R: Beveridge–Nelson Trend and Cycle Decomposition

This post shows how to extract trend and cyclical components from a univariate time series based on the Beveridge–Nelson (BN) decomposition using tsm R package.

Beveridge–Nelson decomposition

Beveridge and Nelson (1981) provide an way of decomposing a non-stationary time series (\(y_t\)) into a permanent and a temporary (cyclical) component by applying ARIMA methods. \[\begin{align} y_t = \tau_t + c_t \end{align}\] where \(\tau_t\) and \(c_t\) are trend and cyclical components respectively.

The trend is modeled as a random walk with drift, and the cycle is treated as stationary process with zero mean.

Let \(y_t\) be integrated of first order, so that its first difference, \(\Delta y_t\) is stationary and has the following moving average representation \[\begin{align} (1-L)y_t = \Delta y_t = \mu + B(L)\epsilon_t \end{align}\] where \(L\) is the lag operator and \(B(L)\) is a polynomial of order \(q\) in the lag operator. \[\begin{align} L X_t = X_{t-1}, L^2 X_t = X_{t-2}, ... \\ B(L) =1 + B_1 L + B_2 L^2 + ... +B_q L^q \end{align}\]
Consider now the polynomial \(B(L)-B(1)\). Given that \(B(1)\) is just a constant, \(B(L)-B(1)\) will be of order \(q\). As \(1\) is a root of \(B(L)-B(1)\) since \(B(1)-B(1)=0\), the following expression can be obtained. \[\begin{align} B(L)-B(1) = B^*(L)(1-L) \\ \rightarrow B^*(L) = (1-L)^{-1}[B(L)-B(1)] \end{align}\] where \(B^*(L)\) is a polynomial of order \(q-1\) and \(B(1) = \sum_{s=0}^{\infty} B_s\).

Using the above derivations, we can express \(\Delta y_t\) as \[\begin{align} \Delta y_t &= \mu + [B(1) + (1-L)B^*(L)]\epsilon_t \\ &= \mu + B(1)\epsilon_t + (1-L)B^*(L)\epsilon_t \end{align}\] As \(y_t = \tau_t + c_t\), \(\Delta y_t = \Delta \tau_t + \Delta c_t\). Hence, changes in the trend and cycle components of \(y_t\) are respectively, \[\begin{align} \Delta \tau_t &= \mu + B(1)\epsilon_t \\ \Delta c_t &= (1-L)B^*(L)\epsilon_t \end{align}\] As the trend follows a random walk with drift, This expression can be solved to yield. \[\begin{align} \tau_t &= \underbrace{\tau_0 + \mu t}_{\text{deterministic term}} + \underbrace{B(1)\sum_{s=1}^{t} \epsilon_s }_{\text{stochastic term}} \\ c_t &= B^*(L)\epsilon_t = (1-L)^{-1}[B(L)-B(1)]\epsilon_t \end{align}\]

R code

The Beveridge-Nelson decomposition is easily estimated by using tsm R package. You can find the installation instruction and working examples at https://kevinkotze.github.io/ts-5-tut/.

#========================================================# # Quantitative ALM, Financial Econometrics & Derivatives  # ML/DL using R, Python, Tensorflow by Sang-Heon Lee  # # https://shleeai.blogspot.com #----------------------------------------------------------- # Beveridge-Nelson decomposition #========================================================#   graphics.off(); rm(list = ls())   #devtools::install_github("KevinKotze/tsm") library(tsm) #library(mFilter)   #----------------------------------------------------------- # U.S. Real GDP data from Greene Text book #-----------------------------------------------------------   # read data file st <- c(1950, 1); fq <- 4 fn <- "http://people.stern.nyu.edu/wgreene/Text/Edition7/TableF5-2.txt" macro <- read.csv(fn, sep="", header = TRUE) gdp <- macro$realgdp   # logarithm of real gdp multiplied by 100 lngdp <- ts(log(gdp)*100, start = st, frequency = fq)   #----------------------------------------------------------- # Beveridge-Nelson decomposition function arguments #----------------------------------------------------------- # bnd(data, nlag) # data : A vector of first-order  #        integrated numeric values # nlag : A numberic scalar for the lag-order  #        of the autoregressive (cyclical) part #-----------------------------------------------------------   # apply the BN decomposition bn.decomp <- bnd(lngdp, nlag = 2)   # extract the BN trend and cycle bn.trend <- ts(bn.decomp[, 1], start = st, frequency = fq) bn.cycle <- ts(bn.decomp[, 2], start = st, frequency = fq)   # graph x11(width = 7, height = 7);  par(mfrow = c(2,1), mar = c(2.2, 2.2, 1, 1), cex = 1) plot.ts(lngdp, ylab = "", col = "blue", lwd=2) lines(bn.trend, col = "red", lwd=2) legend("topleft", legend = c("log(GDP)", "Trend(BN)"),         lty = 1, col = c("blue", "red"), bty = "n", lwd=3) plot.ts(bn.cycle, ylab = "", ylim = c(-2.5,2.5), lwd=3, col = 4) legend("topright", legend = c("Cycle(BN,%)"),         lty = 1, col = 4, bty = "n", lwd=3)   Colored by Color Scripter

cs

We can get the following the outputs (trend and cycle) of the BN decomposition.

Reference

Beveridge, S. and C. R. Nelson (1981). A new approach to decomposition of economic time series into permanent and transitory components with particular attention to measurement of the `business cycle', Journal of Monetary Economics 7-2, 151-174. \(\blacksquare\)