ARBIMETRA · Financial intelligence, made practical.

Financial Modelling Scenario and Risk Modelling

Monte Carlo Simulation in Finance

How Monte Carlo simulation works in finance: choosing distributions, running thousands of trials, reading percentiles and probabilities, and the method’s limitations.

Advanced4 min readUpdated Published

Share

A single forecast says what happens if everything goes as expected. Reality rarely does. Monte Carlo simulation replaces fixed assumptions with probability distributions, then recalculates the model thousands of times with randomly drawn values. The result is not one answer but a distribution of possible outcomes — showing the range, the most likely values and the chance of reaching (or missing) a target. The technique is used to value complex derivatives, plan retirement savings, appraise projects and measure portfolio risk.

ARBIMETRA’s own investment scenario simulation uses this method, with its methodology documented in the Research section.

What you will learn

  • Explain the steps of a Monte Carlo simulation
  • Choose and justify input distributions
  • Interpret percentiles, probabilities and the spread of outcomes
  • Recognise the limitations and common misuses of simulation

How it works

Steps of a Monte Carlo simulation

  1. Build the modelOutput as a function of inputs
  2. Define distributionsFor each uncertain input
  3. Draw random valuesOne set per trial
  4. Calculate the outcomeRun the model once
  5. RepeatThousands of trials
  6. AnalysePercentiles, probabilities, risk
  1. Build a deterministic model — for example, the value of an investment after ten years, or a project’s NPV.
  2. Identify uncertain inputs and describe each with a probability distribution.
  3. Generate random values for every uncertain input.
  4. Calculate the output for that set of values — one "trial" or "path".
  5. Repeat thousands of times.
  6. Analyse the results — average, median, percentiles, probability of loss or of meeting a goal.

Choosing distributions

DistributionTypical use
NormalVariables that can move symmetrically up or down, such as forecast errors
LognormalAsset prices and growth of wealth — values cannot fall below zero and gains compound
UniformAn input known only to lie within a range
TriangularExpert estimates of minimum, most likely and maximum
DiscreteEvents that happen or do not, such as winning a contract

Distributions should be grounded in data where possible (for example, historical volatility of returns) and in expert judgement where not. Correlations between inputs matter: if sales volume and price tend to fall together, simulating them independently understates the risk.

Worked example: a ten-year investment

Worked example: simulating investment growth

An investor places 100,000 in a portfolio for 10 years. Annual returns are assumed to be lognormally distributed with an expected return of 8% and a standard deviation of 15%, independent from year to year. The simulation runs 10,000 trials with a fixed random seed so the results can be reproduced (illustrative assumptions).

MeasurePortfolio value after 10 years
5th percentile95,737
10th percentile112,500
25th percentile146,592
Median (50th percentile)198,407
75th percentile265,567
90th percentile345,030
95th percentile402,631
Mean217,032
  • A deterministic calculation at a steady 8% gives 100,000 × 1.08^10 = 215,892 — close to the simulated mean, but the median is lower, because a few very good paths pull the average up.
  • About 6% of trials end below the original 100,000.
  • About 49% of trials reach 200,000 or more.

These are model outputs based on stated assumptions — not forecasts of any real portfolio.

Distribution of simulated outcomes (10,000 trials)

Bar chart of how many of 10,000 simulated outcomes fall into each value range after 10 years: below 100,000: 612; 100,000 to 150,000: 2,069; 150,000 to 200,000: 2,379; 200,000 to 250,000: 1,992; 250,000 to 300,000: 1,265; 300,000 to 400,000: 1,163; 400,000 and above: 520.05001,0001,5002,0002,500<100k100–150k150–200k200–250k250–300k300–400k400k+
Illustrative figures for explanation only — not real market or company data.
Show the data
Number of trials
<100k612
100–150k2,069
150–200k2,379
200–250k1,992
250–300k1,265
300–400k1,163
400k+520

Reading the results

  • Percentiles answer "what is a reasonably bad, typical and reasonably good outcome?" The range between the 10th and 90th percentiles is a practical summary of uncertainty.
  • Probabilities answer goal-based questions: "what is the chance of having at least 200,000?" This is often more useful than an average.
  • Mean versus median: for compounding returns the distribution is skewed, so the median is a better guide to a "typical" result.
  • Convergence: run enough trials that the results barely change when the number of trials increases.

Applications in finance

  • Personal planning: the chance that savings last through retirement — see retirement planning.
  • Project appraisal: the distribution of NPV and the probability of a negative NPV, extending scenario analysis.
  • Risk measurement: simulating portfolio losses to estimate value at risk and expected shortfall.
  • Derivative pricing: valuing options whose payoffs depend on the path of prices, a technique developed extensively in financial engineering.
  • Credit portfolios and insurance: modelling the combined losses from many uncertain events.

Limitations

  • Garbage in, garbage out: results are only as good as the distributions and correlations assumed.
  • Fat tails and regime changes: real returns have more extreme events than a normal or lognormal model suggests, and relationships can change in crises — complement simulation with stress testing.
  • False precision: results with many decimal places can look more certain than they are.
  • Model risk: complex simulations are harder to check; document the model and test it — see model best practices.

Common mistakes

  • Using a normal distribution for quantities that cannot be negative, such as prices.
  • Ignoring correlations between inputs.
  • Running too few trials, so results change each time.
  • Presenting the mean as the expected experience of a typical investor when the distribution is skewed.
  • Treating simulated percentiles as guarantees.

Key terms

  • Monte Carlo simulation — repeated random sampling of inputs to estimate the distribution of outcomes.
  • Trial (path) — one run of the model with one set of random inputs.
  • Percentile — the value below which a given percentage of outcomes fall.
  • Lognormal distribution — a distribution of values whose logarithms are normally distributed.
  • Random seed — a starting value that makes a pseudo-random sequence reproducible.

Frequently asked questions

How many trials are enough?

Enough that key results stabilise — often several thousand to tens of thousands, depending on how precise the tail estimates need to be.

Is Monte Carlo better than scenario analysis?

It complements it. Scenarios describe coherent stories; simulation describes the full spread of outcomes given input distributions.

Can I run a Monte Carlo simulation in a spreadsheet?

Yes, for simple models, using random number functions and data tables. For large or repeated simulations, programming languages are more efficient and easier to test.

Key takeaways

  • Monte Carlo replaces fixed inputs with distributions and runs the model thousands of times.
  • Results are distributions: use percentiles and probabilities, not just averages.
  • Distributions and correlations must be justified; results inherit their weaknesses.
  • Combine simulation with scenario analysis and stress testing.