# Efficiently Simulating Geometric Brownian Motion in R

For simulating stock prices, Geometric Brownian Motion (GBM) is the de-facto go-to model.

It has some nice properties which are generally consistent with stock prices, such as being log-normally distributed (and hence bounded to the downside by zero), and that expected returns don’t depend on the magnitude of price.

Of course, GBM is just a model and no model is a perfect representation of reality. In particular, GBM uses constant volatility, which is clearly at odds with reality. It also doesn’t account for jumps, such as those caused by news.

In spite of those limitations, GBM is a useful starting point for modeling the behaviour of stock prices. In particular, it’s great for building intuition about various finance concepts – notably, options pricing.

Normally when we’re modeling stock prices, our use case requires running a large number of simulations in order to generate a distribution of possible outcomes.

Since such a use-case requires running a GBM simulator numerous times, it can pay to think about optimising code for speed. A small amount of forethought can save a ton of time down the track.

Of course, you want to avoid the temptation to optimise early lest you spend more time optimising code than what you gain in time saved by the optimisation itself.

But R lends itself to some simple out-of-the-box optimisations that provide great speed-up for little invested time.

In this post, I’ll demonstrate two approaches to simulating price paths using GBM:

• Using for loops to iterate over the number of price paths and the number of time-steps in each
• Vectorisation, where we operate on an entire vector or matrix at once

## Loop-based GBM simulation

Here’s some code for running a GBM simulation in a nested for loop:

gbm_loop <- function(nsim = 100, t = 25, mu = 0, sigma = 0.1, S0 = 100, dt = 1./365) {
gbm <- matrix(ncol = nsim, nrow = t)
for (simu in 1:nsim) {
gbm[1, simu] <- S0
for (day in 2:t) {
epsilon <- rnorm(1)
dt = 1 / 365
gbm[day, simu] <- gbm[(day-1), simu] * exp((mu - sigma * sigma / 2) * dt + sigma * epsilon * sqrt(dt))
}
}

return(gbm)
}


If I run it say, 50 times for 100 time-steps, with annaulised volatility of 10%, drift of 0 and a starting price of 100, I get price paths that look like this:

library(tidyverse)

nsim <- 50
t <- 100
mu <- 0
sigma <- 0.1
S0 <- 100

gbm <- gbm_loop(nsim, t, mu, sigma, S0)

gbm_df <- as.data.frame(gbm) %>%
mutate(ix = 1:nrow(gbm)) %>%
pivot_longer(-ix, names_to = 'sim', values_to = 'price')

gbm_df %>%
ggplot(aes(x=ix, y=price, color=sim)) +
geom_line() +
theme(legend.position = 'none') This looks like a reasonable representation of a random price process described by the parameters specified above. And that loop actually ran pretty quickly.

Let’s see how fast this thing runs if we ask it for 50,000 simulations:

start <- Sys.time()
gbm <- gbm_loop(nsim = 50000, t, mu, sigma, S0)
Sys.time() - start


## Time difference of 9.983079 secs


About ten seconds. Not the end of the world, but one could imagine this quickly becoming tedious.

## Vectorised approach to GBM simulation

Many operations in R are vectorised – which means that operations can occur in parallel under the hood, or at least can run much faster using tight loops written in C and hidden from the user.

The classic example of vectorisation in action is elementwise addition of two vectors. The for-loop version of such an operation looks like this:

x <- c(1:10)
y <- c(10:1)

z <- numeric(length(x))
for(i in c(1:length(x))) {
z[i] <- x[i] + y[i]
}

z


##   11 11 11 11 11 11 11 11 11 11


That’s quite a lot of code….

With vectorisation, we can simply do:

z <- x + y
z


##   11 11 11 11 11 11 11 11 11 11


Lots of operations in R are vectorised – in fact, R was designed with this in mind.

Let’s vectorise an operation in our GBM simulator to demonstrate.

Instead of generating a new random number for each simulation for each day as we did in the loop version, we’ll generate a matrix of all the random numbers we’ll need for the entire simulation, at the outset. That’s the matrix epsilon in the code below.

Then, we can transform that matrix in a single operation to nsim * t realisations of a GBM with our desired parameters.

As a final step, we add an initial price given by S0 to the first element of each simulation, then we take the cumulative product through time to get our price paths.

Here’s the code:

gbm_vec <- function(nsim = 100, t = 25, mu = 0, sigma = 0.1, S0 = 100, dt = 1./365) {

# matrix of random draws - one for each day for each simulation
epsilon <- matrix(rnorm(t*nsim), ncol = nsim, nrow = t)

# get GBM and convert to price paths
gbm <- exp((mu - sigma * sigma / 2) * dt + sigma * epsilon * sqrt(dt))
gbm <- apply(rbind(rep(S0, nsim), gbm), 2, cumprod)

return(gbm)
}


If I run it 50 times, I get price paths that look like this:

nsim <- 50
t <- 100
mu <- 0
sigma <- 0.1
S0 <- 100

gbm <- gbm_vec(nsim, t, mu, sigma, S0)

gbm_df <- as.data.frame(gbm) %>%
mutate(ix = 1:nrow(gbm)) %>%
pivot_longer(-ix, names_to = 'sim', values_to = 'price')

gbm_df %>%
ggplot(aes(x=ix, y=price, color=sim)) +
geom_line() +
theme(legend.position = 'none') All good so far.

Let’s ask it for 50,000 simulated price paths and see if we get a speed-up over our loop version:

start <- Sys.time()
gbm <- gbm_vec(nsim = 50000, t, mu, sigma, S0)
Sys.time() - start


## Time difference of 0.894985 secs


Nice! That’s the best part of an order of magnitude speed-up.

## What could a GBM simulator be used for?

We could use it to estimate the distribution of prices at some point in the future, given our model assumptions:

data.frame(price = gbm[t, ]) %>%
ggplot(aes(x = price)) +
geom_histogram(aes(y = ..density..), binwidth = 0.1) +
geom_density() +
ggtitle('terminal price distribution') And from there, estimate the probability-weighted payoff curve for an option on the stock being simulated, say a call option struck at 105 (again, given our model assumptions, and ignoring forward interest rates and dividends):

D <- gbm[t, ] %>%
density()

strike = 105

profile <- tibble(
price = D$x, value = case_when(price <= strike ~ 0, TRUE ~ price - strike) ) # dataframe of payoffs and probabilities prob_wieghted_payoff_profile <- profile %>% mutate(density = D$y/sum(D\$y))

prob_wieghted_payoff_profile %>%
ggplot(aes(x = price, y = value*density)) +
geom_line() +
xlab('price') +
ylab('probability-weighted payoff') And finally, we can get the expected value of our option by summing the area under the probability-weighted payoff curve:

expected_value <- prob_wieghted_payoff_profile %>%
summarise(ev = sum(density * value))

expected_value


## # A tibble: 1 x 1
##      ev
##
## 1 0.510


## Conclusion

A Geometric Brownian Motion simulator is one of the first tools you reach for when you start modeling stock prices.

In particular, it’s a useful tool for building intuition about concepts such as options pricing.

Leveraging R’s vectorisation tools, we can run tens of thousands of simulations in no time at all.

### 3 thoughts on “Efficiently Simulating Geometric Brownian Motion in R”

1. Nice

2. 