| Title: | Time Series Exploration, Modelling and Forecasting |
|---|---|
| Description: | Includes: (i) tests and visualisations that can help the modeller explore time series components and perform decomposition; (ii) modelling shortcuts, such as functions to construct lagmatrices and seasonal dummy variables of various forms; (iii) an implementation of the Theta method; (iv) tools to facilitate the design of the forecasting process, such as ABC-XYZ analyses; and (v) "quality of life" functions, such as treating time series for trailing and leading values. |
| Authors: | Nikolaos Kourentzes [aut, cre], Ivan Svetunkov [ctb], Oliver Schaer [ctb] |
| Maintainer: | Nikolaos Kourentzes <[email protected]> |
| License: | GPL-3 |
| Version: | 0.9.5 |
| Built: | 2026-11-07 18:04:50 UTC |
| Source: | https://github.com/trnnick/tsutils |
Perform ABC analysis on a set of time series.
abc(x, prc = c(0.2, 0.3, 0.5)) ## S3 method for class 'abc' plot(x, cex.prc = 0.8, ...)abc(x, prc = c(0.2, 0.3, 0.5)) ## S3 method for class 'abc' plot(x, cex.prc = 0.8, ...)
x |
this can either be an array, where each column is a series, or a vector of values. If |
prc |
a vector of percentages indicating how many items are included in each class. By default this is |
cex.prc |
font size of percentages reported in plot. |
... |
additional arguments passed to the plot. |
Return object of class abc and contains:
value: a vector containing the importance value of each series.
class: a vector containing the class membership of each series.
rank: a vector containing the rank of each series, with 1 being the highest ranking series.
conc: the importance concentration of each class, as percentage of total value.
plot(abc): plot ABC or XYZ analyses.
Nikolaos Kourentzes, [email protected].
Ord K., Fildes R., Kourentzes N. (2017) Principles of Business Forecasting, 2e. Wessex Press Publishing Co., p.515-518.
x <- abs(matrix(cumsum(rnorm(5400,0,1)),36,150)) z <- abc(x) print(z) plot(z)x <- abs(matrix(cumsum(rnorm(5400,0,1)),36,150)) z <- abc(x) print(z) plot(z)
Jointly visualise ABC and XYZ analyses.
abcxyz(imp, frc, outplot = c(TRUE, FALSE), error = NULL, ...)abcxyz(imp, frc, outplot = c(TRUE, FALSE), error = NULL, ...)
imp |
an obkect of class |
frc |
an obkect of class |
outplot |
if |
error |
vector of forecast errors for each series that will be distributed in each class, presented as an average. |
... |
additional arguments passed to the plot. |
A list containing:
class: a matrix containing the number of time series in each class.
error: a matrix containing the averaged error for each class, if the argument error was used.
Nikolaos Kourentzes, [email protected].
Ord K., Fildes R., Kourentzes N. (2017) Principles of Business Forecasting, 2e. Wessex Press Publishing Co., p.515-518.
x <- abs(matrix(cumsum(rnorm(5400,0,1)),36,150)) abcxyz(abc(x),xyz(x,type="cv"))x <- abs(matrix(cumsum(rnorm(5400,0,1)),36,150)) abcxyz(abc(x),xyz(x,type="cv"))
Calculate the Centred Moving Average (CMA) for time series.
cmav( y, ma = NULL, fill = c(TRUE, FALSE), outplot = c(FALSE, TRUE), fast = c(TRUE, FALSE) )cmav( y, ma = NULL, fill = c(TRUE, FALSE), outplot = c(FALSE, TRUE), fast = c(TRUE, FALSE) )
y |
input time series. Can be |
ma |
length of centred moving average. If |
fill |
if |
outplot |
if |
fast |
if |
Centred moving average. If y is a ts object, then cma has the same properties.
Nikolaos Kourentzes, [email protected].
Ord K., Fildes R., Kourentzes N. (2017) Principles of Business Forecasting, 2e. Wessex Press Publishing Co., p.109.
cmav(referrals,outplot=TRUE)cmav(referrals,outplot=TRUE)
Produce the opaque colour that looks like a semi-transparent colour drawn over a background, by mixing the two colours. This is useful for graphics devices that do not support semi-transparency, such as postscript (EPS), where semi-transparent elements are not drawn.
colour.blend(col, alpha, bg = NULL)colour.blend(col, alpha, bg = NULL)
col |
colours to blend. Any form accepted by |
alpha |
opacity of |
bg |
background colour. If |
Each colour channel is mixed as alpha * col + (1 - alpha) * bg, which is how a semi-transparent colour is drawn over an opaque background. The result is identical to using adjustcolor(col, alpha.f=alpha) on that background, but overlapping elements do not add up: the colour drawn last covers the ones below it. If the plot is placed on a different background afterwards, the blended colours keep the tint of bg.
A character vector of opaque colours in hexadecimal format.
Nikolaos Kourentzes, [email protected]
colour.blend("red", 0.25) colour.blend("steelblue", c(0.2, 0.5, 0.8)) colour.blend(c("red", "blue"), 0.5, bg="black") plot(NA, xlim=c(0,1), ylim=c(0,1)) rect(0.1, 0.1, 0.9, 0.9, col=colour.blend("steelblue", 0.3))colour.blend("red", 0.25) colour.blend("steelblue", c(0.2, 0.5, 0.8)) colour.blend(c("red", "blue"), 0.5, bg="black") plot(NA, xlim=c(0,1), ylim=c(0,1)) rect(0.1, 0.1, 0.9, 0.9, col=colour.blend("steelblue", 0.3))
Perform Cox-Stuart test for location or dispersion.
coxstuart(y, type = c("trend", "deviation", "dispersion"), alpha = 0.05)coxstuart(y, type = c("trend", "deviation", "dispersion"), alpha = 0.05)
y |
input data. |
type |
type of test. Can be:
|
alpha |
significance level. |
A list containing:
H: hypothesis outcome.
p.value: corresponding p-value.
Htxt: textual description of the hypothesis outcome.
Nikolaos Kourentzes, [email protected].
coxstuart(referrals)coxstuart(referrals)
Perform classical time series decomposition. Optionally, the seasonal component is only estimated when there is evidence of seasonality.
decomp( y, m = NULL, s = NULL, trend = NULL, outplot = c(FALSE, TRUE), decomposition = c("multiplicative", "additive", "auto"), h = 0, type = c("mean", "median", "pure.seasonal"), w = NULL, seastest = c(FALSE, TRUE), alpha = 0.05, ... )decomp( y, m = NULL, s = NULL, trend = NULL, outplot = c(FALSE, TRUE), decomposition = c("multiplicative", "additive", "auto"), h = 0, type = c("mean", "median", "pure.seasonal"), w = NULL, seastest = c(FALSE, TRUE), alpha = 0.05, ... )
y |
input time series. Can be |
m |
seasonal period. If |
s |
starting period in the season. If |
trend |
vector of the level/trend of |
outplot |
if |
decomposition |
type of decomposition. This can be |
h |
forecast horizon for seasonal component. |
type |
calculation for seasonal component:
|
w |
percentage or number of observations to winsorise in the calculation of mean seasonal indices. If w>1 then it is the number of observations, otherwise it is a percentage. If |
seastest |
if |
alpha |
significance level for the seasonality test. Used only if |
... |
additional arguments to control the plot:
|
A list containing:
trend: trend component.
season: season component.
irregular: irregular component.
f.season: forecasted seasonal component if h>0.
g: pure seasonal model parameters.
decomposition: type of decomposition used.
season.exist: TRUE/FALSE result of the seasonality test, or NULL if not tested.
season.pval: p-value of the seasonality test, or NULL if not tested.
Nikolaos Kourentzes, [email protected].
Ord K., Fildes R., Kourentzes N. (2017) Principles of Business Forecasting, 2e. Wessex Press Publishing Co., p.106-111.
decomp(referrals) decomp(referrals,outplot=TRUE,h=12,seastest=TRUE)decomp(referrals) decomp(referrals,outplot=TRUE,h=12,seastest=TRUE)
Quarterly US real gross domestic product and its expenditure components in chained 2017 dollars, forming a hierarchy that is not coherent: chain-linked volumes are not additive, so each aggregate differs from the sum of its components.
A list with:
y: quarterly ts matrix of 26 series from 2007Q1 to 2026Q2 (78 quarters), in billions of chained 2017 dollars, seasonally adjusted annual rate. Columns are ordered as the rows of S.
S: summing matrix () of the expenditure-side hierarchy, following NIPA Table 1.1.6: 11 aggregates and 15 bottom level series. Imports enter net exports, and hence GDP, with -1.
levels: level of each row of S in the hierarchy, for use with plotS.
codes: FRED series code of each row of S.
titles: FRED series title of each row of S.
The hierarchy is: GDP = PCE + GPDI + NetEx + Gov; personal consumption expenditures, PCE = Goods + Services, with Goods = Durable + Nondurable; gross private domestic investment, GPDI = FPI + CBI, with fixed investment FPI = PNFI + Residential and nonresidential PNFI = Structures + Equipment + IPP; net exports, NetEx = Exports - Imports, each split into goods and services; government, Gov = Federal + StateLocal, with Federal = Defense + Nondefense.
Because of chain-linking, the observed series do not satisfy y = B %*% t(S), with B the bottom level series. The discrepancy of GDP against the bottom level equals, up to rounding, the residual published by the Bureau of Economic Analysis (FRED code A960RX1Q020SBEA). Net exports are computed as exports minus imports, so their discrepancy is implied by those of exports and imports. Use Saugment to obtain a coherent augmented hierarchy. The data are rounded to three decimals.
The data are the vintage of 30 September 2026. National accounts are revised, and comprehensive revisions change the reference year, so later vintages differ, including the discrepancies.
Bureau of Economic Analysis, retrieved from FRED, Federal Reserve Bank of St. Louis, https://fred.stlouisfed.org.
Kourentzes, N., et al. (in preparation). Forecast reconciliation under observational incoherence. Working paper.
S <- gdpchain$S y <- gdpchain$y cols <- RColorBrewer::brewer.pal(3,"Pastel1") # Hierarchy, highlighting the series that enter with -1 (imports) neg <- rownames(S) %in% colnames(S)[apply(S < 0,2,any)] plotS(S,levels=gdpchain$levels,fill=ifelse(neg,cols[1],"grey88")) # Produce plot similar to reference paper plotS(S,levels=gdpchain$levels,fill=ifelse(neg,cols[1],"grey88"), view="block",skip="expand",vgap=0.1) # Augment with discrepancy series. The data are rounded to three decimals. aug <- Saugment(S,y,tol=0.01) aug$status aug$error # Augmented hierarchy, highlighting the discrepancy series. Discrepancies are # measured against the bottom level, so they are not nested, and the block # view is clearer than the tree. k <- ncol(aug$d) lev <- c(gdpchain$levels,rep(max(gdpchain$levels),k)) neg <- rownames(aug$S) %in% colnames(aug$S)[apply(aug$S < 0,2,any)] fill <- ifelse(neg,cols[1],"grey88") fill[nrow(S)+(1:k)] <- cols[2] plotS(aug$S,levels=lev,fill=fill,view="block",skip="expand",vgap=0.1) # Discrepancy series: close to zero around the 2017 reference year plot(aug$d,main="")S <- gdpchain$S y <- gdpchain$y cols <- RColorBrewer::brewer.pal(3,"Pastel1") # Hierarchy, highlighting the series that enter with -1 (imports) neg <- rownames(S) %in% colnames(S)[apply(S < 0,2,any)] plotS(S,levels=gdpchain$levels,fill=ifelse(neg,cols[1],"grey88")) # Produce plot similar to reference paper plotS(S,levels=gdpchain$levels,fill=ifelse(neg,cols[1],"grey88"), view="block",skip="expand",vgap=0.1) # Augment with discrepancy series. The data are rounded to three decimals. aug <- Saugment(S,y,tol=0.01) aug$status aug$error # Augmented hierarchy, highlighting the discrepancy series. Discrepancies are # measured against the bottom level, so they are not nested, and the block # view is clearer than the tree. k <- ncol(aug$d) lev <- c(gdpchain$levels,rep(max(gdpchain$levels),k)) neg <- rownames(aug$S) %in% colnames(aug$S)[apply(aug$S < 0,2,any)] fill <- ifelse(neg,cols[1],"grey88") fill[nrow(S)+(1:k)] <- cols[2] plotS(aug$S,levels=lev,fill=fill,view="block",skip="expand",vgap=0.1) # Discrepancy series: close to zero around the 2017 reference year plot(aug$d,main="")
Calculate the geometric mean.
geomean(x, na.rm = c(FALSE, TRUE), ...)geomean(x, na.rm = c(FALSE, TRUE), ...)
x |
input data (will be considered as a vector). |
na.rm |
a logical value indicating whether NA values should be stripped before the computation proceeds. |
... |
further arguments passed to or from other methods. |
The geometirc mean of the values in x.
Nikolaos Kourentzes, [email protected].
geomean(c(0.5,1,1.5))geomean(c(0.5,1,1.5))
Calculate the theoretically optimal temporal aggregation level for AR(1), MA(1) and ARMA(1,1) time series.
getOptK(y, m = 12, type = c("ar", "ma", "arma"))getOptK(y, m = 12, type = c("ar", "ma", "arma"))
y |
a time series that must be of either |
m |
maximum aggregation level. |
type |
type of data generating process. Can be:
|
Identified optimal temporal aggregation level.
Nikolaos Kourentzes, [email protected].
Kourentzes, N., Rostami-Tabar, B., & Barrow, D. K. (2017). Demand forecasting by temporal aggregation: using optimal or multiple aggregation levels?. Journal of Business Research, 78, 1-9.
Rostami-Tabar, B., Babai, M. Z., Syntetos, A., & Ducq, Y. (2013). Demand forecasting by temporal aggregation. Naval Research Logistics (NRL), 60(6), 479-498.
Rostami-Tabar, B., Babai, M. Z., Syntetos, A., & Ducq, Y. (2014). A note on the forecast performance of temporal aggregation. Naval Research Logistics (NRL), 61(7), 489-500.
getOptK(referrals)getOptK(referrals)
Create an array with lead/lags of an input variable.
lagmatrix(x, lag)lagmatrix(x, lag)
x |
input variable. |
lag |
vector of leads and lags. Positive numbers are lags, negative are leads. O is the original |
An array with the resulting leads and lags (columns).
Nikolaos Kourentzes, [email protected]
x <- rnorm(10) lagmatrix(x,c(0,1,-1))x <- rnorm(10) lagmatrix(x,c(0,1,-1))
Calculates the lambdaMax value, which is the penalty term (lambda) beyond which coefficients are guaranteed to be all zero and provides a sequence of nLambda values to lambdaMin in logarithmic descent.
lambdaseq( x, y, weight = NA, alpha = 1, standardise = TRUE, lambdaRatio = 1e-04, nLambda = 100, addZeroLambda = FALSE )lambdaseq( x, y, weight = NA, alpha = 1, standardise = TRUE, lambdaRatio = 1e-04, nLambda = 100, addZeroLambda = FALSE )
x |
matrix of regressors. See |
y |
response variable. See |
weight |
vector of |
alpha |
elastic net mixing value. See |
standardise |
if |
lambdaRatio |
ratio between |
nLambda |
length of the lambda sequence. |
addZeroLambda |
if |
A list that contains:
lambda: sequence of lambda values, from lambdaMax to lambdaMin.
lambdaMin: minimal lambda value.
lambdaMax: maximal lambda value.
nullMSE: MSE of the fit using just a constant term.
Oliver Schaer, [email protected],
Nikolaos Kourentzes, [email protected].
Hastie, T., Tibshirani, R., & Wainwright, M. (2015). Statistical learning with sparsity: the lasso and generalizations. CRC press.
y <- mtcars[,1] x <- as.matrix(mtcars[,2:11]) lambda <- lambdaseq(x, y)$lambda ## Not run: library(glmnet) fit.lasso <- cv.glmnet(x, y, lambda = lambda) coef.lasso <- coef(fit.lasso, s = "lambda.1se") ## End(Not run)y <- mtcars[,1] x <- as.matrix(mtcars[,2:11]) lambda <- lambdaseq(x, y)$lambda ## Not run: library(glmnet) fit.lasso <- cv.glmnet(x, y, lambda = lambda) coef.lasso <- coef(fit.lasso, s = "lambda.1se") ## End(Not run)
Remove leading or trailing zeros or NAs from a vector.
leadtrail( x, rm = c("zeros", "na"), lead = c(TRUE, FALSE), trail = c(TRUE, FALSE) )leadtrail( x, rm = c("zeros", "na"), lead = c(TRUE, FALSE), trail = c(TRUE, FALSE) )
x |
vector of values to check. |
rm |
what to remove, can be |
lead |
If |
trail |
If |
Resulting vector.
Nikolaos Kourentzes, [email protected].
x <- c(rep(0,5),rnorm(100),rep(0,5)) leadtrail(x)x <- c(rep(0,5),rnorm(100),rep(0,5)) leadtrail(x)
Perform nonparametric multiple comparisons, across columns, using the Friedman and the post-hoc Nemenyi tests.
nemenyi( data, conf.level = 0.95, sort = c(TRUE, FALSE), plottype = c("vline", "none", "mcb", "vmcb", "line", "matrix", "lmatrix", "ranges"), select = NULL, labels = NULL, group = NULL, line = c(main = NA, xlab = NA, ylab = NA, xaxis = NA, yaxis = NA, group = NA), ... )nemenyi( data, conf.level = 0.95, sort = c(TRUE, FALSE), plottype = c("vline", "none", "mcb", "vmcb", "line", "matrix", "lmatrix", "ranges"), select = NULL, labels = NULL, group = NULL, line = c(main = NA, xlab = NA, ylab = NA, xaxis = NA, yaxis = NA, group = NA), ... )
data |
an array that includes values to be compared for several treatments (in columns) for several observations (rows), of size n x k. For example, if these are forecast errors, different methods should be in columns and errors for different time series or forecast origins in rows. |
conf.level |
the confidence level used for the comparison. Default is 0.95. |
sort |
if |
plottype |
type of plot to produce:
|
select |
highlight selected treatment (column). Number 1 to k. Use NULL for no highlighting. |
labels |
optional labels for models. If NULL column names of |
group |
if |
line |
margin lines (as in the
Elements that are |
... |
additional arguments passed to the |
Plots respect the graphical parameters currently set with par() (for example mfrow, cex, mgp, fg, bg, family or las for numerical axes). The only parameter changed is mar: margins that are too small for the treatment labels, the group letters or the lines requested in line are enlarged, using the actual width of the labels, while all other margins are kept as set by the user. The user's mar is restored when the function exits. As the plot region depends on the margins, low-level additions to the plot after the call (e.g. abline) may not align; set par(mar=...) large enough for the labels beforehand to avoid this.
Return object of class nemenyi and contains:
means: mean rank of each treatment.
intervals: intervals within there is no evidence of significance difference according to the Nemenyi test at requested confidence level.
fpval: Friedman test p-value.
fH: Friedman test hypothesis outcome.
cd: Nemenyi critical distance. Output intervals is calculate as means +/- cd.
conf.level: confidence level used for testing.
k: number of treatments (columns).
n: number of observations (rows).
group: group membership as a compact letter display; treatments sharing a letter are not significantly different.
diff: k x k matrix of absolute differences in mean ranks. Pairs with a difference smaller than cd are not significantly different.
Nikolaos Kourentzes, [email protected],
Ivan Svetunkov, [email protected].
The tests are detailed by Hollander, M., Wolfe, D.A. and Chicken, E. (2014) Nonparametric Statistical Methods. 3rd Edition, John Wiley & Sons, Inc., New York.
The line plot is introduced here and a first example of its use, along with a short description is provided by Kourentzes, N. (2013). Intermittent demand forecasts with neural networks. International Journal of Production Economics, 143(1), 198-206.
The matrix plot is introduced by Kourentzes, N., & Athanasopoulos, G. (2018). Cross-temporal coherent forecasts for Australian tourism (No. 24/18). Monash University, Department of Econometrics and Business Statistics.
The MCB plot is described by Koning, A. J., Franses, P. H., Hibon, M., & Stekler, H. O. (2005). The M3 competition: Statistical tests of the results. International Journal of Forecasting, 21(3), 397-409.
x <- matrix( rnorm(50*4,mean=0,sd=1), 50, 4) x[,2] <- x[,2]+1 x[,3] <- x[,3]+0.7 x[,4] <- x[,4]+0.5 colnames(x) <- c("Method A","Method B","Method C - long name","Method D") nemenyi(x,conf.level=0.95,plottype="vline") nemenyi(x,plottype="matrix") nemenyi(x,plottype="lmatrix") nemenyi(x,plottype="ranges") nemenyi(x,plottype="line",ylab="Groups",line=c(main=0.5,group=0.2))x <- matrix( rnorm(50*4,mean=0,sd=1), 50, 4) x[,2] <- x[,2]+1 x[,3] <- x[,3]+0.7 x[,4] <- x[,4]+0.5 colnames(x) <- c("Method A","Method B","Method C - long name","Method D") nemenyi(x,conf.level=0.95,plottype="vline") nemenyi(x,plottype="matrix") nemenyi(x,plottype="lmatrix") nemenyi(x,plottype="ranges") nemenyi(x,plottype="line",ylab="Groups",line=c(main=0.5,group=0.2))
Shows how the grouping of treatments by the Nemenyi test changes with the confidence level. As the test uses a single critical distance for all pairs, each pair of treatments changes from significantly different to not significantly different at exactly one confidence level, given by its pairwise p-value. The groupings therefore change only at these levels and are constant in between, so all distinct groupings can be found exactly.
nemenyi.sens( data, conf.level = NULL, conf.range = c(0.8, 0.999), ref = NULL, plottype = c("groups", "letters", "pvalue", "none"), labels = NULL, line = c(main = NA, xlab = NA, ylab = NA, xaxis = NA, yaxis = NA, group = NA), ... )nemenyi.sens( data, conf.level = NULL, conf.range = c(0.8, 0.999), ref = NULL, plottype = c("groups", "letters", "pvalue", "none"), labels = NULL, line = c(main = NA, xlab = NA, ylab = NA, xaxis = NA, yaxis = NA, group = NA), ... )
data |
an array that includes values to be compared for several treatments (in columns) for several observations (rows), of size n x k. See |
conf.level |
confidence levels to show. If |
conf.range |
range of confidence levels searched for distinct groupings, when |
ref |
optional reference confidence level, e.g. |
plottype |
type of plot to produce:
Changes from the previous column are highlighted. In |
labels |
optional labels for models. If NULL column names of |
line |
margin lines where the plot annotation is placed, see |
... |
additional arguments passed to the |
The pairwise p-value of treatments i and j is 1 - ptukey(|R_i - R_j| / sqrt(k(k+1)/(12n)), k, Inf), where R are the mean ranks. The pair is not significantly different at confidence levels above 1 - p, which is the same outcome as nemenyi at that level. Graphical parameters and colours are handled as in nemenyi.
Return object of class nemenyi.sens and contains:
means: mean rank of each treatment.
pval: k x k matrix of pairwise Nemenyi p-values.
pairs: table of all pairs, ordered by p-value, with the absolute difference in mean ranks, the p-value and the confidence level above which the pair is not significantly different.
fpval: Friedman test p-value. The Friedman test is significant for confidence levels below 1 - fpval.
levels: confidence levels shown.
range: range of confidence levels that each shown level represents. For user specified levels this is the level itself.
cd: Nemenyi critical distance at each shown level.
group: k x levels matrix of compact letter displays.
changed: list of two k x levels logical matrices, TRUE where a treatment changed from the previous level: groups, if its set of not significantly different treatments changed (highlighted in the "groups" plot), and letters, if its letters changed (highlighted in the "letters" plot).
k: number of treatments (columns).
n: number of observations (rows).
Nikolaos Kourentzes, [email protected]
x <- matrix( rnorm(50*4,mean=0,sd=1), 50, 4) x[,2] <- x[,2]+1 x[,3] <- x[,3]+0.7 x[,4] <- x[,4]+0.5 colnames(x) <- c("Method A","Method B","Method C - long name","Method D") nemenyi.sens(x) nemenyi.sens(x,plottype="letters") nemenyi.sens(x,plottype="pvalue",ref=0.95) nemenyi.sens(x,conf.level=c(0.8,0.9,0.95,0.99),ref=0.95)x <- matrix( rnorm(50*4,mean=0,sd=1), 50, 4) x[,2] <- x[,2]+1 x[,3] <- x[,3]+0.7 x[,4] <- x[,4]+0.5 colnames(x) <- c("Method A","Method B","Method C - long name","Method D") nemenyi.sens(x) nemenyi.sens(x,plottype="letters") nemenyi.sens(x,plottype="pvalue",ref=0.95) nemenyi.sens(x,conf.level=c(0.8,0.9,0.95,0.99),ref=0.95)
Plots a hierarchy from its summing matrix S, either as a tree or as blocks. This is a convenience function that calls plotStree or plotSblock. These can also be used directly, and their help files describe all the options available.
plotS( S, view = c("tree", "block"), labels = TRUE, parse = c(FALSE, TRUE), skip = c("empty", "expand"), levels = NULL, fill = "grey88", border = "grey30", level.labels = FALSE, level.font = 1, main = NULL, ... )plotS( S, view = c("tree", "block"), labels = TRUE, parse = c(FALSE, TRUE), skip = c("empty", "expand"), levels = NULL, fill = "grey88", border = "grey30", level.labels = FALSE, level.font = 1, main = NULL, ... )
S |
summing matrix ( |
view |
type of plot:
|
labels |
|
parse |
if |
skip |
how to plot nodes that skip a level: |
levels |
optional vector with the level of each row of |
fill |
fill colour of the nodes. Either a single colour or one per row of |
border |
border colour of the nodes. |
level.labels |
|
level.font |
font of the level labels, as in |
main |
optional plot title. |
... |
additional arguments passed to |
Invisibly, the output of plotStree or plotSblock.
Nikolaos Kourentzes, [email protected].
plotStree, plotSblock, plotThief.
# Small hierarchy: A = H1 + H2, H1 = Q1 + Q2, H2 = Q3 + Q4 S <- rbind(c(1,1,1,1),c(1,1,0,0),c(0,0,1,1),diag(4)) rownames(S) <- c("A","H1","H2","Q1","Q2","Q3","Q4") plotS(S) plotS(S,view="block") # Pass additional arguments to plotStree plotS(S,edges="elbow",shape="circle",level.labels=TRUE)# Small hierarchy: A = H1 + H2, H1 = Q1 + Q2, H2 = Q3 + Q4 S <- rbind(c(1,1,1,1),c(1,1,0,0),c(0,0,1,1),diag(4)) rownames(S) <- c("A","H1","H2","Q1","Q2","Q3","Q4") plotS(S) plotS(S,view="block") # Pass additional arguments to plotStree plotS(S,edges="elbow",shape="circle",level.labels=TRUE)
Plots a hierarchy from its summing matrix S, with each node (row of S) shown as the block of bottom level series it aggregates. Levels are arranged from the top to the bottom of the hierarchy. The horizontal axis has one unit per bottom level series (columns of S). This view remains clean for large hierarchies. For the more traditional view use plotStree.
plotSblock( S, labels = TRUE, parse = c(FALSE, TRUE), skip = c("empty", "expand"), levels = NULL, fill = "grey88", border = "grey30", hgap = 0.08, vgap = 0.35, cex = 1, min.cex = 0.35, level.labels = FALSE, level.font = 1, main = NULL )plotSblock( S, labels = TRUE, parse = c(FALSE, TRUE), skip = c("empty", "expand"), levels = NULL, fill = "grey88", border = "grey30", hgap = 0.08, vgap = 0.35, cex = 1, min.cex = 0.35, level.labels = FALSE, level.font = 1, main = NULL )
S |
summing matrix ( |
labels |
|
parse |
if |
skip |
how to treat the space left by a node that skips a level:
|
levels |
optional vector with the level of each row of |
fill |
fill colour of the blocks. Either a single colour or one per row of |
border |
border colour of the blocks. |
hgap |
horizontal space between blocks, as a fraction of a column width. |
vgap |
vertical space between levels, as a fraction of the height of a level. |
cex |
maximum label size. |
min.cex |
labels that would need to be smaller than this to fit in their block are not plotted. |
level.labels |
|
level.font |
font of the level labels, as in |
main |
optional plot title. |
By default, rows are assigned to levels by scanning S from the top and starting a new level whenever a row overlaps with a row already in the current level. Therefore, rows within a level aggregate different bottom level series. This works with the usual ordering of S for cross-sectional, grouped and temporal hierarchies. For unusual row orders, provide the level of each row with levels.
In grouped hierarchies some nodes aggregate bottom level series that are not next to each other (e.g., a product node in a region x product grouping). These are plotted as several blocks, all with the same label and colour.
Invisibly, a data frame with one row per block and columns:
node: row name of S.
row: row index in S.
level: level of the node.
top: highest level reached, after any expansion.
first, last: first and last column of S covered by the block.
Nikolaos Kourentzes, [email protected].
For the design and explanation of this plot see arXiv:2609.23223.
# Helper to construct S from a list of aggregates makeS <- function(aggs,bottom){ A <- t(sapply(aggs,function(a){as.numeric(bottom %in% a)})) S <- rbind(A,diag(length(bottom))) dimnames(S) <- list(c(names(aggs),bottom),bottom) return(S) } # Simple hierarchy S1 <- makeS(list(A=c("Q1","Q2","Q3","Q4"),H1=c("Q1","Q2"),H2=c("Q3","Q4")), c("Q1","Q2","Q3","Q4")) plotSblock(S1) # Row names parsed as plotmath and level labels rownames(S1) <- c("A","H[1]","H[2]","Q[1]","Q[2]","Q[3]","Q[4]") plotSblock(S1,parse=TRUE,level.labels=expression(bold(S)[4],bold(S)[2],bold(S)[1])) # Labels mixing text and expressions plotSblock(S1,labels=list("Total",quote(H[1]),quote(H[2]),"Q1","Q2","Q3",quote(Q[4]^"*"))) # Q5 skips a level and connects directly to A S2 <- makeS(list(A=c("Q1","Q2","Q3","Q4","Q5"),H1=c("Q1","Q2"),H2=c("Q3","Q4")), c("Q1","Q2","Q3","Q4","Q5")) plotSblock(S2,skip="empty") plotSblock(S2,skip="expand") # Temporal hierarchy for monthly data plotSblock(Sthief(12),labels=FALSE,level.labels=paste("k =",c(12,6,4,3,2,1))) # Grouped hierarchy: region x product bottom <- c("R1P1","R1P2","R2P1","R2P2") S3 <- makeS(list(Total=bottom,R1=c("R1P1","R1P2"),R2=c("R2P1","R2P2"), P1=c("R1P1","R2P1"),P2=c("R1P2","R2P2")),bottom) cmp <- RColorBrewer::brewer.pal(5,"Pastel1") plotSblock(S3,fill=c("grey88",cmp[2],cmp[2],cmp[5],cmp[5],rep("grey88",4))) # Larger hierarchy without labels b <- paste0("b",1:60) S4 <- makeS(c(list(Total=b), setNames(lapply(0:4,function(g){b[g*12+1:12]}),paste0("G",1:5)), setNames(lapply(0:19,function(g){b[g*3+1:3]}),paste0("SG",1:20))),b) plotSblock(S4,labels=FALSE)# Helper to construct S from a list of aggregates makeS <- function(aggs,bottom){ A <- t(sapply(aggs,function(a){as.numeric(bottom %in% a)})) S <- rbind(A,diag(length(bottom))) dimnames(S) <- list(c(names(aggs),bottom),bottom) return(S) } # Simple hierarchy S1 <- makeS(list(A=c("Q1","Q2","Q3","Q4"),H1=c("Q1","Q2"),H2=c("Q3","Q4")), c("Q1","Q2","Q3","Q4")) plotSblock(S1) # Row names parsed as plotmath and level labels rownames(S1) <- c("A","H[1]","H[2]","Q[1]","Q[2]","Q[3]","Q[4]") plotSblock(S1,parse=TRUE,level.labels=expression(bold(S)[4],bold(S)[2],bold(S)[1])) # Labels mixing text and expressions plotSblock(S1,labels=list("Total",quote(H[1]),quote(H[2]),"Q1","Q2","Q3",quote(Q[4]^"*"))) # Q5 skips a level and connects directly to A S2 <- makeS(list(A=c("Q1","Q2","Q3","Q4","Q5"),H1=c("Q1","Q2"),H2=c("Q3","Q4")), c("Q1","Q2","Q3","Q4","Q5")) plotSblock(S2,skip="empty") plotSblock(S2,skip="expand") # Temporal hierarchy for monthly data plotSblock(Sthief(12),labels=FALSE,level.labels=paste("k =",c(12,6,4,3,2,1))) # Grouped hierarchy: region x product bottom <- c("R1P1","R1P2","R2P1","R2P2") S3 <- makeS(list(Total=bottom,R1=c("R1P1","R1P2"),R2=c("R2P1","R2P2"), P1=c("R1P1","R2P1"),P2=c("R1P2","R2P2")),bottom) cmp <- RColorBrewer::brewer.pal(5,"Pastel1") plotSblock(S3,fill=c("grey88",cmp[2],cmp[2],cmp[5],cmp[5],rep("grey88",4))) # Larger hierarchy without labels b <- paste0("b",1:60) S4 <- makeS(c(list(Total=b), setNames(lapply(0:4,function(g){b[g*12+1:12]}),paste0("G",1:5)), setNames(lapply(0:19,function(g){b[g*3+1:3]}),paste0("SG",1:20))),b) plotSblock(S4,labels=FALSE)
Plots the temporal hierarchy for a given time series of seasonal periodicity.
plotSthief(y, labels = c(TRUE, FALSE), ...)plotSthief(y, labels = c(TRUE, FALSE), ...)
y |
input time series (a |
labels |
if |
... |
additional arguments passed to the plotting function. |
Produces a plot of the temporal hierarchy.
Nikolaos Kourentzes, [email protected].
Athanasopoulos, G., Hyndman, R. J., Kourentzes, N., & Petropoulos, F. (2017). Forecasting with temporal hierarchies. European Journal of Operational Research, 262(1), 60-74.
plotSthief(AirPassengers)plotSthief(AirPassengers)
Plots a hierarchy from its summing matrix S as the familiar tree of nodes and edges. Each node is a row of S, levels are arranged from the top to the bottom of the hierarchy, and each node is connected to the nodes directly below it. For large hierarchies plotSblock is usually cleaner.
plotStree( S, labels = TRUE, parse = c(FALSE, TRUE), skip = c("empty", "expand"), levels = NULL, layout = c("span", "even"), edges = c("straight", "elbow"), shape = c("box", "circle"), fill = "grey88", border = "grey30", edge.col = NULL, edge.lwd = 1.5, hgap = 0.2, vgap = 0.5, node.size = 0.2, cex = 1, min.cex = 0.35, level.labels = FALSE, level.font = 1, main = NULL )plotStree( S, labels = TRUE, parse = c(FALSE, TRUE), skip = c("empty", "expand"), levels = NULL, layout = c("span", "even"), edges = c("straight", "elbow"), shape = c("box", "circle"), fill = "grey88", border = "grey30", edge.col = NULL, edge.lwd = 1.5, hgap = 0.2, vgap = 0.5, node.size = 0.2, cex = 1, min.cex = 0.35, level.labels = FALSE, level.font = 1, main = NULL )
S |
summing matrix ( |
labels |
|
parse |
if |
skip |
how to plot edges that skip levels:
With |
levels |
optional vector with the level of each row of |
layout |
horizontal position of the nodes:
|
edges |
edge style: |
shape |
node shape: |
fill |
fill colour of the nodes. Either a single colour or one per row of |
border |
border colour of the nodes. |
edge.col |
edge colour. Either a single colour, or one per level, in which case the edge takes the colour of the level of its parent. If |
edge.lwd |
edge line width. |
hgap |
horizontal space between nodes, as a fraction of the distance between neighbouring nodes in a level. |
vgap |
vertical space between levels, as a fraction of the height of a level. |
node.size |
size of nodes without labels, in inches. Nodes with labels are sized to fit them. In both cases nodes are not larger than |
cex |
maximum label size. |
min.cex |
labels that would need to be smaller than this to fit in their node are not plotted. |
level.labels |
|
level.font |
font of the level labels, as in |
main |
optional plot title. |
A node is connected to another node below it if it aggregates all the bottom level series of that node, and there is no other node in between that does so as well. Therefore, each node is connected only to its nearest parents. In temporal hierarchies this gives the usual lattice, e.g., a monthly node is connected to both its 2-monthly and its quarterly node, but not to the annual node. In grouped hierarchies each node has one parent per grouping. Edges are shaded by the level of their parent, which helps to follow the different groupings.
Invisibly, a list with:
nodes: data frame with one row per node and columns node (row name of S), row (row index in S), level, x, y (centre of the node), width and height. Positions and sizes are in plot coordinates, with one unit per column of S and per level.
edges: data frame with one row per edge and columns from, to (row names of S), from.row and to.row (row indices in S).
Nikolaos Kourentzes, [email protected].
# Helper to construct S from a list of aggregates makeS <- function(aggs,bottom){ A <- t(sapply(aggs,function(a){as.numeric(bottom %in% a)})) S <- rbind(A,diag(length(bottom))) dimnames(S) <- list(c(names(aggs),bottom),bottom) return(S) } # Simple hierarchy S1 <- makeS(list(A=c("Q1","Q2","Q3","Q4"),H1=c("Q1","Q2"),H2=c("Q3","Q4")), c("Q1","Q2","Q3","Q4")) plotStree(S1) plotStree(S1,edges="elbow") # Row names parsed as plotmath, circles and level labels rownames(S1) <- c("A","H[1]","H[2]","Q[1]","Q[2]","Q[3]","Q[4]") plotStree(S1,parse=TRUE,shape="circle", level.labels=expression(bold(S)[4],bold(S)[2],bold(S)[1])) # Q5 skips a level and connects directly to A S2 <- makeS(list(A=c("Q1","Q2","Q3","Q4","Q5"),H1=c("Q1","Q2"),H2=c("Q3","Q4")), c("Q1","Q2","Q3","Q4","Q5")) plotStree(S2,skip="empty") tr <- plotStree(S2,skip="expand") tr$edges # Temporal hierarchy for monthly data plotStree(Sthief(12),labels=FALSE,shape="circle",level.labels=paste("k =",c(12,6,4,3,2,1))) # Grouped hierarchy: region x product, with colours for each grouping bottom <- c("R1P1","R1P2","R2P1","R2P2") S3 <- makeS(list(Total=bottom,R1=c("R1P1","R1P2"),R2=c("R2P1","R2P2"), P1=c("R1P1","R2P1"),P2=c("R1P2","R2P2")),bottom) cmp.fill <- RColorBrewer::brewer.pal(5,"Pastel1") cmp.edge <- RColorBrewer::brewer.pal(5,"Set1") plotStree(S3,fill=c("grey88",cmp.fill[2],cmp.fill[2],cmp.fill[5],cmp.fill[5],rep("grey88",4)), edge.col=c("grey20",cmp.edge[2],cmp.edge[5])) # Larger hierarchy without labels b <- paste0("b",1:60) S4 <- makeS(c(list(Total=b), setNames(lapply(0:4,function(g){b[g*12+1:12]}),paste0("G",1:5)), setNames(lapply(0:19,function(g){b[g*3+1:3]}),paste0("SG",1:20))),b) plotStree(S4,labels=FALSE,shape="circle",fill=RColorBrewer::brewer.pal(3,"Pastel1")[2]) plotStree(S4,labels=FALSE,edges="elbow")# Helper to construct S from a list of aggregates makeS <- function(aggs,bottom){ A <- t(sapply(aggs,function(a){as.numeric(bottom %in% a)})) S <- rbind(A,diag(length(bottom))) dimnames(S) <- list(c(names(aggs),bottom),bottom) return(S) } # Simple hierarchy S1 <- makeS(list(A=c("Q1","Q2","Q3","Q4"),H1=c("Q1","Q2"),H2=c("Q3","Q4")), c("Q1","Q2","Q3","Q4")) plotStree(S1) plotStree(S1,edges="elbow") # Row names parsed as plotmath, circles and level labels rownames(S1) <- c("A","H[1]","H[2]","Q[1]","Q[2]","Q[3]","Q[4]") plotStree(S1,parse=TRUE,shape="circle", level.labels=expression(bold(S)[4],bold(S)[2],bold(S)[1])) # Q5 skips a level and connects directly to A S2 <- makeS(list(A=c("Q1","Q2","Q3","Q4","Q5"),H1=c("Q1","Q2"),H2=c("Q3","Q4")), c("Q1","Q2","Q3","Q4","Q5")) plotStree(S2,skip="empty") tr <- plotStree(S2,skip="expand") tr$edges # Temporal hierarchy for monthly data plotStree(Sthief(12),labels=FALSE,shape="circle",level.labels=paste("k =",c(12,6,4,3,2,1))) # Grouped hierarchy: region x product, with colours for each grouping bottom <- c("R1P1","R1P2","R2P1","R2P2") S3 <- makeS(list(Total=bottom,R1=c("R1P1","R1P2"),R2=c("R2P1","R2P2"), P1=c("R1P1","R2P1"),P2=c("R1P2","R2P2")),bottom) cmp.fill <- RColorBrewer::brewer.pal(5,"Pastel1") cmp.edge <- RColorBrewer::brewer.pal(5,"Set1") plotStree(S3,fill=c("grey88",cmp.fill[2],cmp.fill[2],cmp.fill[5],cmp.fill[5],rep("grey88",4)), edge.col=c("grey20",cmp.edge[2],cmp.edge[5])) # Larger hierarchy without labels b <- paste0("b",1:60) S4 <- makeS(c(list(Total=b), setNames(lapply(0:4,function(g){b[g*12+1:12]}),paste0("G",1:5)), setNames(lapply(0:19,function(g){b[g*3+1:3]}),paste0("SG",1:20))),b) plotStree(S4,labels=FALSE,shape="circle",fill=RColorBrewer::brewer.pal(3,"Pastel1")[2]) plotStree(S4,labels=FALSE,edges="elbow")
Plots a temporal hierarchy, either as a tree or as blocks, and names its levels descriptively, e.g., "Monthly", "2-monthly", "Quarterly", "Yearly", instead of by their temporal aggregation level . The plots are produced by plotS, which calls plotStree or plotSblock.
plotThief( y, freq = NULL, k = NULL, view = c("tree", "block"), labels = FALSE, level.labels = TRUE, show.k = c(FALSE, TRUE), shape = c("circle", "box"), node.size = 0.5, fill = RColorBrewer::brewer.pal(3, "Pastel1")[2], border = RColorBrewer::brewer.pal(3, "Set1")[2], ... )plotThief( y, freq = NULL, k = NULL, view = c("tree", "block"), labels = FALSE, level.labels = TRUE, show.k = c(FALSE, TRUE), shape = c("circle", "box"), node.size = 0.5, fill = RColorBrewer::brewer.pal(3, "Pastel1")[2], border = RColorBrewer::brewer.pal(3, "Set1")[2], ... )
y |
input time series (a |
freq |
frequency of the bottom level series, see details. If |
k |
optional vector of temporal aggregation levels to include, passed to |
view |
type of plot: |
labels |
|
level.labels |
|
show.k |
if |
shape |
node shape for the tree view: |
node.size |
size of nodes without labels for the tree view, in inches. This is larger than the default of |
fill |
fill colour of the nodes. Either a single colour or one per row of the summing matrix. |
border |
border colour of the nodes. |
... |
additional arguments passed to |
Rows of the summing matrix are grouped into levels by the number of bottom level periods they aggregate (), with the largest at the top. Therefore, the order of the rows does not matter.
The level names depend on the frequency of the bottom level. If freq=NULL, this is inferred from the number of bottom level periods in the top level, assuming that this is the usual seasonal cycle:
4, 12, 13, 52: quarterly, monthly, 4-weekly and weekly data, with yearly top level.
260, 364, 365: business daily and daily data, with yearly top level.
5, 7, 168, 336: business daily, daily, hourly and half-hourly data, with weekly top level.
24, 48, 96, 144, 288, 1440: hourly, half-hourly, 15-minute, 10-minute, 5-minute and minutely data, with daily top level.
For other cases, or when the hierarchy does not end in the usual cycle (e.g., 24 months with a 2-yearly top level, which would otherwise be taken as hourly data), use freq. This can be "secondly", "minutely", "quarter-hourly", "half-hourly", "hourly", "daily", "business-daily", "weekly", "fortnightly", "half-monthly", "monthly", "quarterly", "half-yearly" or "yearly", optionally with a multiple, e.g., "15-minute", "2-hourly" or "4-weekly". Case, spaces, hyphens and underscores are ignored and nouns such as "minute" or "month" work as well.
Each level is named using the largest time unit that divides it exactly, e.g., 120 minutes is "2-hourly" and 14 days is "2-weekly". The following conventions are used: 7 days or 5 business days are a week; 13, 26 and 52 weeks are "Quarterly", "Half-yearly" and "Yearly"; and 365 days are a year. Levels that do not fit any of these keep the unit of the bottom level, e.g., "26-daily".
Invisibly, a list with the output of plotStree (nodes and edges) or of plotSblock (blocks), with row indices referring to the rows of the summing matrix as given, and:
levels: data frame with the level, its temporal aggregation level k and its name.
freq: name of the frequency of the bottom level.
Nikolaos Kourentzes, [email protected].
Athanasopoulos, G., Hyndman, R. J., Kourentzes, N., & Petropoulos, F. (2017). Forecasting with temporal hierarchies. European Journal of Operational Research, 262(1), 60-74.
Sthief, plotStree, plotSblock, plotS.
# Full temporal hierarchies plotThief(AirPassengers) plotThief(4,show.k=TRUE) plotThief(24) plotThief(7,labels=TRUE) # Plain level labels and smaller nodes plotThief(12,level.font=1,node.size=0.25) # Same structure, but half-monthly data plotThief(24,freq="half-monthly") # Hourly data aggregated to days and weeks, but not years plotThief(168,k=c(168,24,12)) # Two years of monthly data plotThief(24,k=c(24,12,6),freq="monthly") # 15-minute data, with numbered boxes and elbow edges plotThief(4,freq="15-minute",labels=TRUE,shape="box",edges="elbow") # Block view, better for large hierarchies plotThief(12,view="block") plotThief(364,view="block",main="Daily data")# Full temporal hierarchies plotThief(AirPassengers) plotThief(4,show.k=TRUE) plotThief(24) plotThief(7,labels=TRUE) # Plain level labels and smaller nodes plotThief(12,level.font=1,node.size=0.25) # Same structure, but half-monthly data plotThief(24,freq="half-monthly") # Hourly data aggregated to days and weeks, but not years plotThief(168,k=c(168,24,12)) # Two years of monthly data plotThief(24,k=c(24,12,6),freq="monthly") # 15-minute data, with numbered boxes and elbow edges plotThief(4,freq="15-minute",labels=TRUE,shape="box",edges="elbow") # Block view, better for large hierarchies plotThief(12,view="block") plotThief(364,view="block",main="Daily data")
Monthly Accident & Emergency referrals for England and Wales.
https://www.england.nhs.uk/statistics/statistical-work-areas/ae-waiting-times-and-activity/
Create a control chart of residuals and identify outliers.
residout(resid, t = 2, outplot = c(TRUE, FALSE))residout(resid, t = 2, outplot = c(TRUE, FALSE))
resid |
vector of residuals. |
t |
threshold value over which standardised residuals are regarded as outliers. |
outplot |
if |
A list containing:
location: locations of outliers.
outliers: values of outliers.
residuals: standardised residuals.
Nikolaos Kourentzes, [email protected].
residout(rnorm(50), outplot=TRUE)residout(rnorm(50), outplot=TRUE)
Augments a hierarchy whose observed series are not coherent with additional bottom level series that capture the observed discrepancies. The augmented hierarchy is coherent, so that standard forecast reconciliation methods can be applied to it without modification.
Saugment( S, y, tol = sqrt(.Machine$double.eps) * max(abs(y)), nodes = NULL, D = NULL, outplot = c(FALSE, TRUE), ... )Saugment( S, y, tol = sqrt(.Machine$double.eps) * max(abs(y)), nodes = NULL, D = NULL, outplot = c(FALSE, TRUE), ... )
S |
summing matrix ( |
y |
matrix or |
tol |
absolute tolerance that separates genuine discrepancies from rounding. Set this to the rounding precision of the data, e.g., |
nodes |
optional vector of names or row indices of aggregate nodes that must receive their own discrepancy series. See details. |
D |
optional matrix with the known structure of the discrepancies, which replaces their classification. Columns are named by the aggregate nodes that receive their own discrepancy series, and rows by aggregate nodes; each row gives the discrepancy of that node as a combination of the discrepancy series. Aggregate nodes not in the rows are coherent. See details. |
outplot |
if |
... |
additional arguments passed to |
Bottom level series are identified as the rows of S that are unit vectors. All other rows are aggregate nodes. For aggregate node the discrepancy is measured against the bottom level, , where is the row of S and the bottom level series.
Aggregate nodes are processed from the smallest to the largest, in terms of the number of bottom level series they aggregate, and are classified as:
"coherent": the discrepancy is within tol of zero.
"implied": the node is a linear combination of nodes that already have their own discrepancy series, both in S and, within tol, in its discrepancy. For example, net exports defined as exports minus imports. The coefficients are taken from S and stored in the corresponding row of D.
"own": the node receives its own discrepancy series.
When discrepancies are linearly dependent, which nodes receive their own discrepancy series is not unique. Use nodes to choose them.
Implied discrepancies must follow from the structure of the hierarchy: the coefficients are obtained from the rows of S, and the data are only used to check that the discrepancy follows them. Estimating the coefficients from the data instead is not reliable when the discrepancy series have few observations or smooth histories, e.g., in temporal hierarchies where each observation is a complete seasonal cycle, as a few discrepancy series can then reproduce any other one exactly with arbitrary coefficients.
When the structure of the discrepancies is known, it can be given with D, and no classification is done. The function checks that the discrepancies agree with D within tol.
A list with:
S: augmented summing matrix (). Rows are the rows of the input S, in the same order, followed by the discrepancy series. Columns are the bottom level series followed by the discrepancy series.
y: input series with the discrepancy series appended. This is a ts object if y is one.
D: matrix () that maps the discrepancy series to the aggregate nodes.
d: the discrepancy series ().
eps: discrepancies of all aggregate nodes ().
status: classification of each aggregate node.
error: maximum absolute incoherence of the augmented series, i.e., max(abs(y - B %*% t(S))), with B the augmented bottom level series. This should be within tol.
Nikolaos Kourentzes, [email protected].
# Small hierarchy: Total = A + B, A = A1 + A2, B = B1 + B2 S <- rbind(c(1,1,1,1),c(1,1,0,0),c(0,0,1,1),diag(4)) rownames(S) <- c("Total","A","B","A1","A2","B1","B2") set.seed(1) b <- matrix(rnorm(40,10),10,4) y <- b %*% t(S) colnames(y) <- rownames(S) # Add discrepancies at Total and A y[,"Total"] <- y[,"Total"] + rnorm(10) y[,"A"] <- y[,"A"] + rnorm(10) aug <- Saugment(S,ts(y,frequency=4),outplot=TRUE) aug$status aug$S # Block view, passing arguments to plotS aug <- Saugment(S,y,outplot=TRUE,view="block") # Total is published as A + B, while A and B differ from their bottom level # series. The discrepancy of Total is then the sum of those of A and B. y <- b %*% t(S) colnames(y) <- rownames(S) y[,"A"] <- y[,"A"] + rnorm(10) y[,"B"] <- y[,"B"] + rnorm(10) y[,"Total"] <- y[,"A"] + y[,"B"] # By default A and B get their own discrepancy series and Total is implied Saugment(S,y)$status # Use nodes to give Total its own discrepancy series instead. B is now implied # as the difference between the discrepancies of Total and A aug <- Saugment(S,y,nodes="Total") aug$status aug$D # Short sample: a temporal hierarchy of a seasonal series together with its # seasonally adjusted version, with only 5 years of data. The discrepancies of the # adjusted aggregates are sums of those of the adjusted months, following S. y <- ts(100 + 10*sin(2*pi*(1:60)/12) + rnorm(60),frequency=12) s <- stl(y,s.window=7)$time.series[,"seasonal"] Sn <- Sthief(y) rownames(Sn) <- paste0("n",1:nrow(Sn)) S2 <- rbind(Sn,Sn) rownames(S2) <- c(paste0("NSA.",rownames(Sn)),paste0("SA.",rownames(Sn))) stack <- function(x){matrix(x,ncol=12,byrow=TRUE) %*% t(Sn)} Y <- cbind(stack(y),stack(y-s)) colnames(Y) <- rownames(S2) aug <- Saugment(S2,Y) table(aug$status) # The same structure given explicitly with D D <- Sn dimnames(D) <- list(paste0("SA.",rownames(Sn)),paste0("SA.n",17:28)) aug <- Saugment(S2,Y,D=D) table(aug$status)# Small hierarchy: Total = A + B, A = A1 + A2, B = B1 + B2 S <- rbind(c(1,1,1,1),c(1,1,0,0),c(0,0,1,1),diag(4)) rownames(S) <- c("Total","A","B","A1","A2","B1","B2") set.seed(1) b <- matrix(rnorm(40,10),10,4) y <- b %*% t(S) colnames(y) <- rownames(S) # Add discrepancies at Total and A y[,"Total"] <- y[,"Total"] + rnorm(10) y[,"A"] <- y[,"A"] + rnorm(10) aug <- Saugment(S,ts(y,frequency=4),outplot=TRUE) aug$status aug$S # Block view, passing arguments to plotS aug <- Saugment(S,y,outplot=TRUE,view="block") # Total is published as A + B, while A and B differ from their bottom level # series. The discrepancy of Total is then the sum of those of A and B. y <- b %*% t(S) colnames(y) <- rownames(S) y[,"A"] <- y[,"A"] + rnorm(10) y[,"B"] <- y[,"B"] + rnorm(10) y[,"Total"] <- y[,"A"] + y[,"B"] # By default A and B get their own discrepancy series and Total is implied Saugment(S,y)$status # Use nodes to give Total its own discrepancy series instead. B is now implied # as the difference between the discrepancies of Total and A aug <- Saugment(S,y,nodes="Total") aug$status aug$D # Short sample: a temporal hierarchy of a seasonal series together with its # seasonally adjusted version, with only 5 years of data. The discrepancies of the # adjusted aggregates are sums of those of the adjusted months, following S. y <- ts(100 + 10*sin(2*pi*(1:60)/12) + rnorm(60),frequency=12) s <- stl(y,s.window=7)$time.series[,"seasonal"] Sn <- Sthief(y) rownames(Sn) <- paste0("n",1:nrow(Sn)) S2 <- rbind(Sn,Sn) rownames(S2) <- c(paste0("NSA.",rownames(Sn)),paste0("SA.",rownames(Sn))) stack <- function(x){matrix(x,ncol=12,byrow=TRUE) %*% t(Sn)} Y <- cbind(stack(y),stack(y-s)) colnames(Y) <- rownames(S2) aug <- Saugment(S2,Y) table(aug$status) # The same structure given explicitly with D D <- Sn dimnames(D) <- list(paste0("SA.",rownames(Sn)),paste0("SA.n",17:28)) aug <- Saugment(S2,Y,D=D) table(aug$status)
Create binary or trigonometric seasonal dummies.
seasdummy(n, m = NULL, y = NULL, type = c("bin", "trg"), full = c(FALSE, TRUE))seasdummy(n, m = NULL, y = NULL, type = c("bin", "trg"), full = c(FALSE, TRUE))
n |
number of observations to create. |
m |
seasonal periodicity. If |
y |
this is an optional time series input that can be used to get seasonal periodicity ( |
type |
type of seasonal dummies to create.
|
full |
If full is |
An array with seasonal dummies, where rows correspond observations and columns to dummy variables.
If the seasonal periodicity is fractional then the the type will be overriden to trigonometric and only two seasonal dummies with be produced. One cosine and one sine.
Nikolaos Kourentzes, [email protected]
seasdummy(24,12)seasdummy(24,12)
Construct seasonal plots of various styles for a given time series. The series can be automatically detrended as needed.
seasplot( y, m = NULL, s = NULL, trend = NULL, colour = NULL, alpha = 0.05, outplot = c(1, 0, 2, 3, 4, 5), decomposition = c("multiplicative", "additive", "auto"), cma = NULL, labels = NULL, legend = TRUE, ... )seasplot( y, m = NULL, s = NULL, trend = NULL, colour = NULL, alpha = 0.05, outplot = c(1, 0, 2, 3, 4, 5), decomposition = c("multiplicative", "additive", "auto"), cma = NULL, labels = NULL, legend = TRUE, ... )
y |
input time series. Can be |
m |
seasonal period. If |
s |
starting period in the season. If |
trend |
if |
colour |
colour override for plots. A single colour, or a vector of colours, used as follows:
|
alpha |
significance level for statistical tests. |
outplot |
type of seasonal plot
|
decomposition |
type of seasonal decomposition. This can be |
cma |
input precalculated level/trend for the analysis. If |
labels |
external labels for the seasonal periods. Use |
legend |
if |
... |
additional arguments passed to plotting functions. Arguments that refer to the plotted data, |
The seasonality test is the same as in seastest, which can be used to run it without producing a plot.
An object of class seasexpl containing:
season: matrix of (detrended) seasonal elements.
season.exist: TRUE/FALSE results of seasonality test, or NULL if not tested.
season.pval: p-value of seasonality test (Friedman test), or NULL if not tested.
trend: CMA estimate (using cmav) or NULL if no trend was removed.
trend.exist: TRUE/FALSE results of trend test.
trend.pval: p-value of trend test (Cox-Stuart).
decomposition: type of decomposition used. "none" if no trend was removed.
Nikolaos Kourentzes, [email protected].
seasplot(referrals,outplot=1) seasplot(referrals,outplot=1,colour=c("grey","red"),lwd=2,legend="bottomright") seasplot(referrals,outplot=3,labels=month.abb,las=2,main="Referrals",ylab="Index")seasplot(referrals,outplot=1) seasplot(referrals,outplot=1,colour=c("grey","red"),lwd=2,legend="bottomright") seasplot(referrals,outplot=3,labels=month.abb,las=2,main="Referrals",ylab="Index")
Test a time series for deterministic seasonality using the Friedman test on the distributions of the (detrended) seasonal periods. The series can be automatically detrended as needed. This is the test reported by seasplot.
seastest( y, m = NULL, s = NULL, trend = NULL, alpha = 0.05, decomposition = c("multiplicative", "additive", "auto"), cma = NULL )seastest( y, m = NULL, s = NULL, trend = NULL, alpha = 0.05, decomposition = c("multiplicative", "additive", "auto"), cma = NULL )
y |
input time series. Can be |
m |
seasonal period. If |
s |
starting period in the season. If |
trend |
if |
alpha |
significance level for statistical tests. |
decomposition |
type of seasonal decomposition. This can be |
cma |
input precalculated level/trend for the analysis. If |
The series is first detrended (if needed) using a centred moving average (see cmav), with the presence of trend tested using the Cox-Stuart test (see coxstuart) when trend=NULL. The detrended series is then arranged in a matrix, with one row per season and one column per seasonal period, and the Friedman test is used to test whether the seasonal periods come from the same distribution. Only complete seasons are used for the test, so at least two complete seasons are needed.
An object of class seastest containing:
is.seasonal: TRUE/FALSE result of the seasonality test. NA if the test could not be performed.
statistic: Friedman test statistic.
pvalue: p-value of the seasonality test.
alpha: significance level used.
season: matrix of (detrended) seasonal elements.
trend: CMA estimate (using cmav) or NULL if no trend was removed.
trend.exist: TRUE/FALSE results of trend test, or NULL if not tested.
trend.pval: p-value of trend test (Cox-Stuart), or NULL if not tested.
decomposition: type of decomposition used. "none" if no trend was removed.
Nikolaos Kourentzes, [email protected].
seasplot, mseastest, trendtest.
seastest(referrals) seastest(referrals,trend=TRUE,decomposition="auto")$pvalueseastest(referrals) seastest(referrals,trend=TRUE,decomposition="auto")$pvalue
Construct the summing matrix S of a grouped hierarchy from the labels of the bottom level series on one or more dimensions, e.g., product type, region, and organic. Redundant aggregate nodes are removed.
Sgroup( x, names = NULL, sep = "_", cross = c("all", "none"), na = c("exclude", "group"), total = "Total", sign = NULL, sort.levels = c(FALSE, TRUE), outplot = c(FALSE, TRUE), ... )Sgroup( x, names = NULL, sep = "_", cross = c("all", "none"), na = c("exclude", "group"), total = "Total", sign = NULL, sort.levels = c(FALSE, TRUE), outplot = c(FALSE, TRUE), ... )
x |
data frame or matrix of labels, with one row per bottom level series and one column per dimension. Column names are used as dimension names. |
names |
optional names of the bottom level series. If |
sep |
separator used to construct the names of the nodes. |
cross |
which combinations of dimensions to include, in addition to the total and each dimension on its own:
|
na |
how to treat missing labels:
|
total |
name of the top level node. |
sign |
optional vector of signs ( |
sort.levels |
if |
outplot |
if |
... |
additional arguments passed to |
For the total, each dimension on its own, and each combination of dimensions requested in cross, the series are grouped by their labels, giving one aggregate node per combination of values. A combination of dimensions forms one level of the hierarchy. Levels are ordered from the total, to single dimensions, to combinations of two dimensions, and so on, with the bottom level series last.
With sign, series enter aggregate nodes with their sign. Nodes whose series are all negative are gross totals and are kept positive, e.g., total imports, while nodes that mix signs are net values, e.g., net exports. Bottom level series are always positive.
Aggregate nodes are redundant and removed when they:
aggregate a single series, as they repeat a bottom level series;
aggregate the same series as a node higher up, e.g., a dimension with a single value repeats the total, or a combination of nested dimensions (country within region) repeats the finer dimension. The first, higher, node is kept.
Levels left without any nodes disappear. If sort.levels=TRUE, the levels are reordered by their number of nodes, after removing the redundant nodes. Levels are named by their dimensions, e.g., "product x region", and the bottom level is named "Bottom". Nodes are named by their values joined with sep. If the same name arises from different dimensions, the dimension name is added, e.g., "organic:yes".
S matrix (), with row and column names. The attribute "levels" gives the level (combination of dimensions) of each row, which can be passed to plotS. The attribute "removed" gives the names of the removed redundant nodes.
Nikolaos Kourentzes, [email protected].
# Six products in five regions, organic or not x <- expand.grid(organic=c("no","yes"), region=c("denmark","sweden","norway","iceland","finland"), product=c("milk","yoghurt","cheese","quark","whey","casein"), stringsAsFactors=FALSE)[,3:1] S <- Sgroup(x,outplot=TRUE) dim(S) table(attr(S,"levels")) # Only the total and each dimension on its own S <- Sgroup(x,cross="none",outplot=TRUE) # Add only the product by region combination S <- Sgroup(x,cross="product:region",outplot=TRUE) # Levels with fewer nodes higher up: organic, region, product, ... S <- Sgroup(x,cross="product:region",sort.levels=TRUE,outplot=TRUE) unique(attr(S,"levels")) # Block view, passing arguments to plotS S <- Sgroup(x,cross="product:region",outplot=TRUE,view="block") # Redundant dimensions: market has a single value and union is nested in region x$market <- "nordics" x$union <- ifelse(x$region %in% c("norway","iceland"),"non-eu","eu") S <- Sgroup(x) head(attr(S,"removed")) # Missing labels x <- x[,1:3] x$region[1:4] <- NA S <- Sgroup(x,cross="none") S[c("Total","denmark"),1:6] # Net values: exports and imports of goods and services x <- data.frame(flow=c("exports","exports","imports","imports"), type=c("goods","services","goods","services"), row.names=c("X.goods","X.services","M.goods","M.services")) S <- Sgroup(x,cross="none",sign=ifelse(x$flow=="imports",-1,1),total="NetExports") S # NetExports, goods and services are net values, exports and imports are gross totals# Six products in five regions, organic or not x <- expand.grid(organic=c("no","yes"), region=c("denmark","sweden","norway","iceland","finland"), product=c("milk","yoghurt","cheese","quark","whey","casein"), stringsAsFactors=FALSE)[,3:1] S <- Sgroup(x,outplot=TRUE) dim(S) table(attr(S,"levels")) # Only the total and each dimension on its own S <- Sgroup(x,cross="none",outplot=TRUE) # Add only the product by region combination S <- Sgroup(x,cross="product:region",outplot=TRUE) # Levels with fewer nodes higher up: organic, region, product, ... S <- Sgroup(x,cross="product:region",sort.levels=TRUE,outplot=TRUE) unique(attr(S,"levels")) # Block view, passing arguments to plotS S <- Sgroup(x,cross="product:region",outplot=TRUE,view="block") # Redundant dimensions: market has a single value and union is nested in region x$market <- "nordics" x$union <- ifelse(x$region %in% c("norway","iceland"),"non-eu","eu") S <- Sgroup(x) head(attr(S,"removed")) # Missing labels x <- x[,1:3] x$region[1:4] <- NA S <- Sgroup(x,cross="none") S[c("Total","denmark"),1:6] # Net values: exports and imports of goods and services x <- data.frame(flow=c("exports","exports","imports","imports"), type=c("goods","services","goods","services"), row.names=c("X.goods","X.services","M.goods","M.services")) S <- Sgroup(x,cross="none",sign=ifelse(x$flow=="imports",-1,1),total="NetExports") S # NetExports, goods and services are net values, exports and imports are gross totals
Converts a summing matrix S, where aggregate nodes are sums of bottom level series, to one where aggregate nodes are means of the bottom level series they aggregate. The means can use equal weights or a vector of bottom level weights. Negative entries, e.g., imports in net exports, are allowed.
Smean(S, w = NULL)Smean(S, w = NULL)
S |
summing matrix ( |
w |
optional vector of non-negative bottom level weights, one per column of |
Each aggregate row of S is replaced by , so that every aggregate node is a convex combination of the bottom level series it aggregates, entering with their sign. For example, net exports with equal weights become . For non-negative S the normalisation is the usual weighted sum. The weights do not need to sum to one, as they are normalised within each node. Bottom level rows (unit rows) are not changed, so a bottom level series with zero weight is excluded from the means but remains in the hierarchy. Entries of S other than one, e.g., from Sgroup or temporal hierarchies, are kept as relative weights.
The result is still a linear aggregation, , so it can be used in place of the summing matrix for forecast reconciliation, e.g., for series that are rates, prices, or stocks rather than flows.
If structural scaling is used with the resulting S, it should be derived from its assumptions rather than computed from S: with uncorrelated bottom level errors of equal variance, the error variance of node is proportional to , which is for an equal weight mean over series. Using diag(rowSums(S)), as for sums, gives the identity matrix, which corresponds to OLS rather than structural scaling.
S matrix with weighted mean aggregation, with the same dimensions, names and attributes as S.
Nikolaos Kourentzes, [email protected].
# Small hierarchy: A = H1 + H2, H1 = Q1 + Q2, H2 = Q3 + Q4 S <- rbind(c(1,1,1,1),c(1,1,0,0),c(0,0,1,1),diag(4)) rownames(S) <- c("A","H1","H2","Q1","Q2","Q3","Q4") # Equal weight means Smean(S) # Weighted means, e.g., by sales volume of each bottom level series Smean(S,w=c(10,30,20,40)) # Temporal hierarchy for quarterly data where the annual value is the average Smean(Sthief(4)) # Net exports: exports and imports of goods and services S <- rbind(NetExports=c(1,1,-1,-1),Exports=c(1,1,0,0),Imports=c(0,0,1,1),diag(4)) colnames(S) <- rownames(S)[4:7] <- c("X.goods","X.services","M.goods","M.services") Smean(S) # Structural scaling for the mean hierarchy, derived from its assumptions Sm <- Smean(S) W <- diag(rowSums(Sm^2)/rowSums(abs(Sm))^2) diag(W)# Small hierarchy: A = H1 + H2, H1 = Q1 + Q2, H2 = Q3 + Q4 S <- rbind(c(1,1,1,1),c(1,1,0,0),c(0,0,1,1),diag(4)) rownames(S) <- c("A","H1","H2","Q1","Q2","Q3","Q4") # Equal weight means Smean(S) # Weighted means, e.g., by sales volume of each bottom level series Smean(S,w=c(10,30,20,40)) # Temporal hierarchy for quarterly data where the annual value is the average Smean(Sthief(4)) # Net exports: exports and imports of goods and services S <- rbind(NetExports=c(1,1,-1,-1),Exports=c(1,1,0,0),Imports=c(0,0,1,1),diag(4)) colnames(S) <- rownames(S)[4:7] <- c("X.goods","X.services","M.goods","M.services") Smean(S) # Structural scaling for the mean hierarchy, derived from its assumptions Sm <- Smean(S) W <- diag(rowSums(Sm^2)/rowSums(abs(Sm))^2) diag(W)
Calculate the temporal hierarchy summing matrix S for a given time series of seasonal periodicity.
Sthief(y, k = NULL)Sthief(y, k = NULL)
y |
input time series (a |
k |
optional vector of temporal aggregation levels to include. These must be factors of the seasonal periodicity, i.e., |
S matrix. Rows are ordered from the top level (largest k) to the bottom level (k=1). The top level does not need to be a single node, for example Sthief(24,k=c(12,1)) gives two annual nodes over 24 months.
Nikolaos Kourentzes, [email protected].
Athanasopoulos, G., Hyndman, R. J., Kourentzes, N., & Petropoulos, F. (2017). Forecasting with temporal hierarchies. European Journal of Operational Research, 262(1), 60-74.
Sthief(AirPassengers) # Only some aggregation levels: monthly, quarterly, and annual Sthief(12,k=c(12,3)) # Two years of monthly data, with half-yearly, annual and 2-yearly levels S <- Sthief(24,k=c(24,12,6)) plotThief(S,freq="monthly")Sthief(AirPassengers) # Only some aggregation levels: monthly, quarterly, and annual Sthief(12,k=c(12,3)) # Two years of monthly data, with half-yearly, annual and 2-yearly levels S <- Sthief(24,k=c(24,12,6)) plotThief(S,freq="monthly")
Estimate Theta method.
Forecast with fitted Theta method.
Produce a plot of the fitted Theta method.
theta( y, m = NULL, sign.level = 0.05, cost0 = c("MSE", "MdSE", "MAE", "MdAE"), cost2 = c("MSE", "MdSE", "MAE", "MdAE"), costs = c("MSE", "MdSE", "MAE", "MdAE"), season = c("multiplicative", "additive", "auto"), cma = NULL, outliers = NULL, multiplicative = NULL ) ## S3 method for class 'theta' forecast(object, h = NULL, ...) ## S3 method for class 'theta' plot(x, thetalines = c(TRUE, FALSE), h = 0, legend = TRUE, ...) theta.thief(y, h = NULL, ...)theta( y, m = NULL, sign.level = 0.05, cost0 = c("MSE", "MdSE", "MAE", "MdAE"), cost2 = c("MSE", "MdSE", "MAE", "MdAE"), costs = c("MSE", "MdSE", "MAE", "MdAE"), season = c("multiplicative", "additive", "auto"), cma = NULL, outliers = NULL, multiplicative = NULL ) ## S3 method for class 'theta' forecast(object, h = NULL, ...) ## S3 method for class 'theta' plot(x, thetalines = c(TRUE, FALSE), h = 0, legend = TRUE, ...) theta.thief(y, h = NULL, ...)
y |
input time series. Can be |
m |
seasonal period. If |
sign.level |
significance level for trend and seasonality tests. |
cost0 |
cost function of theta0 line. Can be:
|
cost2 |
cost function of theta2 line. Same options as |
costs |
cost function of seasonal element. Same options as |
season |
type of seasonal decomposition. This can be |
cma |
input precalculated level/trend for the analysis. Use |
outliers |
provide vector of location of observations that are considered outliers (see |
multiplicative |
deprecated, use |
object |
object of class |
h |
forecast horizon. If |
... |
additional arguments passed to functions. |
x |
object of class |
thetalines |
if |
legend |
if |
This implementation of the Theta method tests automatically for seasonality and trend. Seasonal decomposition can be done either additively or multiplicatively and the seasonality is treated as a pure seasonal model. The various Theta components can be optimised using different cost functions. The originally proposed Theta method always assumed multiplicative seasonality and presence of trend, while all theta lines were optimised using MSE and seasonality was estimated using classical decomposition.
The argument season replaces the argument multiplicative of earlier versions. Calls that use multiplicative still work, but give a deprecation warning. The output element multiplicative is unchanged.
For plot.theta, if h > 0 the forecasts are added to the plot, together with the forecasts of the theta lines if thetalines=TRUE. Outliers provided in outliers are marked on the data. Arguments in ... control the plot: col sets the colours of the data, the fit and forecast, theta0, theta2 and the outliers, in this order (recycled as needed); lwd and lty the lines; everything else, such as main, xlab, ylab, xlim or ylim, is passed to plot. The default colours are from the Set1 palette of brewer.pal.
An object of class theta, containing:
"method": "Theta".
"y": the input time series.
"m": seasonal periods.
"exist": Statistical testing results, exist[1] is the result for trend, exist[2] is for season.
"multiplicative": If TRUE, then seasonality is modelled multiplicatively.
"theta0": fitted theta0 line values.
"theta2": fitted theta2 line values.
"season": fitted season values.
"x.out": modelled outliers.
"cost": cost functions for theta0, theta2 and season components.
"a": SES parameters of theta2.
"b": regression parameters of theta0.
"p": coefficients of outliers from theta0 and theta2 estimation.
"g": pure seasonal exponential smoothing parameters.
"fitted": fitted values.
"residuals": in-sample residuals.
"MSE": in-sample Mean Squared Error.
"outliers": location of outliers provided in outliers.
theta.thief(): Wrapper function to use Theta with thief.
Nikolaos Kourentzes, [email protected].
The original Theta method was proposed by: Assimakopoulos, V., & Nikolopoulos, K. (2000). The theta model: a decomposition approach to forecasting. International journal of forecasting, 16(4), 521-530. See details in how the implementation here differs.
The THieF forecasting methodology used for theta.thief is proposed by: Athanasopoulos, G., Hyndman, R. J., Kourentzes, N., & Petropoulos, F. (2017). Forecasting with temporal hierarchies. European Journal of Operational Research, 262(1), 60-74.
fit <- theta(referrals) plot(fit) forecast.theta(fit,h=12) # Or simply use forecast(fit) plot(fit,h=24) plot(fit,thetalines=FALSE,h=12,main="Referrals",legend="bottomright") ## Not run: library(thief) thief(referrals,forecastfunction=theta.thief) ## End(Not run)fit <- theta(referrals) plot(fit) forecast.theta(fit,h=12) # Or simply use forecast(fit) plot(fit,h=24) plot(fit,thetalines=FALSE,h=12,main="Referrals",legend="bottomright") ## Not run: library(thief) thief(referrals,forecastfunction=theta.thief) ## End(Not run)
Test a time series for trend by either fitting exponential smoothing models and comparing then using the AICc, or by using the non-parametric Cox-Stuart test. The tests can be augmented by using multiple temporal aggregation.
trendtest( y, extract = c("FALSE", "TRUE"), type = c("aicc", "cs"), mta = c(FALSE, TRUE) )trendtest( y, extract = c("FALSE", "TRUE"), type = c("aicc", "cs"), mta = c(FALSE, TRUE) )
y |
a time series that must be of either |
extract |
if |
type |
type of test. Can be:
|
mta |
If |
All tests are performed at 5
The function returns TRUE when there is evidence of trend and FALSE otherwise.
Nikolaos Kourentzes, [email protected].
The multiple temporal aggregation follows the construction approach suggested by Kourentzes, N., Petropoulos, F., & Trapero, J. R. (2014). Improving forecasting by estimating time series structural components across multiple frequencies. International Journal of Forecasting, 30(2), 291-302.
trendtest(referrals,TRUE)trendtest(referrals,TRUE)
The tsutils package provides functions to support various aspects of time series and forecasting modelling. In particular this package includes: (i) tests and visualisations that can help the modeller explore time series components and perform decomposition; (ii) modelling shortcuts, such as functions to construct lagmatrices and seasonal dummy variables of various forms; (iii) an implementation of the Theta method; (iv) tools to facilitate the design of the forecasting process, such as ABC-XYZ analyses; and (v) "quality of life" tools, such as treating time series for trailing and leading values.
cmav: centred moving average.
coxstuart: Cox-Stuart test for location/dispersion.
decomp: classical time series decomposition.
seasplot: construct seasonal plots.
seastest: seasonality test based on the Friedman test.
trendtest: test a time series for trend.
getOptK: optimal temporal aggregation level for AR(1), MA(1), ARMA(1,1).
lagmatrix: create leads/lags of variable.
nemenyi: nonparametric multiple comparisons.
residout: construct control chart of residuals.
seasdummy: create seasonal dummies.
theta: Theta method.
plotSthief: plot temporal hierarchy S matrix. Obsolete, replaced by plotThief.
plotS: general plotting function for S-matrices. Calls plotSblock or plotStree.
plotSblock: plot an S-matrix using the block representation.
plotStree: plot an S-matrix using the standard tree representation.
plotThief: plot temporal hierarchy S matrix. Calls plotS and handles temporal names automatically.
Saugment: augment an incoherent hierarchy with discrepancy series.
Sgroup: S matrix from dimensions.
Smean: Mean (equal or unequal weights) aggregation S matrix.
Sthief: temporal hierarchy S matrix.
geomean: geometric mean.
lambdaseq: generate sequence of lambda for LASSO regression.
leadtrail: remove leading/training zeros/NAs.
wins: winsorisation, including vectorised versions colWins and rowWins.
Winsorise either by number or percentage of observations.
wins(x, p = 0.05) colWins(x, p = 0.05) rowWins(x, p = 0.05)wins(x, p = 0.05) colWins(x, p = 0.05) rowWins(x, p = 0.05)
x |
input data. NAs will be removed. |
p |
percentage or number of observations to be winsorised. If value is <1 then it is used as a percentages. Otherwise it is the number of observations to winsorise. If the resulting p > floor((length(x)-1)/2), then it is set equal to floor((length(x)-1)/2). |
Winsorised vector.
colWins(): Vectorised version of wins by columns.
rowWins(): Vectorised version of wins by rows.
Nikolaos Kourentzes, [email protected].
x <- rnorm(100,mean=0,sd=1) xW <- wins(x)x <- rnorm(100,mean=0,sd=1) xW <- wins(x)
Perform XYZ analysis on a set of time series.
xyz(x, m = NULL, prc = c(0.2, 0.3, 0.5), type = c("naive", "ets", "cv"))xyz(x, m = NULL, prc = c(0.2, 0.3, 0.5), type = c("naive", "ets", "cv"))
x |
this can either be an array, where each column is a series, or a vector of values. If |
m |
seasonal length for time series. Required when type is |
prc |
a vector of percentages indicating how many items are included in each class. By default this is |
type |
the type of forecastability calculation. This can be:
|
Return object of class abc and contains:
value: a vector containing the forecastability value of each series.
class: a vector containing the class membership of each series.
rank: a vector containing the rank of each series, with 1 being the lowest forecastability series.
conc: the forecastability concentration of each class, as percentage of total value.
model: fitted model for each series.
Nikolaos Kourentzes, [email protected].
Ord K., Fildes R., Kourentzes N. (2017) Principles of Business Forecasting, 2e. Wessex Press Publishing Co., p.515-518.
x <- abs(matrix(cumsum(rnorm(5400,0,1)),36,150)) z <- xyz(x,m=12) print(z) plot(z)x <- abs(matrix(cumsum(rnorm(5400,0,1)),36,150)) z <- xyz(x,m=12) print(z) plot(z)