure. Do Definition 3.10, Theorem 3.11 and Examples 1–2 (the √n law and "drift is invisible day to day"). Discuss the EMH briefly — what it claims and what it does not. Move to log prices, Definition 3.12 and Theorem 3.12 with the complete-the-square derivation, and Example 3, stressing the asymmetric range.
  2. GBM and Monte Carlo. Definitions 3.13–3.14 and Theorems 3.13–3.14; Examples 4–6. Project Interactive 1: start with σ = 18%, then raise σ to 50% and ask why the green median line falls away from the purple mean line. Then Theorem 3.15 and the 🧭 simulation recipe, Example 7 (touch vs end) and Example 8. Finish with Interactive 2 and the 📈 Market case: change the SIP until the goal probability is 75%, and then lower μ by three points to show parameter risk.

Class activity. Each student picks a goal and a SIP from §1.2. In pairs, they use Interactive 2 to find the SIP for a 50% and an 80% chance under two sets of inputs (μ = 12%, σ = 18% and μ = 9%, σ = 22%), and present the four numbers as a "probability menu" rather than one figure.

Discussion questions. (1) If prices are a random walk, why do technical analysts see patterns? (2) Is a 3% chance of loss over 10 years "safe"? For whom? (3) Why might a planning tool prefer to show the mean, and why is that misleading?

Misconceptions to watch for. Random walk = no drift; mean = typical outcome; long horizon = no risk; Monte Carlo output = truth; more simulations = more accuracy about the real world; forgetting σ²/2 when coding.

Suggested homework. I(2, 3, 5, 6), II(1–6), III(1, 2, 3, 6), IV(1, 3, 4, 5, 7); Practice corner E3, E4, E6.

🛠️ Do it yourselfSpreadsheet or Python mini-project · about 60–90 minutes

Project. Build your own Monte Carlo simulator for a SIP, check it against the exact formulas of this section, and turn it into a goal-probability table for a goal of your choice.

Why it matters. Once you have built a simulator yourself, you can test any plan — a different SIP, a step-up, a lump sum, a shorter horizon — and you will read the output of any planning tool with the right mixture of interest and scepticism.

What to do (spreadsheet).

  1. Inputs: SIP in B1, μ in B2, σ in B3, years in B4, goal in B5. Monthly log drift B6 =(B2-B3^2/2)/12, monthly SD B7 =B3/SQRT(12).
  2. One path per row: in row 10, column A holds a path number; columns B onward hold months. B10 =$B$1*EXP($B$6+$B$7*NORM.S.INV(RAND())); C10 =(B10+$B$1)*EXP($B$6+$B$7*NORM.S.INV(RAND())), copied right for all months. The last column is the corpus.
  3. Copy row 10 down 1,000 rows. Then: median =MEDIAN(corpus column), 10th and 90th percentiles =PERCENTILE.INC(range,0.1) and 0.9, goal probability =COUNTIF(range,">="&B5)/1000, and its standard error =SQRT(p*(1-p)/1000).
  4. Check the code: set the SIP to a single lump sum in month 1 only, and compare the simulated loss probability and median with Theorem 3.14. Then compare the simulated mean with the exact $E[FV]=A\sum_k e^{\mu k/12}$ of Example 8.
  5. Press F9 (recalculate) several times and watch the estimates move by about the standard error — that is Monte Carlo noise.

Python version: rng=np.random.default_rng(1); v=np.zeros(N); for k in range(n): v=(v+A)*np.exp(nu/12+sig*np.sqrt(1/12)*rng.standard_normal(N)), then np.percentile(v,[10,50,90]) and (v>=G).mean(). For the streak fact of Exercise I(5), track the current run length: a state vector p[j] = probability that the current run of up days has length j (j = 0…6) and that no 7-run has occurred yet; each day, p'[0]=0.5*sum(p), p'[j+1]=0.5*p[j], and the mass leaving from p[6] is the hit probability.

Hand in: a fan chart (10th, 25th, 50th, 75th, 90th percentiles by year), a "probability menu" table of SIPs for 50%, 75% and 90% success, and a short note on how the menu changes if μ is lowered by 3 percentage points.

Going further. Replace the normal draws with a fat-tailed distribution with the same mean and SD (for example a scaled Student-t with 4 degrees of freedom), or resample monthly returns from a fund's own NAV history downloaded from its AMC or AMFI (a bootstrap). How do the 10th percentile and the goal probability change?

Instructor note: with ₹10,000 a month for 15 years, μ = 12%, σ = 18%, a large simulation gives a median of about ₹44.4 lakh, 10th and 90th percentiles of about ₹24.7 and ₹84.0 lakh, and a mean close to the exact ₹50.75 lakh; with 1,000 paths expect the percentiles to wobble by a lakh or two between recalculations. The most common bugs are using μ instead of μ − σ²/2 as the log drift and mixing annual and monthly parameters.

🧪 Self-check quiz

0/5 correct

Q1. In a random walk with independent steps of standard deviation $s$, the standard deviation of the total change after $n$ steps is

Q2. If $\log(S_T/S_0)\sim N(a,b^2)$, the median of $S_T$ is

Q3. Under GBM with $\mu=12\%$ and $\sigma=18\%$, the median value grows at a continuously compounded rate of about

Q4. A Monte Carlo estimate from 1,000 paths has a standard error of 1.5 percentage points. To bring it down to 0.5 points you need about

Q5. A SIP calculator that grows every instalment at the expected rate of return gives a figure that, under GBM, is