A Framework for Variance Reduction.

VarRedOpt

• Installation
• Example
• Details and Algorithms
• simulate outer
• Antithetic Variates
• Inner Control Variates
• Outer Control Variates
• Importance Sampling
• Adding Custom Function to VarRedOpt
• License

The increase in computing power has made us capable to run bigger simulations. We can choose bigger sample sizes with bigger dimensions. Nevertheless, this phenomenon does not make the need for an efficient simulation disappear. We still have to choose the most efficient way to make our simulations in order to get the most robust results with the computing power at the hand.

The reliability of the simulation lies in the variance of the simulation. As the simulation size increase, variance is expected to be decreased. We can increase simulation size to the point but after a certain point the simulation time will be infeasible to get results. This problem reveals the need for a different approach. We need a set of tools to get more robust simulation results with the same simulation size. This is where variance reduction (VR) algorithms come to our help.

A Variance Reduction Algorithm is an algorithm that behaves like the simulation itself. These algorithms uses simulations as an input and returns another simulation with approximately the same expected value and less variance value.

In this library, we are sharing different VR algorithms as a framework. Antithetic Variates, Inner Control Variates, Outer Control Variates and Importance Sampling algorithms are applied and presented as ready-to-use manner. Any user can run their simulations with different combinations of these methods and get advantage of these variance reduction algorithms.

Simulation is needed to approximate the most probable behavior of the system or the solution of the problem at the hand. In order to do that, we use different simulation techniques. These techniques can take random variables as input and attempt to approximate the solution of the our problem.

Our variance reduction framework makes it easier to conduct experiments without writing variance reduction algorithms. The idea is to get desired variance reduction algorithms from the user via a function that launches the simulation process and prints the simulation results. The launcher function is the only function user needs to fill with parameters. The parameters of this function are the name of the variance reduction algorithms to be applied, the naive simulation and parameters to be passed on the simulation function. The name of this launcher function in our framework is sim.outer(. . . ). Variance reduction algorithms are already implemented in our framework and the user can easily take advantage of these methods by simply writing their function names.

Installation

You can install the released version of VarRedOpt from CRAN with:

install.packages("VarRedOpt")

And the development version from GitHub with:

# install.packages("devtools")
devtools::install_github("onurboyar/VarRedOpt")

Example

To make things more concrete, let’s specify the parameters needed to simulate an Asian Call Option. To simulate an Asian Call Option, we need to have

• Strike price (K)
• Interest rate (riskfree rate) (r)
• Starting price (S0)
• Sigma, yearly volatility (sigma)
• Time (ti)

To launch a simulation process via sim.outer function, we need to give the simulation size (n) and the dimension values (d).

In order to get naive simulation without applying any of the variance reduction algorithms, we need just need to give the above parameters to sim.outer function in the following way.

devtools::install_github("onurboyar/VarRedOpt")
#>      checking for file ‘/private/var/folders/x0/92n0x33d1q33x2l2th2d_2yr0000gn/T/RtmpuTmCH8/remotes8d5937a4381e/onurboyar-VarRedOpt-d9b29a0/DESCRIPTION’ ...  ✓  checking for file ‘/private/var/folders/x0/92n0x33d1q33x2l2th2d_2yr0000gn/T/RtmpuTmCH8/remotes8d5937a4381e/onurboyar-VarRedOpt-d9b29a0/DESCRIPTION’
#>   ─  preparing ‘VarRedOpt’:
#>      checking DESCRIPTION meta-information ...  ✓  checking DESCRIPTION meta-information
#>   ─  checking for LF line-endings in source and make files and shell scripts
#>   ─  checking for empty or unneeded directories
#>   ─  building ‘VarRedOpt_0.1.0.tar.gz’
#>      
#> 
library(VarRedOpt)
sim.outer(n=1e5, d=3, q.outer = myq_asian,
               K = 100, ti=(1:3)/12, r = 0.03, sigma = 0.3, S0 = 100)
#>    Estimation StandardError 
#>    4.54300000    0.04283238

sim.outer function creates a matrix with 10⁵ rows and 3 columns. Values in that matrix are drawn from standard normal distribution. This matrix is sent to myq_asian function as an input and Asian Call Option prices are simulated. The simulated values are sent back to sim.outer function to calculate estimation and standard error values. User sees these values as the output of the simulation.

There are a lot of different parameters to be used in different functions in our framework. In order to handle different parameters and to create a flexible framework, we are taking advantage of ellipsis parameter (. . . ) of R inside our functions.

Details and Algorithms

The main function of our framework is sim.outer() function. It simulates the input variables, which are standard normal random variables. The size of this simulated in determined by the given parameters, n and d. Given these values, sim.outer() creates Z matrix which includes standard normal random variables. The Z matrix is the input of our target simulation. If we stick to our previous example, Z matrix will be passed to myq_asian() function within sim.outer() function. Asian Option function will simulate Asian Option prices using Z matrix and return calculated prices to our main function. The main function calculates expected value and variance of the returning values and prints them as final output.

simulate outer

If we set simulation size to 10⁷ we already have big simulation size and it is hard to run this simulation few times to check if we are getting consistent results. We can compare expected values obtained from these simulations with confidence interval and see if expected values are within the confidence interval. It is hard to perform this task if our simulation size is equal to 10⁶ or 10⁷ but it is not hard if it is equal to 10³. Besides, another problem is that all functions need to store several vectors of length n, this grows bigger if we have a dimension greater than 1, and it makes simulation hard to run due to memory constraints. We can run simulations of size 10³ few times, let’s say 10³ after running our main simulation with simulation size 10⁶. After estimated mean of our bigger simulation we can check if it lays within confidence intervals of these 10³ simulations and come up with different measure to evaluate our simulation. In our framework it is very simple to make such analysis. All needed to be done is to set auto.repetition parameter to a value rather than 1 in the following way.

sim.outer(n=1e5, d=3, auto_repetition = 100, q.outer = myq_asian, 
               K = 100, ti = (1:3)/12, r = 0.03, sigma = 0.3, S0 = 100)

Above function will run simulation with n = 1000 auto_repetition times, which is 100. Another aspect need to be mentioned of the sim.outer() function is the q.outer parameter. In the above example it is set to function that simulates Asian Option. In order to perform variance reduction via the algorithms within our framework, we need to give different parameter(s) to our main function. If we are to add Antithetic Variates within our framework, q.outer parameter will be set to sim.AV and the new parameter will appear. It is the parameter that we need inside sim.AV function to call in order to perform variance reduction. When we set q.outer to sim.AV, we need to specify another parameter to tell our framework the function to be simulated, like Asian Option. Since we are using sim.AV function in this example, it must be called inside of the sim.AV function.

Antithetic Variates

In order to add antithetic variates to our framework and simulate asian options we need to slightly change the sim.outer function given above. Antithetic Variates calls the simulation function twice by using opposite signed inputs. The q.outer parameter of our main function will be sim.AV. In order to specify the function to be simulated we need another parameter. This parameter is named as q.av. It’s name includes av because this parameter is only needed when we are using sim.AV within our framework. We will give q.av = myq_asian this time. The function now becomes

sim.outer(n=1e5, d=3, q.outer = sim.AV, 
               q.av = myq_asian, K = 100, ti = (1:3)/12, r = 0.03, sigma = 0.3, S0 = 100)
#>    Estimation StandardError 
#>    4.55000000    0.02285073

Inner Control Variates

Like antithetic variates, inner control variates algorithm does not require any additional parameters. It can be directly applied to naive simulation since it uses only input variables as control variates. To run our simulation with inner control variates, we need to assign different function to q.outer parameter. Our framework has a built-in function named sim.InnerCV(). In order to simulate Asian Option with Inner Control Variates, the following function can be used.

sim.outer(n=1e5, d=3, q.outer = sim.InnerCV, 
               q.cv = myq_asian,K = 100, ti =(1:3)/12, r = 0.03, sigma = 0.3, S0 = 100)
#>    Estimation StandardError 
#>    4.53800000    0.01698056

Note that we are using q.cv parameter this time.

Outer Control Variates

Outer Control Variates approach is using the result of a similar problem to the task at hand in which the exact solution is known. Deciding good outer control variate results in a great amount of variance reduction. The main disadvantage of this method is that it is difficult and requires domain knowledge to come up with such a control variate; also the additional computations may be quite slow depending on the control variate.

In Option Pricing, we can use prices of other options as Outer Control Variates. Since we use expected values of the control variates in our calculations, the exact solution of the future prices must be known.

As an example we can take Asian Options again. If we want to simulate Asian Option prices using Outer Control variates we can use Asian Call Option with Geometric Mean as an Outer Control Variate since it’s exact solution is known.

sim.outer(n=1e5, d=3, q.outer = sim.AV, q.av = sim.GeometricAvg, 
               q.ga = myq_asian,K = 100, ti = (1:3)/12, r = 0.03, sigma=0.3, S0=100)
#>    Estimation StandardError 
#>  4.5400000000  0.0006227557

Outer Control Variates function is the same as Inner Control Variates function in several ways. It has the same control for returning list lengths and application logic of the IS weight is the same. The major difference is the control variate itself. In Inner Control Variates function, we use input columns as control variates. In Geometric Average Outer Control Variate, we use product of the prices as control variates and it is returned from our myq_asian function. The expectation of this control variate is calculated by BS_Asian_geom function.

Using IS and CV together is not as straightforward as using AV and CV or AV and IS together. Applying IS weight and returning updated simulation values to CV does not work well. Following the opposite approach, applying CV and returning updated values to IS and applying IS weight to these values does not work well either. What we do is that applying multiplying target values with IS weight after decreasing variance by applying Control Variates.

Importance Sampling

Importance Sampling is a variance reduction technique that is especially useful in rare event simulation. In an Option Pricing problem like Asian Call Options, when we have strike price like 140 we have a rare event simulation problem.

To be able to sample from tails, we sample from the different distribution and fix the error of not sampling from the correct distribution in Importance Sampling.

In our framework, in order to apply Importance Sampling, user should specify two parameters. One is muis and the other is sis. muis is the mean value of the importance density and sis is the standard deviation of the importance density.

If user do not want to specify muis value, use_pilot_study parameter can be used to look for an optimum muis value by using a pilot study. In current version, muis values in an interval [1.01, 1.2] is used as candidate values for muis values. Starting from 1.01, values are incremented by 0.01 until it is reached to 1.2. sis, standard deviation of the IS density, is not recommended to be changed. We do not offer optimization for sis value. Nevertheless, our framework look for optimum muis value by conducting a pilot study inside the function.

sim.outer(1e6,d=3,q.outer=myq_asian,K=120,
               ti=(1:3)/12,r=0.03,sigma=0.3,S0=100)
#>    Estimation StandardError 
#>   0.255000000   0.003160006

Now, let’s add Inner Control Variates and obtain a little variance reduction.

sim.outer(n=1e6,d=3,q.outer=sim.InnerCV,
               q.cv=myq_asian,K=120,ti=(1:3)/12,r=0.03,sigma=0.3,S0=100)
#>    Estimation StandardError 
#>    0.25500000    0.00261813

Because we set strike price to 120, simulated results are rare. We expect IS to reduce variance. Let’s use IS and Inner CV together. We apply IS weights after calculating simulated values in Control Variates function.

sim.outer(n=1e6,d=3,q.outer=sim.InnerCV,q.cv=sim.IS,muis=1.03,sis=1,
               q.is=myq_asian,K=120,ti=(1:3)/12,r=0.03,sigma=0.3,S0=100)
#>    Estimation StandardError 
#>   0.256000000   0.000614824

Adding Custom Function to VarRedOpt

In this section motivating examples will be given to show how to perform naive simulation using our framework. If we want to simulate Euclidean distance of iid N(0, 1) vector to given point, we can write following function.

myq_euclidean <- function(zm,point=c(1,2,1)){
  # returns Euclidean distance of iid N(0,1) vector to "point"
  d <- length(point)
  sumDist2 <- 0
  for(i in 1:d) sumDist2 <- sumDist2 + (point[i]-zm[,i])^2
  
  returning_list = list(sqrt(sqrt(sumDist2)))
  return(returning_list)
}

This function, myq_euclidean, takes two parameters. The length of the point vector and the dimension of the z.matrix should be the same in order to find euclidean distance between these points. Note that returning value type have to be list.

Now, let’s simulate myq using our framework.

sim.outer(n=1e6,d=2,q.outer=myq_euclidean,point=c(1,3))
#>    Estimation StandardError 
#>  1.8020000000  0.0005443944

Let’s see the output when we use auto repetition.

sim.outer(n=1e6,d=2,auto_repetition=1000,q.outer=myq_euclidean,point=c(1,3))
#>              Estimation           StandardError ConfidenceIntervalRatio 
#>            1.8020000000            0.0005443944            0.9470000000

We observe that estimated value is in the confidence interval 947 times out of 1000.

LICENSE

MIT © Wolfgang Hörmann, Onur Boyar.