CFA Level 1 · Module 02 Quantitative Methods · Chapter 9
Some questions in finance have clean formula answers. Many do not. What is the chance a retirement portfolio runs out of money over thirty years of uncertain returns and withdrawals? What is the loss a complex portfolio might suffer on a bad day? For questions like these, where the moving parts interact in ways no tidy equation captures, the practical answer is to let a computer play out the uncertainty thousands of times and look at the range of outcomes. That is simulation, and this reading covers its three main forms.
The three methods differ mainly in where they get their randomness. Historical simulation replays actual past returns. Bootstrap resampling draws repeatedly, with replacement, from a set of observed returns. Monte Carlo simulation generates returns from an assumed probability distribution. All three share the same spirit: build many possible futures, then read the distribution of results rather than a single point forecast.
This is a largely conceptual reading, so the exam tests understanding over calculation. You need to describe each method, know its strengths and weaknesses, match it to a use, and recognize the limitation they all share: a simulation is only as good as the assumptions and data fed into it.
A closed-form formula gives an exact answer in one step, like the present value of a bond. Simulation is what you reach for when no such formula exists or when writing one would be hopelessly complex. Problems with many interacting uncertainties, path-dependent outcomes, or long horizons rarely yield to a single equation, but they yield readily to brute force: specify how the uncertainty behaves, then let a computer generate a large number of possible outcomes and summarize them.
Simulation is old in idea but modern in practice, because it depends on cheap computing. Running thousands or millions of trials by hand is impossible, so the methods only became everyday tools once computers could generate and process large numbers of random outcomes quickly. That is why simulation now appears throughout finance, from a planner’s laptop projecting a retirement to a bank’s risk system stress-testing a trading book overnight. The conceptual recipe has not changed, but the ability to execute it at scale has.
The output of a simulation is not a single number but a whole distribution of outcomes, and that is exactly its value. Instead of a point estimate that hides the risk, a simulation shows the full range, the median case, the good cases, and crucially the bad tail. An analyst can then read off any statistic of interest, the average outcome, the chance of falling below a threshold, the size of a rare loss, from the same set of simulated results.
The general recipe is the same across all three methods. Decide what quantity you want to study, such as the value of a portfolio at a future date. Specify the source of randomness for the returns that drive it. Generate a large number of trials, computing the outcome in each. Finally, collect the outcomes into a distribution and read the answer from it. The methods differ only in the second step, how the random returns are produced.
The point of simulation is to replace a single forecast with a whole distribution of outcomes. A formula that returns one number hides the risk; a simulation that returns thousands of outcomes shows the median, the good cases, and above all the bad tail, which is usually the part a decision actually turns on.
Historical simulation uses actual past returns as the set of possible future outcomes. To estimate the risk of a portfolio, you take a window of its historical returns and treat each historical day, or month, as an equally likely scenario for the future. Reapplying those historical returns to the current portfolio produces a distribution of possible outcomes drawn entirely from what has really happened, with no assumption about the shape of the distribution required.
Its great strength is realism. Because it uses genuine market history, it automatically captures the fat tails, the skewness, and the way different assets actually moved together during real events, features that a neat theoretical distribution would smooth away. Nothing has to be assumed about how returns are distributed, because the historical record supplies the distribution directly. This makes historical simulation a natural tool for estimating value at risk, the loss a portfolio might suffer at a given confidence level.
Its weakness is the mirror image of its strength: it assumes the future will resemble the past. If the historical window did not contain a particular kind of shock, the simulation cannot produce one, so a portfolio can look safe simply because the sample period was calm. The results are also sensitive to the length of the window chosen, a short recent window may miss a crisis, while a long one may include regimes no longer relevant. History is a guide, not a guarantee, and historical simulation inherits that limitation completely.
Setup. An analyst has 20 historical daily returns for a portfolio and wants a one-day value at risk at the 95% confidence level using historical simulation. How is it found, and if the worst return in the window is negative 3.2%, what does the estimate look like?
Answer: a one-day 95% value at risk of roughly 3.2%. Notice that no distribution was assumed; the estimate came straight from the ranked history. The obvious caution is that a 20-day window is tiny, and a longer, more varied history would give a more reliable tail.
Historical simulation is defined by using the actual past returns as the scenario set, with no distribution assumed. If a question describes estimating value at risk by ranking a portfolio’s real historical returns and reading off a low percentile, it is describing historical simulation, not Monte Carlo. The absence of any assumed distribution is the giveaway.
The choice of window is itself a judgment with consequences. A window that stops just before a crisis will make a portfolio look far safer than one that includes it, so the estimate can swing sharply with the period chosen. Some practitioners weight recent observations more heavily than distant ones, on the view that the near past is more relevant to the near future, but this reintroduces a modeling choice into a method whose appeal was that it made none. The honest reading is that historical simulation is only as representative as the slice of history behind it.
Bootstrap resampling is a clever extension of the historical idea. Instead of using each historical return once, it draws returns repeatedly, with replacement, from the observed set. Because each draw is put back before the next, the same historical return can appear several times in one simulated sample, and the method can generate as many simulated observations as desired from a limited history. Drawing with replacement is the defining feature, and it is what separates the bootstrap from plain historical simulation.
The payoff is that the bootstrap can build a rich distribution of a statistic, such as a portfolio’s average return or its value at risk, even from a modest sample. By resampling thousands of times and computing the statistic each time, an analyst obtains a distribution for that statistic and can attach a confidence interval to it without assuming the data follow any particular theoretical shape. This makes the bootstrap especially useful when the data are not normal and no formula for the statistic’s distribution is available.
Like historical simulation, the bootstrap is anchored to the observed data, so it still assumes that the sample is representative of the future. It cannot invent outcomes more extreme than the range present in the data, since every draw comes from that set. But by reshuffling and reusing the observations it extracts more information from a small sample than a single pass through the history could, which is why it is a workhorse of modern statistics.
The bootstrap is used wherever a statistic’s uncertainty is needed but no clean formula exists for it. It can put a confidence interval around a Sharpe ratio, a maximum drawdown, or the correlation between two strategies, quantities whose sampling distributions are awkward to derive by hand. By resampling the data and recomputing the statistic thousands of times, the bootstrap lets its distribution reveal itself empirically. This is why it has become a standard tool in performance analysis, where the questions are often about odd statistics computed from limited track records.
Setup. An analyst has only 12 monthly returns and wants a distribution for the average monthly return. Explain how a bootstrap produces one.
Answer: repeated resampling with replacement turns 12 observations into a full distribution of the average, from which a confidence interval is read directly. The bootstrap squeezes a rich answer from a thin sample, at the cost of never producing an outcome outside the original 12.
Monte Carlo simulation takes a different route to the randomness. Rather than reuse historical data, it generates returns from an assumed probability distribution, typically a normal distribution for returns or a lognormal for prices, with parameters the analyst specifies. Because the returns come from a model rather than a fixed history, Monte Carlo can produce outcomes never seen in the past, including extreme scenarios, and it can be run for as many trials and as long a horizon as desired.
The process has clear steps. Choose a model for how the asset behaves, such as prices following a lognormal path. Specify the parameters, the expected return, the volatility, and any correlations. Draw random values from the chosen distribution, convert them into simulated returns, and roll the asset forward over the horizon. Repeat for thousands of paths, then collect the ending values into a distribution. The flexibility to model almost any process is what makes Monte Carlo the most powerful of the three methods.
That power comes with a dependence on the model. Monte Carlo results are only as good as the assumptions behind them: choose the wrong distribution or the wrong parameters and the simulated futures, however numerous, will be systematically wrong. If returns are assumed normal but reality has fatter tails, the simulation will understate the odds of a crash no matter how many paths are run. The method does not discover the true distribution; it faithfully explores the one it is told to use.
where S0 is the starting price, S1 is the simulated next price, μ is the expected continuously compounded return, σ is the volatility, and z is a random draw from the standard normal distribution. Each trial uses a fresh z, producing a new simulated price.
Setup. A price of 100 is modeled with a continuously compounded annual return that is normal with a mean of 8% and a standard deviation of 20%. In one simulated trial the drawn standard normal value is positive 0.5. What is the simulated one-year price?
Answer: one simulated path ends at about 119.72. A full Monte Carlo run repeats this thousands of times with fresh random draws, producing a whole distribution of year-end prices from which any probability, such as the chance of a loss, can be read.
Monte Carlo does not discover reality; it explores the model it is given. Feed it a normal distribution and it will faithfully generate normal outcomes, thin tails and all, even if real returns are fatter-tailed. The method’s power to model anything is inseparable from its dependence on the analyst choosing the right something.
Monte Carlo extends naturally to many assets at once, which is where much of its real value lies. To simulate a portfolio, the draws for the different assets must respect their correlations, so the model uses the covariance structure from the earlier statistics reading to generate returns that move together in a realistic way. A single uncorrelated draw per asset would understate the risk of the portfolio falling all at once, so building the correlations into the draws is essential. This is why the statistics of covariance and correlation feed directly into a credible multi-asset simulation.
All three methods need a source of randomness, and it starts with a random number generator producing values that are uniformly distributed between zero and one, each equally likely. This is where the uniform distribution from the earlier reading finally earns its keep: it is the raw randomness from which every other distribution is built. On its own a uniform draw is not a return, but it is the starting material.
To turn a uniform draw into a draw from another distribution, such as the normal, the standard trick is the inverse transform. The uniform value is treated as a probability and fed into the inverse of the target distribution’s cumulative function, which returns the value at that percentile. A uniform draw of 0.5 maps to the median of the target distribution, a draw of 0.95 to its ninety-fifth percentile, and so on. Repeating this converts a stream of uniform values into a stream of normal, or lognormal, returns ready to drive the simulation.
where u is a uniform random draw between zero and one, F−1 is the inverse of the target distribution’s cumulative function, and x is the resulting draw from that distribution. Feeding a stream of uniform values through this inverse transform produces a stream of normal, lognormal, or other draws.
The quality of the random number generator matters more than it might seem. A poor generator whose values repeat in patterns can introduce hidden structure that corrupts the results, which is why simulation software relies on carefully designed generators. For the exam the key point is simpler: simulation begins with uniform random numbers, which are then transformed into draws from whatever distribution the model requires.
The three methods form a spectrum from fully data-driven to fully model-driven. Historical simulation uses the actual sequence of past returns. The bootstrap reuses the same data but reshuffles it with replacement to build richer distributions. Monte Carlo cuts the tie to history entirely and generates returns from an assumed distribution. Moving along this spectrum trades realism for flexibility: the historical methods stay faithful to what happened, while Monte Carlo can explore what has never happened but might.
| Method | Source of randomness | Main strength | Main weakness |
|---|---|---|---|
| Historical simulation | Actual past returns, used once | Realistic; no distribution assumed | Limited to what history contained |
| Bootstrap resampling | Observed returns, drawn with replacement | Rich distributions from small samples | Cannot exceed the observed range |
| Monte Carlo | An assumed probability distribution | Flexible; can model any process | Only as good as the assumed model |
There is no single best method; the right choice depends on the question and the data. When a long, representative history exists and realism is paramount, historical simulation is natural. When the sample is small and a statistic’s distribution is needed, the bootstrap shines. When the future may differ from the past, or when a specific model of the asset’s behavior is required, Monte Carlo is the tool. In practice, analysts often run more than one and compare, treating agreement between methods as reassurance and disagreement as a signal to dig deeper.
In practice the boundaries blur, and sophisticated systems blend the methods. A filtered historical simulation, for instance, rescales past returns to reflect current volatility before replaying them, borrowing a modeling idea to keep a data-driven method relevant. At this level you are not expected to know the hybrids, only the three pure methods and their trade-offs, but it is worth knowing that real risk systems mix them to get the realism of history and the flexibility of a model at once.
Mixing up which method assumes a distribution. The bootstrap and historical simulation both stay inside the observed data and assume nothing about its shape; only Monte Carlo draws from an assumed distribution. Attributing an assumed normal or lognormal model to the bootstrap, or claiming Monte Carlo is limited to observed outcomes, reverses the defining features of the two.
It also helps to distinguish simulation from simple scenario analysis. Scenario analysis examines a handful of hand-picked cases, a mild recession, a rate spike, and reports the outcome of each. Simulation instead generates thousands of outcomes across the full range of possibilities and reports their distribution. Scenario analysis answers “what happens in these particular situations”, while simulation answers “what is the whole distribution of what could happen”, including the probability of each region. The two are complementary, but only simulation yields a probability for an outcome.
Simulation appears wherever uncertainty compounds over time or across many moving parts. In risk management it estimates value at risk and the size of losses in stressed conditions, especially for portfolios too complex for a formula. In retirement and financial planning it projects the range of outcomes for a portfolio subject to uncertain returns and withdrawals over decades, answering the crucial question of how likely a plan is to succeed. In derivative pricing it values options and other instruments whose payoffs depend on the whole path of an underlying, not just its ending point.
What unites these uses is that each involves a distribution of outcomes rather than a single deterministic answer. A retirement plan does not have one future; it has thousands, and simulation is how you see them all and count the fraction that end badly. This is why the output of a simulation is often expressed as a probability, the chance of running out of money, the chance of a loss beyond a threshold, rather than a single forecast. That probabilistic framing is exactly what decision-makers need when the stakes are uncertain.
The shift to probabilistic thinking is itself part of simulation’s value. A client told that a plan “should work” learns little, but a client told that a plan succeeds in 85 of 100 simulated futures, and fails in 15, can weigh that against their own comfort and adjust. Simulation forces the uncertainty into the open and attaches numbers to it, which supports better decisions than a single confident-sounding forecast ever could. That honesty about what is unknown, rather than any single output, is often the deepest reason to run a simulation at all.
Setup. A planner runs 10,000 Monte Carlo paths of a retirement portfolio over 30 years and finds that in 1,500 of them the portfolio is exhausted before the end. How is the result read, and what does it depend on?
Answer: an estimated 15% chance of depletion, expressed as a probability precisely because the future is a distribution of paths, not one. The number is decision-useful, but only as trustworthy as the return and spending assumptions feeding the simulation.
| Application | Question it answers |
|---|---|
| Risk management | How large a loss might a complex portfolio suffer at a given confidence level? |
| Retirement planning | How likely is a plan to succeed over decades of uncertain returns and withdrawals? |
| Derivative pricing | What is a path-dependent option worth across many possible price paths? |
Every simulation shares one overriding limitation: it can only reflect the assumptions and data put into it. A Monte Carlo run with the wrong distribution, a historical simulation from an unrepresentative window, or a bootstrap from a biased sample will all produce confident-looking distributions that are quietly wrong. The precision of thousands of trials can lull a user into trusting a result that rests on a flawed input, which is the oldest warning in computing: garbage in, garbage out.
A second caution is that more trials reduce sampling noise but never fix a bad model. Running a million Monte Carlo paths instead of a thousand makes the simulated distribution smoother and more stable, but if the underlying assumptions are wrong the smoother answer is just a more precisely stated error. The number of trials controls how well the simulation approximates its own model, not how well that model matches reality. Judgment about the inputs, not raw computing power, is what makes a simulation trustworthy.
Because a simulation can be wrong in ways that are invisible on the screen, its results should be checked against reality wherever possible. A risk model can be backtested by asking how often actual losses exceeded the value at risk it predicted; if breaches happen far more often than the confidence level allows, the model’s assumptions are too optimistic. Comparing the output of different methods is another safeguard, since agreement between a historical simulation and a Monte Carlo run is reassuring, while a large gap flags an assumption worth questioning. A simulation that is never validated is an opinion dressed up as a calculation.
Simulation replaces a single forecast with a distribution of outcomes, which is its whole advantage. But the distribution is only as honest as the data or model behind it. Historical and bootstrap methods are trapped inside the observed range; Monte Carlo is trapped inside its assumed distribution. Knowing which trap applies is what separates useful simulation from false confidence.
Match the method to its defining feature. Historical simulation replays actual returns; the bootstrap draws from observed returns with replacement; Monte Carlo draws from an assumed distribution. A question that stresses drawing repeatedly with replacement is pointing at the bootstrap, while one that specifies a distribution and parameters is pointing at Monte Carlo.
Believing that a very large number of trials makes a simulation accurate. More trials only reduce the noise in approximating the assumed model; they do nothing to correct a wrong distribution or unrepresentative data. A tightly estimated wrong answer is still wrong.
Seen together, the three methods answer the same question in three voices. Historical simulation says the future will look like this stretch of the past. The bootstrap says the future will be built from these observed pieces, reshuffled. Monte Carlo says the future will follow this model I have specified. Each voice is useful, and each can mislead if trusted blindly, which is why the mature practitioner reaches for whichever fits the question, checks it against the others, and never forgets that all three are only as good as what they are fed.
Which simulation method draws returns from an assumed probability distribution rather than from historical data?
Monte Carlo simulation. It generates returns from a specified distribution, such as normal or lognormal, which lets it produce outcomes never observed in history but makes it dependent on the chosen model being correct.
What single feature distinguishes bootstrap resampling from plain historical simulation?
Drawing with replacement. The bootstrap samples repeatedly from the observed returns, putting each back so it can be drawn again, which lets a small sample generate a rich distribution. Plain historical simulation uses each observation once.
A price of 50 follows a continuously compounded return that is normal with a mean of 6% and a standard deviation of 10%. If a trial draws a standard normal value of negative 1.0, what is the simulated price?
Return = 0.06 + 0.10 × (−1.0) = −0.04. Price = 50 × e−0.04 = 50 × 0.9608 = about 48.04.
Why is a uniform random number the starting point for generating returns from any distribution?
Because a uniform draw between zero and one can be treated as a probability and fed into the inverse cumulative function of any target distribution, returning the value at that percentile. Uniform randomness is the raw material from which normal, lognormal, or any other draws are built.
Because many realistic problems, long horizons, many interacting risks, path-dependent payoffs, have no practical closed-form solution. Simulation sidesteps the need for a formula by generating many possible outcomes and reading the answer from their distribution.
Historical simulation uses each past return once, in effect replaying history. The bootstrap draws from the same returns with replacement, so observations can repeat, which lets a small sample generate a much larger set of simulated outcomes and a distribution for a statistic.
It generates returns from an assumed distribution rather than from fixed data, so it can produce outcomes never seen in the past and can model almost any process over any horizon. The trade-off is that its results depend entirely on the distribution and parameters chosen being correct.
Because a uniform value between zero and one can be treated as a percentile and converted, through the inverse cumulative function, into a draw from any target distribution. Uniform randomness is the common raw material behind normal, lognormal, and other draws.
Only up to a point. More trials reduce the random noise in approximating the assumed model, giving a smoother, more stable distribution. They do nothing to correct a wrong model, so a simulation built on flawed assumptions is only more precisely wrong.
Estimating value at risk and stressed losses in risk management, projecting the range of outcomes for retirement and financial plans over decades, and pricing derivatives whose payoffs depend on the whole path of an underlying. Each involves a distribution of outcomes rather than a single answer.
They can only reflect their inputs. Unrepresentative data or a wrong model produces a confident-looking but incorrect distribution. The precision of many trials can mask a flawed assumption, so the quality of the inputs, not the number of trials, determines whether the result can be trusted.
Through description and comparison rather than heavy calculation: identifying each method by its defining feature, matching a method to a use, and recognizing the shared limitation. Expect to distinguish the bootstrap by its drawing with replacement and Monte Carlo by its assumed distribution.
Loading comments...
Add your Thoughts: