Monte Carlo VaR

Historical VaR has a fatal flaw: it can only ever show you tomorrows that already happened. Your one year of data is your one year of data — 252 days, take it or leave it, and the crash you’re really afraid of probably isn’t in there. Parametric VaR fixes the data shortage by assuming a tidy normal bell curve, but then it lies about anything non-linear (hello, options). Monte Carlo VaR is the method that says: fine — if reality won’t give me enough tomorrows, I’ll manufacture them.

The trick is gloriously simple in spirit. Pick a statistical model for how your risk factors move. Use a random-number generator to roll thousands of synthetic tomorrows out of that model. Reprice your entire portfolio in each one. Now you have ten thousand fake P&Ls — sort them, walk into the loss tail, and read off the percentile. It’s historical VaR, except the history is one you invented, and you can invent as much of it as your laptop can stand.

Simulate a thousand tomorrows

Analogy. A pilot doesn’t wait for ten real engine fires to learn how the plane handles one — she runs the flight simulator a thousand times, each with slightly different gremlins. Monte Carlo VaR is the flight simulator for your portfolio: you can’t afford to live through ten thousand market days, so you generate them in software and watch how your book reacts.

Definition. Monte Carlo VaR estimates Value at Risk by (1) choosing a statistical model for the risk factors that drive your portfolio, (2) drawing many random scenarios from that model, (3) revaluing the full portfolio in each scenario to get a profit-or-loss, and (4) taking the appropriate percentile of that simulated loss distribution. The name comes from the casino — it’s the dice-rolling, random-sampling family of methods, applied to risk.

A risk factor is any market variable your portfolio’s value depends on: a stock price, an FX rate, an interest rate, an implied volatility. The model is your assumption about how those factors move — most commonly that one-day returns are normal (or that prices are lognormal, i.e. their logs are normal, which keeps prices positive) with a chosen mean $\mu$, volatility $\sigma$, and — when you have more than one factor — a correlation matrix tying them together so they move realistically in lockstep rather than independently.

That correlation piece is the soul of the method for a real book. If your portfolio holds tech and oil, and your simulator lets them wander independently, you’ll badly misjudge the days they crater together. The correlation matrix is what makes your synthetic tomorrows hang together like real market days.

The recipe, step by step

Here is the whole method as a five-step cookbook. Memorise the shape — every Monte Carlo VaR engine on earth is some elaboration of these five lines.

1. Specify the model and its parameters. Decide the distribution (e.g. lognormal prices), and pin down each factor’s $\mu$ and $\sigma$ plus the covariance/correlation matrix linking them. This is the one creative, dangerous step — everything downstream inherits whatever you assume here.
2. Generate $M$ scenarios. Draw $M$ random scenarios — say $M = 10{,}000$ — from that model. Each scenario is one synthetic “tomorrow”: a complete set of new values for every risk factor, drawn jointly so the correlations hold.
3. Fully reprice the portfolio under each scenario. For every one of the $M$ scenarios, recompute the portfolio’s value with the new factor values, and subtract today’s value to get that scenario’s P&L. “Fully” is the operative word — you re-run each instrument’s actual pricing formula, not a linear approximation.
4. Sort the $M$ P&Ls from worst to best. You now hold an empirical distribution of simulated outcomes — a histogram of ten thousand fake tomorrows.
5. Read the percentile. The VaR at confidence $c$ is the loss at the $(1-c)$ percentile of that sorted list. For $c = 99\%$ and $M = 10{,}000$, that’s the 100th-worst simulated loss (1% of 10,000) — every scenario below it is in the tail.

The chart below is exactly the object you build in steps 4–5: the histogram of simulated outcomes, with the VaR line sitting at the chosen percentile and the loss tail beyond it shaded. Drag the confidence slider and watch the line march deeper into the manufactured tail.

The killer feature: nonlinear portfolios

This is the reason Monte Carlo VaR exists, and the one argument that wins the meeting. Parametric (delta-normal) VaR approximates every position as a straight line: it takes each instrument’s delta (its first-order sensitivity to the underlying) and assumes the P&L is just delta × move, with moves drawn from a normal bell. That’s fine for a portfolio of plain stocks, whose payoff really is a straight line. It is dangerously wrong for options.

Analogy. Parametric VaR draws the tangent line to a curve and pretends the curve is the line. Near the touch point, the lie is harmless. Out in the tail — exactly where VaR lives — the curve has peeled far away from its tangent, and the line is telling you a fairy tale.

An option’s payoff curves: it has gamma, the curvature of its value as the underlying moves (delta’s own rate of change). Monte Carlo doesn’t approximate — in step 3 it fully reprices every option at each simulated price, running the actual pricing model. So when a scenario throws a violent move, the simulated P&L reflects the option’s true curved payoff, gamma and all. The tail of an optioned book — short straddles, structured notes, anything convex — comes out right instead of linearised into nonsense.

Worked intuition. You’re short a put, delta $-0.5$ today, on a $100 stock. Parametric VaR sees a 10% down-move and predicts a loss of about 0.5 × $10 = $5 per share — straight-line. But as the stock falls, the put’s delta steepens toward $-1$ (that’s gamma at work), so the real loss accelerates past $5, maybe to $8 or $9 per share. Monte Carlo, repricing the put at the simulated $90, sees that accelerating loss. Delta-normal never does — it’s still riding its flat tangent line off the cliff, reporting a comfortable number while the real book bleeds.

Convergence and standard error

Monte Carlo gives a different answer every time you run it, because it’s built on random draws. Run 10,000 paths today and 10,000 tomorrow and your VaR estimate wobbles a little between them. That wobble is sampling error, and the central fact about it is how it shrinks.

Definition. The standard error of a Monte Carlo estimate falls like $1/\sqrt{M}$, where $M$ is the number of paths. In plain terms: to halve the error, you must quadruple the paths. Accuracy is expensive — it improves only with the square root of effort, so the last digit of precision costs brutally more compute than the first.

Analogy. It’s like polling. A poll of 10,000 people isn’t twice as precise as one of 5,000 — it’s only about $\sqrt 2 \approx 1.4$ times as precise. Doubling the sample barely moves the margin of error; you have to quadruple it to cut that margin in half. Monte Carlo paths are voters, and the VaR is the election.

Worked example. Suppose at $M = 10{,}000$ paths your VaR estimate has a sampling standard error of $1.0 million. Push to $M = 40{,}000$ — a 4× increase. The error scales by $1/\sqrt{4} = 1/2$, so it drops to about $0.5 million. Quadruple the work, halve the noise. Want to halve it again, down to $0.25 million? That’s another 4×, to 160,000 paths. The compute bill compounds fast.

Because brute force is so costly, quants reach for variance-reduction tricks — ways to get the same accuracy from fewer paths:

• Antithetic variates. For every random scenario you draw, also use its mirror image (flip the sign of the random shocks). The pair partly cancels each other’s noise, so the average converges faster.
• Control variates. Lean on a related quantity whose answer you know exactly (say, the parametric VaR of a simplified version), and use the known error in that to correct your simulated estimate.

You don’t need to implement these today — just know they exist, and that a good engine reaches a target precision with far fewer paths than the naive $1/\sqrt M$ count would demand.

Strengths and weaknesses

Monte Carlo is the most flexible VaR method, full stop. You can feed it any distribution — fat-tailed Student-t instead of the thin-tailed normal, skewed, regime-switching, whatever fits the data. It handles non-linearity (options, convexity) by full repricing. It handles path-dependence (barrier options, anything whose payoff depends on the route, not just the endpoint) by simulating whole paths. Multiple correlated factors? Just enlarge the covariance matrix. Nothing else in the VaR toolkit is this general.

But that power comes with two heavy tolls.

The first is the one that bites in practice: it is only as good as the model you feed it. A normal model with too-low a $\sigma$, a stale correlation matrix, or a thin-tailed distribution doesn’t produce a vague or uncertain VaR — it produces a confidently precise wrong number. Ten thousand beautifully-sorted scenarios drawn from a bad model give you ten thousand beautifully-sorted lies. The pretty histogram launders the bad assumption into something that looks rigorous.

The second is mundane but real: it’s compute-heavy. Full repricing of a large book across tens of thousands of paths — sometimes nested, if instruments themselves need simulation to price — can mean millions of valuations per VaR run. Historical and parametric VaR are nearly free by comparison.

The three methods, side by side

You now hold all three VaR engines. Here’s the cheat-sheet that decides which to reach for.

The quick read: parametric is the fast, cheap default for linear books; historical is the assumption-light choice when you trust your data window and your book isn’t wildly optioned; Monte Carlo is the heavyweight you bring out for non-linear, multi-factor, fat-tailed books where getting the tail right is worth the compute.

Putting it together

Monte Carlo VaR manufactures the tomorrows reality won’t give you: pick a model, roll $M$ random scenarios, fully reprice the book in each, sort the P&Ls, and read the $(1-c)$ percentile of the losses. Its superpower is non-linearity — full repricing captures the gamma curvature that delta-normal linearises away — and its flexibility extends to any distribution, correlation structure, or path-dependence you can model. Its price is compute and, more insidiously, total dependence on the input model: feed it garbage and it returns a beautifully-rendered, high-confidence lie.

Next up — Expected Shortfall (CVaR) — we confront VaR’s deepest flaw: it tells you the line you rarely cross, but says nothing about how bad it gets once you do. A $10M VaR is identical whether the worst 1% averages an $11M loss or a $100M wipeout. Expected Shortfall answers the question VaR ducks — given we’ve blown past the threshold, what’s the average loss in that tail? — and it’s exactly the shaded region you’ve been watching to the left of every VaR line.