[Returnanalytics-commits] r3862 - pkg/Dowd/R

noreply at r-forge.r-project.org noreply at r-forge.r-project.org
Tue Jul 28 10:15:23 CEST 2015


Author: dacharya
Date: 2015-07-28 10:15:23 +0200 (Tue, 28 Jul 2015)
New Revision: 3862

Added:
   pkg/Dowd/R/tES.R
   pkg/Dowd/R/tESDFPerc.R
   pkg/Dowd/R/tESFigure.R
Log:
Function tES added.

Added: pkg/Dowd/R/tES.R
===================================================================
--- pkg/Dowd/R/tES.R	                        (rev 0)
+++ pkg/Dowd/R/tES.R	2015-07-28 08:15:23 UTC (rev 3862)
@@ -0,0 +1,145 @@
+#' ES for t distributed P/L
+#' 
+#' Estimates the ES of a portfolio assuming that P/L are 
+#' t-distributed, for specified confidence level and holding period.
+#' 
+#' @param ... The input arguments contain either return data or else mean and 
+#'  standard deviation data. Accordingly, number of input arguments is either 4 
+#'  or 5. In case there 4 input arguments, the mean and standard deviation of 
+#'  data is computed from return data. See examples for details.
+#'  
+#'  returns Vector of daily P/L data
+#' 
+#'  mu Mean of daily geometric return data
+#' 
+#'  sigma Standard deviation of daily geometric return data
+#' 
+#'  df Number of degrees of freedom in the t-distribution
+#' 
+#'  cl ES confidence level
+#' 
+#'  hp ES holding period in days
+#'  
+#' @return Matrix of ES whose dimension depends on dimension of hp and cl. If 
+#' cl and hp are both scalars, the matrix is 1 by 1. If cl is a vector and hp is
+#'  a scalar, the matrix is row matrix, if cl is a scalar and hp is a vector, 
+#'  the matrix is column matrix and if both cl and hp are vectors, the matrix 
+#'  has dimension length of cl * length of hp.
+#'  
+#' @references Dowd, K. Measuring Market Risk, Wiley, 2007.
+#' 
+#' Evans, M., Hastings, M. and Peacock, B. Statistical Distributions, 3rd 
+#' edition, New York: John Wiley, ch. 38,39.
+#'
+#' @author Dinesh Acharya
+#' @examples
+#' 
+#'    # Computes ES given P/L data
+#'    data <- runif(5, min = 0, max = .2)
+#'    tES(returns = data, df = 6, cl = .95, hp = 90)
+#'    
+#'    # Computes ES given mean and standard deviation of P/L data
+#'    tES(mu = .012, sigma = .03, df = 6, cl = .95, hp = 90)
+#'
+#'
+#' @export
+tES <- function(...){
+  if (nargs() < 4) {
+    stop("Too few arguments")
+  }
+  if (nargs() > 5) {
+    stop("Too many arguments")
+  }
+  args <- list(...)
+  if (nargs() == 5) {
+    mu <- args$mu
+    df <- args$df
+    cl <- args$cl
+    sigma <- args$sigma
+    hp <- args$hp
+  }
+  if (nargs() == 4) {
+    mu <- mean(args$returns)
+    df <- args$df
+    cl <- args$cl
+    sigma <- sd(args$returns)
+    hp <- args$hp
+  }
+  
+  # Check that inputs have correct dimensions
+  mu <- as.matrix(mu)
+  mu.row <- dim(mu)[1]
+  mu.col <- dim(mu)[2]
+  if (max(mu.row, mu.col) > 1) {
+    stop("Mean must be a scalar")
+  }
+  sigma <- as.matrix(sigma)
+  sigma.row <- dim(sigma)[1]
+  sigma.col <- dim(sigma)[2]
+  if (max(sigma.row, sigma.col) > 1) {
+    stop("Standard deviation must be a scalar")
+  }
+  cl <- as.matrix(cl)
+  cl.row <- dim(cl)[1]
+  cl.col <- dim(cl)[2]
+  if (min(cl.row, cl.col) > 1) {
+    stop("Confidence level must be a scalar or a vector")
+  }
+  hp <- as.matrix(hp)
+  hp.row <- dim(hp)[1]
+  hp.col <- dim(hp)[2]
+  if (min(hp.row, hp.col) > 1) {
+    stop("Holding period must be a scalar or a vector")
+  }
+  df <- as.matrix(df)
+  df.row <- dim(df)[1]
+  df.col <- dim(df)[2]
+  if (max(df.row, df.col) > 1) {
+    stop("Number of degrees of freedom must be a scalar")
+  }
+  
+  # Check that cl and hp are read as row and column vectors respectively
+  if (cl.row > cl.col) {
+    cl <- t(cl)
+  }
+  if (hp.row > hp.col) {
+    hp <- t(hp)
+  }
+  
+  # Check that inputs obey sign and value restrictions
+  if (sigma < 0) {
+    stop("Standard deviation must be non-negative")
+  }
+  if (df < 3) {
+    stop("Number of degrees of freedom must be at least 3 for first two moments of distribution to be defined")
+  }
+  if (max(cl) >= 1){
+    stop("Confidence level(s) must be less than 1")
+  }
+  if (min(cl) <= 0){
+    stop("Confidence level(s) must be greater than 0")
+  }
+  if (min(hp) <= 0){
+    stop("Holding Period(s) must be greater than 0")
+  }
+  
+  # ES estimation
+  ES <- matrix(0, length(hp), length(cl))
+  for (i in 1:length(cl)) {
+    for (j in 1:length(hp)) {
+      ES[j, i] <- Univariate.tES(mu, sigma, df, cl[i], hp[j])
+    }
+  }
+  return (ES)
+}
+
+# Accessory function
+Univariate.tES <- function (mu, sigma, df, cl, hp) {
+  # This function estimates univariate t-ES using average tail quantile algorithm
+  number.slices <- 1000
+  delta.p <- (1 - cl)/number.slices
+  p <- seq(cl + delta.p, 1 - delta.p, delta.p) # Tail confidence levels or cumulative probs
+  tail.VaRs <- -sigma * sqrt(hp) * sqrt((df - 2)/df) * qt((1 - p), df) - mu * hp # Tail VaRs
+  y <- mean(tail.VaRs)
+  return(y)
+}
\ No newline at end of file

Added: pkg/Dowd/R/tESDFPerc.R
===================================================================
--- pkg/Dowd/R/tESDFPerc.R	                        (rev 0)
+++ pkg/Dowd/R/tESDFPerc.R	2015-07-28 08:15:23 UTC (rev 3862)
@@ -0,0 +1,179 @@
+#' Percentiles of ES distribution function for t-distributed  P/L
+#' 
+#' Estimates percentiles of ES distribution function for t-distributed P/L, 
+#' using the theory of order statistics
+#' 
+#' @param ... The input arguments contain either return data or else mean and 
+#' standard deviation data. Accordingly, number of input arguments is either 5 
+#' or 7. In case there 5 input arguments, the mean, standard deviation and assumed sampel size of 
+#' data is computed from return data. See examples for details.
+#' 
+#'  returns Vector of daily geometric return data
+#' 
+#'  mu Mean of daily geometric return data
+#'  
+#'  sigma Standard deviation of daily geometric return data
+#' 
+#'  n Sample size
+#' 
+#'  df Degrees of freedom
+#' 
+#'  perc Desired percentile
+#' 
+#'  df Number of degrees of freedom in the t distribution
+#' 
+#'  cl ES confidence level and must be a scalar
+#' 
+#'  hp ES holding period and must be a a scalar
+#'  
+#' @return Percentiles of ES distribution function
+#'  
+#' @references Dowd, K. Measuring Market Risk, Wiley, 2007.
+#'
+#'
+#' @author Dinesh Acharya
+#' @examples
+#' 
+#'    # Estimates Percentiles of ES distribution given P/L data
+#'    data <- runif(5, min = 0, max = .2)
+#'    tESDFPerc(returns = data, perc = .7, df = 6, cl = .95, hp = 60)
+#'    
+#'    # Estimates Percentiles of ES distribution given mean, std. deviation and sample size
+#'    tESDFPerc(mu = .012, sigma = .03, n= 10, perc = .8, df = 6, cl = .99, hp = 40)
+#'
+#'
+#' @export
+tESDFPerc <- function(...){
+  if (nargs() < 5) {
+    stop("Too few arguments")
+  }
+  if (nargs() == 6) {
+    stop("Incorrect number of arguments")
+  }
+  if (nargs() > 7) {
+    stop("Too many arguments")
+  }
+  args <- list(...)
+  if (nargs() == 7) {
+    mu <- args$mu
+    df <- args$df
+    cl <- args$cl
+    perc <- args$sigma
+    n <- args$n
+    sigma <- args$sigma
+    hp <- args$hp
+  }
+  if (nargs() == 5) {
+    mu <- mean(args$returns)
+    df <- args$df
+    n <- max(dim(as.matrix(args$returns)))
+    perc <- args$perc
+    cl <- args$cl
+    sigma <- sd(args$returns)
+    hp <- args$hp
+  }
+  
+  # Check that inputs have correct dimensions
+  mu <- as.matrix(mu)
+  mu.row <- dim(mu)[1]
+  mu.col <- dim(mu)[2]
+  if (max(mu.row, mu.col) > 1) {
+    stop("Mean must be a scalar")
+  }
+  sigma <- as.matrix(sigma)
+  sigma.row <- dim(sigma)[1]
+  sigma.col <- dim(sigma)[2]
+  if (max(sigma.row, sigma.col) > 1) {
+    stop("Standard deviation must be a scalar")
+  }
+  n <- as.matrix(n)
+  n.row <- dim(n)[1]
+  n.col <- dim(n)[2]
+  if (max(n.row, n.col) > 1) {
+    stop("Number of observations in a sample must be a scalar")
+  }
+  if (n - round(n) > 0 | n - round(n) < 0) {
+    stop("Number of observations in a sample must be an integer")
+  }
+  perc <- as.matrix(perc)
+  perc.row <- dim(perc)[1]
+  perc.col <- dim(perc)[2]
+  if (max(perc.row, perc.col) > 1) {
+    stop("Chosen percentile of the distribution must be a scalar")
+  }
+  cl <- as.matrix(cl)
+  cl.row <- dim(cl)[1]
+  cl.col <- dim(cl)[2]
+  if (max(cl.row, cl.col) > 1) {
+    stop("Confidence level must be a scalar")
+  }
+  hp <- as.matrix(hp)
+  hp.row <- dim(hp)[1]
+  hp.col <- dim(hp)[2]
+  if (max(hp.row, hp.col) > 1) {
+    stop("Holding period must be a scalar")
+  }
+  
+  # Check that inputs obey sign and value restrictions
+  if (sigma < 0) {
+    stop("Standard deviation must be non-negative")
+  }
+  if (n < 0) {
+    stop("Number of observations must be non-negative")
+  }
+  if (df < 0) {
+    stop("Number of degrees of freedom must be greater than zero")
+  }
+  if (perc > 1){
+    stop("Chosen percentile must not exceed 1")
+  }
+  if (perc <= 0){
+    stop("Chosen percentile must be positive")
+  }
+  if (cl >= 1){
+    stop("Confidence level must be less than 1")
+  }
+  if (cl <= 0){
+    stop("Confidence level must be greater than 0")
+  }
+  if (hp <= 0){
+    stop("Honding period must be greater than 0")
+  }
+  
+  # Derive order statistic and ensure it is an integer
+  w <- n * cl # Derive r-th order statistic
+  r <- round(w) # Round r to nearest integer
+  # Bisection routine
+  a <- 0
+  fa <- -Inf
+  b <- 1
+  fb <- Inf
+  eps <- .Machine$double.eps
+  while (b - a > eps * b) {
+    x <- (a + b) / 2
+    fx <- 1 - pbinom(r - 1, n, x) - perc
+    if (sign(fx) == sign(fa)){
+      a = x
+      fa = fx
+    } else {
+      b = x
+      fb = fx
+    }
+  }
+  
+  # VaR estimation
+  VaR <- - mu * hp + sigma * sqrt((df-2)/df)*sqrt(hp)*qt(x, df) # Value of VaR percentile, note t VaR formula
+  
+  # ES estimation
+  cl <- x
+  n <- 1000 # Number of slices into which tail is divided
+  cl0 <- cl # Initial confidence level
+  delta.cl <- (1 - cl) / n # Increment to confidence level as each slice is taken
+  term <- VaR
+  for (i in 1:(n-1)) {
+    cl <- cl0 + i * delta.cl # Revised cl
+    term <- term - mu * hp - sigma * sqrt((df-2)/df)*sqrt(hp)*qt(1 - cl, df)
+  }
+  y <- term/n
+  return(y) # Value of ES percentile
+}

Added: pkg/Dowd/R/tESFigure.R
===================================================================
--- pkg/Dowd/R/tESFigure.R	                        (rev 0)
+++ pkg/Dowd/R/tESFigure.R	2015-07-28 08:15:23 UTC (rev 3862)
@@ -0,0 +1,173 @@
+#' Figure of t - VaR and ES and pdf against L/P
+#'
+#' Gives figure showing the VaR and ES and probability distribution function assuming P/L is t- distributed, for specified confidence level and holding period.
+#' 
+#' @param ... The input arguments contain either return data or else mean and 
+#'  standard deviation data. Accordingly, number of input arguments is either 4 
+#'  or 5. In case there 4 input arguments, the mean and standard deviation of 
+#'  data is computed from return data. See examples for details.
+#
+#'  returns Vector of daily geometric return data
+#' 
+#'  mu Mean of daily geometric return data
+#'  
+#'  sigma Standard deviation of daily geometric return data
+#'  
+#'  df Number of degrees of freedom
+#'  
+#'  cl VaR confidence level and should be scalar
+#'  
+#'  hp VaR holding period in days and should be scalar
+#'  
+#' @references Dowd, K. Measuring Market Risk, Wiley, 2007.
+#' 
+#' Evans, M., Hastings, M. and Peacock, B. Statistical Distributions, 3rd 
+#' edition, New York: John Wiley, ch. 38,39.
+#'
+#' @author Dinesh Acharya
+#' @examples
+#' 
+#'    # Plots lognormal VaR, ES and pdf against L/P data for given returns data
+#'    data <- runif(5, min = 0, max = .2)
+#'    tESFigure(returns = data, df = 10, cl = .95, hp = 90)
+#'    
+#'    # Plots lognormal VaR, ES and pdf against L/P data with given parameters
+#'    tESFigure(mu = .012, sigma = .03, df = 10, cl = .95, hp = 90)
+#'
+#' @export
+tESFigure <- function(...){
+  # Determine if there are four or five arguments and ensure that arguments are
+  # read as intended
+  if (nargs() < 4) {
+    stop("Too few arguments")
+  }
+  if (nargs() > 5) {
+    stop("Too many arguments")
+  }
+  args <- list(...)
+  if (nargs() == 5) {
+    mu <- args$mu
+    cl <- args$cl
+    df <- args$df
+    sigma <- args$sigma
+    hp <- args$hp
+  }
+  if (nargs() == 4) {
+    mu <- mean(args$returns)
+    cl <- args$cl
+    df <- args$df
+    sigma <- sd(args$returns)
+    hp <- args$hp
+  }
+  
+  # Check that inputs have correct dimensions
+  mu <- as.matrix(mu)
+  mu.row <- dim(mu)[1]
+  mu.col <- dim(mu)[2]
+  if (max(mu.row, mu.col) > 1) {
+    stop("Mean must be a scalar")
+  }
+  sigma <- as.matrix(sigma)
+  sigma.row <- dim(sigma)[1]
+  sigma.col <- dim(sigma)[2]
+  if (max(sigma.row, sigma.col) > 1) {
+    stop("Standard deviation must be a scalar")
+  }
+  cl <- as.matrix(cl)
+  cl.row <- dim(cl)[1]
+  cl.col <- dim(cl)[2]
+  if (min(cl.row, cl.col) > 1) {
+    stop("Confidence level must be a scalar or a vector")
+  }
+  df <- as.matrix(df)
+  df.row <- dim(df)[1]
+  df.col <- dim(df)[2]
+  if (max(df.row, df.col) > 1) {
+    stop("Number of degrees of freedom must be a scalar")
+  }
+  hp <- as.matrix(hp)
+  hp.row <- dim(hp)[1]
+  hp.col <- dim(hp)[2]
+  if (min(hp.row, hp.col) > 1) {
+    stop("Holding period must be a scalar or a vector")
+  }
+  
+  # Check that cl and hp are read as row and column vectors respectively
+  if (cl.row > cl.col) {
+    cl <- t(cl)
+  }
+  if (hp.row > hp.col) {
+    hp <- t(hp)
+  }
+  
+  # Check that inputs obey sign and value restrictions
+  if (sigma < 0) {
+    stop("Standard deviation must be non-negative")
+  }
+  if (df < 3) {
+    stop("Number of degrees of freedom must be at least 3 for first two moments of distribution to be defined")
+  }
+  if (max(cl) >= 1){
+    stop("Confidence level(s) must be less than 1")
+  }
+  if (min(cl) <= 0){
+    stop("Confidence level(s) must be greater than 0")
+  }
+  if (min(hp) <= 0){
+    stop("Holding Period(s) must be greater than 0")
+  }
+  
+  # Message to indicate how matrix of results is to be interpreted, if cl and hp both vary and results are given in matrix form
+  if (max(cl.row, cl.col) > 1 & max(hp.row, hp.col) > 1) {
+    print('VaR results with confidence level varying across row and holding period down column')
+  }
+  
+  # VaR estimation
+  cl.row <- dim(cl)[1]
+  cl.col <- dim(cl)[2]
+  VaR <- - sigma[1,1] * sqrt(hp) * sqrt((df - 2) / df) %*% qt(1 - cl, df)  - mu[1,1] * hp %*% matrix(1, cl.row, cl.col) # VaR
+  
+  # ES Estimation
+  n <- 1000 # Number of slices into which tail is divided
+  cl0 <- cl # Initial confidence level
+  w <- VaR
+  delta.cl <- (1 - cl) / n # Increment to confidence level as each slice is taken
+  for (i in 1:(n-1)) {
+    cl <- cl0 + i * delta.cl
+    w <- w - sigma[1,1] * sqrt(hp) * sqrt((df - 2) / df) %*% qt(1 - cl, df)  - mu[1,1] * hp %*% matrix(1, cl.row, cl.col)
+  }
+  ES <- w/n
+  # Plotting
+  x.min <- -mu - 5 * sigma
+  x.max <- -mu + 5 * sigma
+  delta <- (x.max-x.min) / 100
+  x <- seq(x.min, x.max, delta)
+  p <- dt((x-mu) / sigma, df)
+  plot(x, p, type = "l", xlim = c(x.min, x.max), ylim = c(0, max(p)*1.1), xlab = "Loss (+) / Profit (-)", ylab = "Probability", main = "t- VaR and ES")
+  
+  # VaR line
+  u <- c(VaR, VaR)
+  v <- c(0, .6*max(p))
+  lines(u, v, type = "l", col = "blue")
+
+  # ES line
+  w <- c(ES, ES)
+  z <- c(0, .45*max(p))
+  lines(w, z, type = "l", col = "blue")
+  # Input Labels
+  cl.for.label <- 100 * cl0
+  xpos <- -mu-2.5*sigma
+  text(xpos,.95*max(p), pos = 1, 'Input parameters', cex=.75, font = 2)
+  text(xpos, .875*max(p),pos = 1, paste('Daily mean L/P = ', -mu), cex=.75)
+  text(xpos, .8*max(p),pos = 1, paste('St. dev. of daily L/P = ', sigma), cex=.75)
+  text(xpos, .725*max(p),pos = 1, paste('Degrees of freedom', df), cex=.75)
+  text(xpos, .65*max(p),pos = 1, paste('Holding period = ', hp,' day(s)'), cex=.75)
+  # VaR label
+  text(VaR, .7*max(p),pos = 2, paste('VaR at ', cl.for.label,'% CL'), cex=.75)
+  text(VaR, .65 * max(p),pos = 2, paste('= ',VaR), cex=.75)
+  
+  # ES label
+  text(ES, .55*max(p),pos = 2, 'ES =', cex=.75)
+  text(ES, .65 * max(p),pos = 2, paste(ES), cex=.75)
+  
+}



More information about the Returnanalytics-commits mailing list