Quantitative Finance Stack Exchange is a question and answer site for finance professionals and academics. It's 100% free, no registration required.

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

I've been using Monte Carlo simulation (MC) for pricing vanilla options with non-lognormal underlyings returns.

I'm tempted to start using MC as my primary option-valuating technique as I can get sound results without relying on the assumptions of the analytical methods (Black-Scholes, for example).

Computational costs aside, when to use MC simulation over analytical methods for options pricing?

share|improve this question
up vote 19 down vote accepted

Monte Carlo is most useful when you lack analytic tractability or when you have a highly multidimensional problem.

For example, even using simple lognormal and poisson models, there exist path-dependent payoffs or multi-asset computations such that no analytic solution exists and such that any PDE finite difference solution would require 3 or more dimensions. Other times, you are employing a model where the SDE is not solvable, so an apparently one-dimensional problem still ends up forcing you to generate many incremental paths using Euler or Milstein integration.

Cases Where Monte Carlo Is A Poor Idea

  • Weakly path-dependent options (e.g. lookbacks): Use PDE or series solutions
  • Single-dimensional cases: If your problem is just one dimensional, such as pricing a payoff along the terminal distribution, you should never use Monte Carlo, since numerical quadrature is far superior in this case, even if you just use Riemann sums.

Cases Where Monte Carlo Is A Good Idea

  • Strongly path-dependent options such as ratchet range options
  • Portfolio risks and exotic baskets where high dimensionality comes into play. CDO tranche protection is a classic example. So are tail risk computations, especially for multi-asset portfolios.
  • Intractable models where the terminal distribution cannot be computed, such as some stochastic vol models

To the point about single-dimensional cases -- it sounds like this describes your usage, perhaps because you are using some kind of implied distributional fit to agree with volatility skew. In this case Monte Carlo seems easy, but using a trapezoid rule integrator (or similar) will be as easy and far higher quality by about any measure.

Now Monte Carlo does make it tricky to accurately compute greeks. As with any model, we can compute greeks by using a finite difference "parameter bump", computing our greek

$$ g_\mu =\frac{ V(\dots,\mu+\Delta \mu,\dots) - V(\dots, \mu,\dots)}{\Delta \mu} $$

but if there is a lot of random noise in those two separate computations of $V()$ then our $g_\mu$ will be inaccurate. Instead it is important to bring the differencing inside the Monte Carlo formula. That is, we don't want to be doing

$$ \hat{g}_\mu =\frac{ \frac1M \sum_{i=1}^M V(x_i,\dots,\mu+\Delta \mu,\dots) - \frac1M \sum_{i=1}^MV(y_i, \dots, \mu,\dots)}{\Delta \mu} $$

for two separate sample sets $x_i$ and $y_i$. Instead, we want to use the same $x_i$ for both sums, meaning we effectively compute

$$ g_\mu =\frac1{M {\Delta \mu}} \sum_{i=1}^M V(x_i,\dots,\mu+\Delta \mu,\dots) -V(x_i, \dots, \mu,\dots) $$

and end up with a far more accurate estimate, typically better than our estimate of option value.

I'll make one final note, which is that you feel you "get sound results without relying on the assumptions of the analytical methods". If your terminal distributions are empirically generated, then you are likely to misprice any options because you are not using anything close to a risk-neutral measure. For example, you almost certainly find yourself pricing a forward contract $F$ far higher than the true, arbitrageable value range $F \in [S_0 e^{r_L T}, S_0 e^{r_b T}]$ where $r_b, r_L$ are standard borrow-lend rates.

(Vytautas beat me to some of these points)

share|improve this answer
+1 for mentioning multidimensionality – Alexey Kalmykov Oct 2 '12 at 18:25

Normally, one uses MC methods when:

  • Analytical solutions do not exist
  • PDE style solutions also don't work (they are usually still faster than MC)
  • You need to price some exotic, but computation time does not matter (MC methods are easy(-ier) and fast to code-up)

Note: Using MC is not free of assumptions: you always assume a distribution for the driving asset, as well as structure for its volatility, absence or existence of jumps...

Having analytic solutions gives you an easy way to calibrate your model parameters to market prices - if you use MC, then... try calibrating your model to 50, 100 option quotes - this does not work within seconds... unless you're using FPGAs, GPUs, etc...

share|improve this answer
You may want to add something about Greeks. – John Oct 2 '12 at 15:17
Further more you can estimate models without normality much easier. Just sample from the distribution. For many contingent claims on stochastic processes the PDE equivalent does not really exist in a classical form (PIDE for jump processes for example). I think this is the main strength of MC. – vanguard2k Oct 2 '12 at 15:21
@vanguard2k, you can use fourier time stepping methods for American style options with jump/levy processes. MC is almost never the right tool except as a check for analytic/quasi analytic methods or for portfolio loss modeling (eg VaR for a large options portfolio). – user9403 Apr 14 at 21:25

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.