## Improving Performance of Monte Carlo Simulation with Parallel Computing

This example shows how to improve the performance of a Monte Carlo simulation using Parallel Computing Toolbox™.

Consider a geometric Brownian motion (GBM) process in which you want to incorporate alternative asset price dynamics. Specifically, suppose that you want to include a time-varying short rate and a volatility surface. The process to simulate is written as

$dS\left(t\right)=r\left(t\right)S\left(t\right)dt+V\left(t,S\left(t\right)\right)S\left(t\right)dW\left(t\right)$

for stock price S(t), rate of return r(t), volatility V(t,S(t)), and Brownian motion W(t). In this example, the rate of return is a deterministic function of time and the volatility is a function of both time and current stock price. Both the return and volatility are defined on a discrete grid such that intermediate values are obtained by linear interpolation. For example, such a simulation can be used to support the pricing of thinly traded options.

To include a time series of riskless short rates, suppose that you derive the following deterministic short-rate process as a function of time.

times = [0 0.25 0.5 1 2 3 4 5 6 7 8 9 10];  % in years
rates = [0.1 0.2 0.3 0.4 0.5 0.8 1.25 1.75 2.0 2.2 2.25 2.50 2.75]/100;

Suppose that you then derive the following volatility surface whose columns correspond to simple relative moneyness, or the ratio of the spot price to strike price, and whose rows correspond to time to maturity, or tenor.

surface = [28.1 25.3 20.6 16.3 11.2  6.2  4.9  4.9  4.9  4.9  4.9  4.9
22.7 19.8 15.4 12.6  9.6  6.7  5.2  5.2  5.2  5.2  5.2  5.2
21.7 17.6 13.7 11.5  9.4  7.3  5.7  5.4  5.4  5.4  5.4  5.4
19.8 16.4 12.9 11.1  9.3  7.6  6.2  5.6  5.6  5.6  5.6  5.6
18.6 15.6 12.5 10.8  9.3  7.8  6.6  5.9  5.9  5.9  5.9  5.9
17.4 13.8 11.7 10.8  9.9  9.1  8.5  7.9  7.4  7.3  7.3  7.3
17.1 13.7 12.0 11.2 10.6 10.0  9.5  9.1  8.8  8.6  8.4  8.0
17.5 13.9 12.5 11.9 11.4 10.9 10.5 10.2  9.9  9.6  9.4  9.0
18.3 14.9 13.7 13.2 12.8 12.4 12.0 11.7 11.4 11.2 11.0 10.8
19.2 19.6 14.2 13.9 13.4 13.0 13.2 12.5 12.1 11.9 11.8 11.4]/100;

tenor = [0 0.25 0.50 0.75 1 2 3 5 7 10];   % in years
moneyness = [0.25 0.5 0.75 0.8 0.9 1 1.10 1.25 1.50 2 3 5];

Set the simulation parameters. The following assumes that the price of the underlying asset is initially equal to the strike price and that the price of the underlying asset is simulated monthly for 10 years, or 120 months. As a simple illustration, 100 sample paths are simulated.

price = 100;
strike = 100;
dt = 1/12;
NPeriods = 120;
NTrials = 100;

For reproducibility, set the random number generator to its default, and draw the Gaussian random variates that drive the simulation. Generating the random variates is not necessary to incur the performance improvement of parallel computation, but doing so allows the resulting simulated paths to match those of the conventional (that is, non-parallelized) simulation. Also, generating independent Gaussian random variates as inputs also guarantees that all simulated paths are independent.

rng default
Z = randn(NPeriods,1,NTrials);

Create the return and volatility functions and the GBM model using gbm. Notice that the rate of return is a deterministic function of time, and therefore accepts simulation time as its only input argument. In contrast, the volatility must account for the moneyness and is a function of both time and stock price. Moreover, since the volatility surface is defined as a function of time to maturity rather than simulation time, the volatility function subtracts the current simulation time from the last time at which the price process is simulated (10 years). This ensures that as the simulation time approaches its terminal value, the time to maturity of the volatility surface approaches zero. Although far more elaborate return and volatility functions could be used if desired, the following assumes simple linear interpolation.

mu = @(t) interp1(times,rates,t);
sigma = @(t,S) interp2(moneyness,tenor,surface,S/strike,tenor(end)-t);
mdl = gbm(mu,sigma,'StartState',price);

Simulate the paths of the underlying geometric Brownian motion without parallelization.

tStart = tic;
paths = simBySolution(mdl,NPeriods,'NTrials',NTrials,'DeltaTime',dt,'Z',Z);
time1 = toc(tStart);

Simulate the paths in parallel using a parfor (Parallel Computing Toolbox) loop. For users licensed to access the Parallel Computing Toolbox, the following code segment automatically creates a parallel pool using the default local profile. If desired, this behavior can be changed by first calling the parpool (Parallel Computing Toolbox) function. If a parallel pool is not already created, the following simulation will likely be slower than the previous simulation without parallelization. In this case, rerun the following simulation to assess the true benefits of parallelization.

tStart = tic;
parPaths = zeros(NPeriods+1,1,NTrials);
parfor i = 1:NTrials
parPaths(:,:,i) = simBySolution(mdl,NPeriods,'DeltaTime',dt,'Z',Z(:,:,i));
end
time2 = toc(tStart);

If you examine any given path obtained without parallelization to the corresponding path with parallelization, you see that they are identical. Moreover, although relative performance varies, the results obtained with parallelization will generally incur a significant improvement. To assess the performance improvement, examine the runtime of each approach in seconds and the speedup gained from simulating the paths in parallel.

time1                 % elapsed time of conventional simulation, in seconds
time2                 % elapsed time of parallel simulation, in seconds
speedup = time1/time2 % speedup factor
time1 =
6.1329
time2 =
2.5918
speedup =
2.3663