A few years ago, I was working on a project where we measured various characteristics of a time series and used the information to determine what forecasting method to apply or how to cluster the time series into meaningful groups. The two main papers to come out of that project were:
- Wang, Smith and Hyndman (2006) Characteristic-based clustering for time series data. Data Mining and Knowledge Discovery, 13(3), 335–364.
- Wang, Smith-Miles and Hyndman (2009) “Rule induction for forecasting method selection: meta-learning the characteristics of univariate time series”, Neurocomputing, 72, 2581–2594.
I’ve since had a lot of requests for the code which one of my coauthors has been helpfully emailing to anyone who asked. But to make it easier, we thought it might be helpful if I post some updated code here. This is not the same as the R code we used in the paper, as I’ve improved it in several ways (so it will give different results). If you just want the code, skip to the bottom of the post.
Finding the period of the data
Usually in time series work, we know the period of the data (if the observations are monthly, the period is 12, for example). But in this project, some of our data was of unknown period and we wanted a method to automatically determine the appropriate period. The method we used was based on local peaks and troughs in the ACF. But I’ve since devised a better approach (prompted on crossvalidated.com) using an estimate of the spectral density:
find.freq <- function(x) { n <- length(x) spec <- spec.ar(c(na.contiguous(x)),plot=FALSE) if(max(spec$spec)>10) # Arbitrary threshold chosen by trial and error. { period <- round(1/spec$freq[which.max(spec$spec)]) if(period==Inf) # Find next local maximum { j <- which(diff(spec$spec)>0) if(length(j)>0) { nextmax <- j[1] + which.max(spec$spec[j[1]:500]) if(nextmax <= length(spec$freq)) period <- round(1/spec$freq[nextmax]) else period <- 1 } else period <- 1 } } else period <- 1 return(period) } |
The function is called find.freq because time series people often call the period of seasonality the “frequency” (which is of course highly confusing).
Decomposing the data into trend and seasonal components
We needed a measure of the strength of trend and the strength of seasonality, and to do this we decomposed the data into trend, seasonal and error terms.
Because not all data could be decomposed additively, we first needed to apply an automated Box-Cox transformation. We tried a range of Box-Cox parameters on a grid, and selected the one which gave the most normal errors. That worked ok, but I’ve since found some papers that provide quite good automated Box-Cox algorithms that I’ve implemented in the forecast package. So this code uses Guerrero’s (1993) method instead.
For seasonal time series, we decomposed the transformed data using an stl decomposition with periodic seasonality.
For non-seasonal time series, we estimated the trend of the transformed data using penalized regression splines via the mgcv package.
decomp <- function(x,transform=TRUE) { require(forecast) # Transform series if(transform & min(x,na.rm=TRUE) >= 0) { lambda <- BoxCox.lambda(na.contiguous(x)) x <- BoxCox(x,lambda) } else { lambda <- NULL transform <- FALSE } # Seasonal data if(frequency(x)>1) { x.stl <- stl(x,s.window="periodic",na.action=na.contiguous) trend <- x.stl$time.series[,2] season <- x.stl$time.series[,1] remainder <- x - trend - season } else #Nonseasonal data { require(mgcv) tt <- 1:length(x) trend <- rep(NA,length(x)) trend[!is.na(x)] <- fitted(gam(x ~ s(tt))) season <- NULL remainder <- x - trend } return(list(x=x,trend=trend,season=season,remainder=remainder, transform=transform,lambda=lambda)) } |
Putting everything on a [0,1] scale
We wanted to measure a range of characteristics such as strength of seasonality, strength of trend, level of nonlinearity, skewness, kurtosis, serial correlatedness, self-similarity, level of chaoticity (is that a word?) and the periodicity of the data. But we wanted all these on the same scale which meant mapping the natural range of each measure onto [0,1]. The following two functions were used to do this.
# f1 maps [0,infinity) to [0,1] f1 <- function(x,a,b) { eax <- exp(a*x) if (eax == Inf) f1eax <- 1 else f1eax <- (eax-1)/(eax+b) return(f1eax) } # f2 maps [0,1] onto [0,1] f2 <- function(x,a,b) { eax <- exp(a*x) ea <- exp(a) return((eax-1)/(eax+b)*(ea+b)/(ea-1)) } |
The values of 
and 
in each function were chosen so the measure had a 90th percentile of 0.10 when the data were iid standard normal, and a value of 0.9 using a well-known benchmark time series.
Calculating the measures
Now we are ready to calculate the measures on the original data, as well as on the adjusted data (after removing trend and seasonality).
measures <- function(x) { require(forecast) N <- length(x) freq <- find.freq(x) fx <- c(frequency=(exp((freq-1)/50)-1)/(1+exp((freq-1)/50))) x <- ts(x,f=freq) # Decomposition decomp.x <- decomp(x) # Adjust data if(freq > 1) fits <- decomp.x$trend + decomp.x$season else # Nonseasonal data fits <- decomp.x$trend adj.x <- decomp.x$x - fits + mean(decomp.x$trend, na.rm=TRUE) # Backtransformation of adjusted data if(decomp.x$transform) tadj.x <- InvBoxCox(adj.x,decomp.x$lambda) else tadj.x <- adj.x # Trend and seasonal measures v.adj <- var(adj.x, na.rm=TRUE) if(freq > 1) { detrend <- decomp.x$x - decomp.x$trend deseason <- decomp.x$x - decomp.x$season trend <- ifelse(var(deseason,na.rm=TRUE) < 1e-10, 0, max(0,min(1,1-v.adj/var(deseason,na.rm=TRUE)))) season <- ifelse(var(detrend,na.rm=TRUE) < 1e-10, 0, max(0,min(1,1-v.adj/var(detrend,na.rm=TRUE)))) } else #Nonseasonal data { trend <- ifelse(var(decomp.x$x,na.rm=TRUE) < 1e-10, 0, max(0,min(1,1-v.adj/var(decomp.x$x,na.rm=TRUE)))) season <- 0 } m <- c(fx,trend,season) # Measures on original data xbar <- mean(x,na.rm=TRUE) s <- sd(x,na.rm=TRUE) # Serial correlation Q <- Box.test(x,lag=10)$statistic/(N*10) fQ <- f2(Q,7.53,0.103) # Nonlinearity p <- terasvirta.test(na.contiguous(x))$statistic fp <- f1(p,0.069,2.304) # Skewness sk <- abs(mean((x-xbar)^3,na.rm=TRUE)/s^3) fs <- f1(sk,1.510,5.993) # Kurtosis k <- mean((x-xbar)^4,na.rm=TRUE)/s^4 fk <- f1(k,2.273,11567) # Hurst=d+0.5 where d is fractional difference. H <- fracdiff(na.contiguous(x),0,0)$d + 0.5 # Lyapunov Exponent if(freq > N-10) stop("Insufficient data") Ly <- numeric(N-freq) for(i in 1:(N-freq)) { idx <- order(abs(x[i] - x)) idx <- idx[idx < (N-freq)] j <- idx[2] Ly[i] <- log(abs((x[i+freq] - x[j+freq])/(x[i]-x[j])))/freq if(is.na(Ly[i]) | Ly[i]==Inf | Ly[i]==-Inf) Ly[i] <- NA } Lyap <- mean(Ly,na.rm=TRUE) fLyap <- exp(Lyap)/(1+exp(Lyap)) m <- c(m,fQ,fp,fs,fk,H,fLyap) # Measures on adjusted data xbar <- mean(tadj.x, na.rm=TRUE) s <- sd(tadj.x, na.rm=TRUE) # Serial Q <- Box.test(adj.x,lag=10)$statistic/(N*10) fQ <- f2(Q,7.53,0.103) # Nonlinearity p <- terasvirta.test(na.contiguous(adj.x))$statistic fp <- f1(p,0.069,2.304) # Skewness sk <- abs(mean((tadj.x-xbar)^3,na.rm=TRUE)/s^3) fs <- f1(sk,1.510,5.993) # Kurtosis k <- mean((tadj.x-xbar)^4,na.rm=TRUE)/s^4 fk <- f1(k,2.273,11567) m <- c(m,fQ,fp,fs,fk) names(m) <- c("frequency", "trend","seasonal", "autocorrelation","non-linear","skewness","kurtosis", "Hurst","Lyapunov", "dc autocorrelation","dc non-linear","dc skewness","dc kurtosis") return(m) } |
Here is a quick example applied to Australian monthly gas production:
library(forecast) measures(gas) frequency trend seasonal autocorrelation 0.1096 0.9989 0.9337 0.9985 non-linear skewness kurtosis Hurst 0.4947 0.1282 0.0055 0.9996 Lyapunov dc autocorrelation dc non-linear dc skewness 0.5662 0.1140 0.0538 0.1743 dc kurtosis 0.9992 |
The function is far from perfect, and it is not hard to find examples where it fails. For example, it doesn’t work with multiple seasonality — try measure(taylor) and check the seasonality. Also, I’m not convinced the kurtosis provides anything useful here, or that the skewness measure is done in the best way possible. But it was really a proof of concept, so we will leave it to others to revise and improve the code.
In our papers, we took the measures obtained using R, and produced self-organizing maps using Viscovery. There is now a som package in R for that, so it might be possible to integrate that step into R as well. The dendogram was generated in matlab, although that could now also be done in R using the ggdendro package for example.
Download the code in a single file.
Related Posts:
- Time Series Data Library now on DataMarket
- Flat forecasts
- Initializing the Holt-Winters method
- Cyclic and seasonal time series
- Batch forecasting in R
Pingback: Time Series Decomposition : Box Cox for Additive Decomp | Q&A System