Title: Seasonal Trend Decomposition Using Regression
Version: 0.7
Description: Methods for decomposing seasonal data: STR (a Seasonal-Trend time series decomposition procedure based on Regression) and Robust STR. In some ways, STR is similar to Ridge Regression and Robust STR can be related to LASSO. They allow for multiple seasonal components, multiple linear covariates with constant, flexible and seasonal influence. Seasonal patterns (for both seasonal components and seasonal covariates) can be fractional and flexible over time; moreover they can be either strictly periodic or have a more complex topology. The methods provide confidence intervals for the estimated components. The methods can also be used for forecasting.
Depends: R (≥ 3.5.0)
Imports: compiler, foreach, forecast, graphics, grDevices, Matrix, methods, quantreg, SparseM, stats
Suggests: demography, doParallel, knitr, markdown, rgl, rmarkdown, seasonal, testthat
URL: https://pkg.robjhyndman.com/stR/, https://github.com/robjhyndman/stR
BugReports: https://github.com/robjhyndman/stR/issues
License: GPL-3
LazyData: true
Encoding: UTF-8
Config/testthat/edition: 3
RoxygenNote: 7.3.2
VignetteBuilder: knitr
NeedsCompilation: no
Packaged: 2024-07-28 05:56:50 UTC; hyndman
Author: Alexander Dokumentov ORCID iD [aut], Rob Hyndman ORCID iD [aut, cre]
Maintainer: Rob Hyndman <Rob.Hyndman@monash.edu>
Repository: CRAN
Date/Publication: 2024-07-28 13:30:01 UTC

Automatic STR decomposition for time series data

Description

Automatically selects parameters for an STR decomposition of time series data. The time series should be of class ts or msts.

Usage

AutoSTR(
  data,
  robust = FALSE,
  gapCV = NULL,
  lambdas = NULL,
  reltol = 0.001,
  confidence = NULL,
  nsKnots = NULL,
  trace = FALSE
)

Arguments

data

A time series of class ts or msts.

robust

When TRUE, Robust STR decomposition is used. Default is FALSE.

gapCV

An optional parameter defining the length of the sequence of skipped values in the cross validation procedure.

lambdas

An optional parameter. A structure which replaces lambda parameters provided with predictors. It is used as either a starting point for the optimisation of parameters or as the exact model parameters.

reltol

An optional parameter which is passed directly to optim() when optimising the parameters of the model.

confidence

A vector of percentiles giving the coverage of confidence intervals. It must be greater than 0 and less than 1. If NULL, no confidence intervals are produced.

nsKnots

An optional vector parameter, defining the number of seasonal knots (per period) for each sesonal component.

trace

When TRUE, tracing is turned on.

Value

A structure containing input and output data. It is an S3 class STR, which is a list with the following components:

Author(s)

Alexander Dokumentov

References

Dokumentov, A., and Hyndman, R.J. (2022) STR: Seasonal-Trend decomposition using Regression, INFORMS Journal on Data Science, 1(1), 50-62. https://robjhyndman.com/publications/str/

See Also

STR

Examples


# Decomposition of a multiple seasonal time series
decomp <- AutoSTR(calls)
plot(decomp)

# Decomposition of a monthly time series
decomp <- AutoSTR(log(grocery))
plot(decomp)

Robust STR decomposition

Description

Robust Seasonal-Trend decomposition of time series data using Regression (robust version of STRmodel).

Usage

RSTRmodel(
  data,
  predictors = NULL,
  strDesign = NULL,
  lambdas = NULL,
  confidence = NULL,
  nMCIter = 100,
  control = list(nnzlmax = 1e+06, nsubmax = 3e+05, tmpmax = 50000),
  reportDimensionsOnly = FALSE,
  trace = FALSE
)

Arguments

data

Time series or a vector of length L.

predictors

List of predictors.
According to the paradigm of this implementation, the trend, the seasonal components, the flexible predictors and the seasonal predictors are all presented in the same form (as predictors) and must be described in this list.
Every predictor is a list of the following structures:

  • data – vector of length L (length of input data, see above). For trend or for a seasonal component it is a vector of ones. For a flexible or a seasonal predictor it is a vector of the predictor's data.

  • times – vector of length L of times of observations.

  • seasons – vector of length L. It is a vector of ones for a trend or a flexible predictor. It is vector assigning seasons to every observation (for a seasonal component or a seasonal predictor). Seasons can be fractional for observations in between seasons.

  • timeKnots – vector of times (time knots) where knots are positioned (for a seasonal component or a seasonal predictor a few knots have the same time; every knot is represented by time and season). Usually this vector coincides with times vector described above, or timeKnots is a subset of times vector.

  • seasonalStructure – describes seasonal topology (which can have complex structure) and seasonal knots.The seasonal topology is described by a list of segments and seasonal knots, which are positioned inside the segments, on borders of the segments or, when they are on on borders, they can connect two or more segments.
    seasonalStructure is a list of two elements:

    • segments – a list of vectors representing segments. Each vector must contain two ordered real values which represent left and right borders of a segment. Segments should not intersect (inside same predictor).

    • sKnots – a list of real values (vectors of length one) or vectors of lengths two or greater (seasonal knots) defining seasons of the knots (every knot is represented by time and season). All real values must belong (be inside or on border of) segments listed in segments. If a few values represent a single seasonal knot then all these values must be on borders of some segments (or a single segment). In this case they represent a seasonal knot which connects a few segments (or both sides of one segment).

  • lambdas – a vector with three values representing lambda (smoothing) parameters (time-time, season-season, time-season flexibility parameters) for this predictor.

strDesign

An optional parameter used to create the design matrix. It is used internally in the library to improve performance when the design matrix does not require full recalculation.

lambdas

An optional parameter. A structure which replaces lambda parameters provided with predictors. It is used as either a starting point for the optimisation of parameters or as the exact model parameters.

confidence

A vector of percentiles giving the coverage of confidence intervals. It must be greater than 0 and less than 1. If NULL, no confidence intervals are produced.

nMCIter

Number of Monte Carlo iterations used to estimate confidence intervals for Robust STR decomposition.

control

Passed directly to rq.fit.sfn() during Robust STR decomposition.

reportDimensionsOnly

A boolean parameter. When TRUE the method constructs the design matrix and reports its dimensions without proceeding further. It is mostly used for debugging.

trace

When TRUE, tracing is turned on.

Value

A structure containing input and output data. It is an S3 class STR, which is a list with the following components:

Author(s)

Alexander Dokumentov

References

Dokumentov, A., and Hyndman, R.J. (2022) STR: Seasonal-Trend decomposition using Regression, INFORMS Journal on Data Science, 1(1), 50-62. https://robjhyndman.com/publications/str/

See Also

STRmodel STR

Examples


n <- 70
trendSeasonalStructure <- list(segments = list(c(0, 1)), sKnots = list(c(1, 0)))
ns <- 5
seasonalStructure <- list(
  segments = list(c(0, ns)),
  sKnots = c(as.list(1:(ns - 1)), list(c(ns, 0)))
)
seasons <- (0:(n - 1)) %% ns + 1
trendSeasons <- rep(1, length(seasons))
times <- seq_along(seasons)
data <- seasons + times / 4
set.seed(1234567890)
data <- data + rnorm(length(data), 0, 0.2)
data[20] <- data[20] + 3
data[50] <- data[50] - 5
plot(times, data, type = "l")
timeKnots <- times
trendData <- rep(1, n)
seasonData <- rep(1, n)
trend <- list(
  data = trendData, times = times, seasons = trendSeasons,
  timeKnots = timeKnots, seasonalStructure = trendSeasonalStructure, lambdas = c(1, 0, 0)
)
season <- list(
  data = seasonData, times = times, seasons = seasons,
  timeKnots = timeKnots, seasonalStructure = seasonalStructure, lambdas = c(1, 0, 1)
)
predictors <- list(trend, season)
rstr <- RSTRmodel(data, predictors, confidence = 0.8)
plot(rstr)

Automatic STR decomposition

Description

Automatically selects parameters for an STR decomposition of time series data.

If a parallel backend is registered for use before STR call, STR will use it for n-fold cross validation computations.

Usage

STR(
  data,
  predictors,
  confidence = NULL,
  robust = FALSE,
  lambdas = NULL,
  pattern = extractPattern(predictors),
  nFold = 5,
  reltol = 0.005,
  gapCV = 1,
  solver = c("Matrix", "cholesky"),
  nMCIter = 100,
  control = list(nnzlmax = 1e+06, nsubmax = 3e+05, tmpmax = 50000),
  trace = FALSE,
  iterControl = list(maxiter = 20, tol = 1e-06)
)

Arguments

data

Time series or a vector of length L.

predictors

List of predictors.
According to the paradigm of this implementation, the trend, the seasonal components, the flexible predictors and the seasonal predictors are all presented in the same form (as predictors) and must be described in this list.
Every predictor is a list of the following structures:

  • data – vector of length L (length of input data, see above). For trend or for a seasonal component it is a vector of ones. For a flexible or a seasonal predictor it is a vector of the predictor's data.

  • times – vector of length L of times of observations.

  • seasons – vector of length L. It is a vector of ones for a trend or a flexible predictor. It is vector assigning seasons to every observation (for a seasonal component or a seasonal predictor). Seasons can be fractional for observations in between seasons.

  • timeKnots – vector of times (time knots) where knots are positioned (for a seasonal component or a seasonal predictor a few knots have the same time; every knot is represented by time and season). Usually this vector coincides with times vector described above, or timeKnots is a subset of times vector.

  • seasonalStructure – describes seasonal topology (which can have complex structure) and seasonal knots.The seasonal topology is described by a list of segments and seasonal knots, which are positioned inside the segments, on borders of the segments or, when they are on on borders, they can connect two or more segments.
    seasonalStructure is a list of two elements:

    • segments – a list of vectors representing segments. Each vector must contain two ordered real values which represent left and right borders of a segment. Segments should not intersect (inside same predictor).

    • sKnots – a list of real values (vectors of length one) or vectors of lengths two or greater (seasonal knots) defining seasons of the knots (every knot is represented by time and season). All real values must belong (be inside or on border of) segments listed in segments. If a few values represent a single seasonal knot then all these values must be on borders of some segments (or a single segment). In this case they represent a seasonal knot which connects a few segments (or both sides of one segment).

  • lambdas – a vector with three values representing lambda (smoothing) parameters (time-time, season-season, time-season flexibility parameters) for this predictor.

confidence

A vector of percentiles giving the coverage of confidence intervals. It must be greater than 0 and less than 1. If NULL, no confidence intervals are produced.

robust

When TRUE, Robust STR decomposition is used. Default is FALSE.

lambdas

An optional parameter. A structure which replaces lambda parameters provided with predictors. It is used as either a starting point for the optimisation of parameters or as the exact model parameters.

pattern

An optional parameter which has the same structure as lambdas although with a different meaning. All zero values correspond to lambda (smoothing) parameters which will not be estimated.

nFold

An optional parameter setting the number of folds for cross validation.

reltol

An optional parameter which is passed directly to optim() when optimising the parameters of the model.

gapCV

An optional parameter defining the length of the sequence of skipped values in the cross validation procedure.

solver

A vector with two string values. The only supported combinations are: c("Matrix", "cholesky") (default), and c("Matrix", "qr"). The parameter is used to specify a particular library and method to solve the minimisation problem during STR decompositon.

nMCIter

Number of Monte Carlo iterations used to estimate confidence intervals for Robust STR decomposition.

control

Passed directly to rq.fit.sfn() during Robust STR decomposition.

trace

When TRUE, tracing is turned on.

iterControl

Control parameters for some experimental features. This should not be used by an ordinary user.

Value

A structure containing input and output data. It is an S3 class STR, which is a list with the following components:

Author(s)

Alexander Dokumentov

References

Dokumentov, A., and Hyndman, R.J. (2022) STR: Seasonal-Trend decomposition using Regression, INFORMS Journal on Data Science, 1(1), 50-62. https://robjhyndman.com/publications/str/

See Also

STRmodel RSTRmodel AutoSTR

Examples


TrendSeasonalStructure <- list(
  segments = list(c(0, 1)),
  sKnots = list(c(1, 0))
)
WDSeasonalStructure <- list(
  segments = list(c(0, 48), c(100, 148)),
  sKnots = c(as.list(c(1:47, 101:147)), list(c(0, 48, 100, 148)))
)

TrendSeasons <- rep(1, nrow(electricity))
WDSeasons <- as.vector(electricity[, "WorkingDaySeasonality"])

Data <- as.vector(electricity[, "Consumption"])
Times <- as.vector(electricity[, "Time"])
TempM <- as.vector(electricity[, "Temperature"])
TempM2 <- TempM^2

TrendTimeKnots <- seq(from = head(Times, 1), to = tail(Times, 1), length.out = 116)
SeasonTimeKnots <- seq(from = head(Times, 1), to = tail(Times, 1), length.out = 24)

TrendData <- rep(1, length(Times))
SeasonData <- rep(1, length(Times))

Trend <- list(
  name = "Trend",
  data = TrendData,
  times = Times,
  seasons = TrendSeasons,
  timeKnots = TrendTimeKnots,
  seasonalStructure = TrendSeasonalStructure,
  lambdas = c(1500, 0, 0)
)
WDSeason <- list(
  name = "Dayly seas",
  data = SeasonData,
  times = Times,
  seasons = WDSeasons,
  timeKnots = SeasonTimeKnots,
  seasonalStructure = WDSeasonalStructure,
  lambdas = c(0.003, 0, 240)
)
StaticTempM <- list(
  name = "Temp Mel",
  data = TempM,
  times = Times,
  seasons = NULL,
  timeKnots = NULL,
  seasonalStructure = NULL,
  lambdas = c(0, 0, 0)
)
StaticTempM2 <- list(
  name = "Temp Mel^2",
  data = TempM2,
  times = Times,
  seasons = NULL,
  timeKnots = NULL,
  seasonalStructure = NULL,
  lambdas = c(0, 0, 0)
)
Predictors <- list(Trend, WDSeason, StaticTempM, StaticTempM2)

elec.fit <- STR(
  data = Data,
  predictors = Predictors,
  gapCV = 48 * 7
)

plot(elec.fit,
  xTime = as.Date("2000-01-11") + ((Times - 1) / 48 - 10),
  forecastPanels = NULL
)

#########################################################

n <- 70
trendSeasonalStructure <- list(segments = list(c(0, 1)), sKnots = list(c(1, 0)))
ns <- 5
seasonalStructure <- list(
  segments = list(c(0, ns)),
  sKnots = c(as.list(1:(ns - 1)), list(c(ns, 0)))
)
seasons <- (0:(n - 1)) %% ns + 1
trendSeasons <- rep(1, length(seasons))
times <- seq_along(seasons)
data <- seasons + times / 4
set.seed(1234567890)
data <- data + rnorm(length(data), 0, 0.2)
data[20] <- data[20] + 3
data[50] <- data[50] - 5
plot(times, data, type = "l")
timeKnots <- times
trendData <- rep(1, n)
seasonData <- rep(1, n)
trend <- list(
  data = trendData, times = times, seasons = trendSeasons,
  timeKnots = timeKnots, seasonalStructure = trendSeasonalStructure, lambdas = c(1, 0, 0)
)
season <- list(
  data = seasonData, times = times, seasons = seasons,
  timeKnots = timeKnots, seasonalStructure = seasonalStructure, lambdas = c(1, 0, 1)
)
predictors <- list(trend, season)
rstr <- STR(data, predictors,
  reltol = 0.0000001, gapCV = 10,
  confidence = 0.95, nMCIter = 400, robust = TRUE
)
plot(rstr)

STR decomposition

Description

Seasonal-Trend decomposition of time series data using Regression.

Usage

STRmodel(
  data,
  predictors = NULL,
  strDesign = NULL,
  lambdas = NULL,
  confidence = NULL,
  solver = c("Matrix", "cholesky"),
  reportDimensionsOnly = FALSE,
  trace = FALSE
)

Arguments

data

Time series or a vector of length L.

predictors

List of predictors.
According to the paradigm of this implementation, the trend, the seasonal components, the flexible predictors and the seasonal predictors are all presented in the same form (as predictors) and must be described in this list.
Every predictor is a list of the following structures:

  • data – vector of length L (length of input data, see above). For trend or for a seasonal component it is a vector of ones. For a flexible or a seasonal predictor it is a vector of the predictor's data.

  • times – vector of length L of times of observations.

  • seasons – vector of length L. It is a vector of ones for a trend or a flexible predictor. It is vector assigning seasons to every observation (for a seasonal component or a seasonal predictor). Seasons can be fractional for observations in between seasons.

  • timeKnots – vector of times (time knots) where knots are positioned (for a seasonal component or a seasonal predictor a few knots have the same time; every knot is represented by time and season). Usually this vector coincides with times vector described above, or timeKnots is a subset of times vector.

  • seasonalStructure – describes seasonal topology (which can have complex structure) and seasonal knots.The seasonal topology is described by a list of segments and seasonal knots, which are positioned inside the segments, on borders of the segments or, when they are on on borders, they can connect two or more segments.
    seasonalStructure is a list of two elements:

    • segments – a list of vectors representing segments. Each vector must contain two ordered real values which represent left and right borders of a segment. Segments should not intersect (inside same predictor).

    • sKnots – a list of real values (vectors of length one) or vectors of lengths two or greater (seasonal knots) defining seasons of the knots (every knot is represented by time and season). All real values must belong (be inside or on border of) segments listed in segments. If a few values represent a single seasonal knot then all these values must be on borders of some segments (or a single segment). In this case they represent a seasonal knot which connects a few segments (or both sides of one segment).

  • lambdas – a vector with three values representing lambda (smoothing) parameters (time-time, season-season, time-season flexibility parameters) for this predictor.

strDesign

An optional parameter used to create the design matrix. It is used internally in the library to improve performance when the design matrix does not require full recalculation.

lambdas

An optional parameter. A structure which replaces lambda parameters provided with predictors. It is used as either a starting point for the optimisation of parameters or as the exact model parameters.

confidence

A vector of percentiles giving the coverage of confidence intervals. It must be greater than 0 and less than 1. If NULL, no confidence intervals are produced.

solver

A vector with two string values. The only supported combinations are: c("Matrix", "cholesky") (default), and c("Matrix", "qr"). The parameter is used to specify a particular library and method to solve the minimisation problem during STR decompositon.

reportDimensionsOnly

A boolean parameter. When TRUE the method constructs the design matrix and reports its dimensions without proceeding further. It is mostly used for debugging.

trace

When TRUE, tracing is turned on.

Value

A structure containing input and output data. It is an S3 class STR, which is a list with the following components:

Author(s)

Alexander Dokumentov

References

Dokumentov, A., and Hyndman, R.J. (2022) STR: Seasonal-Trend decomposition using Regression, INFORMS Journal on Data Science, 1(1), 50-62. https://robjhyndman.com/publications/str/

See Also

AutoSTR

Examples

n <- 50
trendSeasonalStructure <- list(segments = list(c(0, 1)), sKnots = list(c(1, 0)))
ns <- 5
seasonalStructure <- list(
  segments = list(c(0, ns)),
  sKnots = c(as.list(1:(ns - 1)), list(c(ns, 0)))
)
seasons <- (0:(n - 1)) %% ns + 1
trendSeasons <- rep(1, length(seasons))
times <- seq_along(seasons)
data <- seasons + times / 4
plot(times, data, type = "l")
timeKnots <- times
trendData <- rep(1, n)
seasonData <- rep(1, n)
trend <- list(
  data = trendData, times = times, seasons = trendSeasons,
  timeKnots = timeKnots, seasonalStructure = trendSeasonalStructure, lambdas = c(1, 0, 0)
)
season <- list(
  data = seasonData, times = times, seasons = seasons,
  timeKnots = timeKnots, seasonalStructure = seasonalStructure, lambdas = c(10, 0, 0)
)
predictors <- list(trend, season)

str1 <- STRmodel(data, predictors)
plot(str1)

data[c(3, 4, 7, 20, 24, 29, 35, 37, 45)] <- NA
plot(times, data, type = "l")
str2 <- STRmodel(data, predictors)
plot(str2)

Number of phone calls dataset

Description

Number of call arrivals per 5-minute interval handled on weekdays between 7:00 am and 9:05 pm from March 3, 2003 in a large North American commercial bank.

Usage

calls

Format

A numerical time series of class msts and ts.

Source

Data file

References

Forecasting time series with complex seasonal patterns using exponential smoothing A.M. De Livera, R.J. Hyndman & R.D. Snyder J American Statistical Association, 106(496), 1513-1527. https://robjhyndman.com/publications/complex-seasonality/

Examples

plot(calls, ylab = "Calls handled")

Extract STR components

Description

components extracts components as time series from the result of an STR decomposition.

Usage

components(object)

Arguments

object

Result of STR decomposition.

Author(s)

Alexander Dokumentov

See Also

STRmodel, RSTRmodel, STR, AutoSTR

Examples


fit <- AutoSTR(log(grocery))
comp <- components(fit)
plot(comp)

Common argument

Description

Common argument

Arguments

confidence

A vector of percentiles giving the coverage of confidence intervals. It must be greater than 0 and less than 1. If NULL, no confidence intervals are produced.

Common argument

Description

Common argument

Arguments

control

Passed directly to rq.fit.sfn() during Robust STR decomposition.

Common argument

Description

Common argument

Arguments

data

Time series or a vector of length L.

Electricity consumption dataset

Description

The data set provides information about electricity consumption in Victoria, Australia during the 115 days starting on 10th of January, 2000, and comprises the maximum electricity demand in Victoria during 30-minute periods (48 observations per day). For each 30-minute period, the dataset also provides the air temperature in Melbourne.

Usage

electricity

Format

An numerical matrix of class msts and ts.

Details

Examples

plot(electricity[, 1:2],
  xlab = "Weeks",
  main = "Electricity demand and temperature in Melbourne, Australia"
)

Common argument

Description

Common argument

Arguments

gapCV

An optional parameter defining the length of the sequence of skipped values in the cross validation procedure.

Grocery and supermarkets turnover

Description

Turnover of supermarkets and grocery stores in New South Wales, Australia.

Usage

grocery

Format

An object of class ts.

References

Australian Bureau of Statistics, CAT 8501.0. (TABLE 11. Retail Turnover, State by Industry Subgroup, Original)

Examples

plot(grocery, ylab = "NSW Grocery, $ 10^6")

Automatic STR decomposition with heuristic search of the parameters

Description

Automatically selects parameters (lambda coefficients) for an STR decomposition of time series data. Heuristic approach can give a better estimate compare to a standard optmisaton methods used in STR.

If a parallel backend is registered for use before STR call, heuristicSTR will use it for n-fold cross validation computations.

Usage

heuristicSTR(
  data,
  predictors,
  confidence = NULL,
  lambdas = NULL,
  pattern = extractPattern(predictors),
  nFold = 5,
  reltol = 0.005,
  gapCV = 1,
  solver = c("Matrix", "cholesky"),
  trace = FALSE,
  ratioGap = 1e+12,
  relCV = 0.01
)

Arguments

data

Time series or a vector of length L.

predictors

List of predictors.
According to the paradigm of this implementation, the trend, the seasonal components, the flexible predictors and the seasonal predictors are all presented in the same form (as predictors) and must be described in this list.
Every predictor is a list of the following structures:

  • data – vector of length L (length of input data, see above). For trend or for a seasonal component it is a vector of ones. For a flexible or a seasonal predictor it is a vector of the predictor's data.

  • times – vector of length L of times of observations.

  • seasons – vector of length L. It is a vector of ones for a trend or a flexible predictor. It is vector assigning seasons to every observation (for a seasonal component or a seasonal predictor). Seasons can be fractional for observations in between seasons.

  • timeKnots – vector of times (time knots) where knots are positioned (for a seasonal component or a seasonal predictor a few knots have the same time; every knot is represented by time and season). Usually this vector coincides with times vector described above, or timeKnots is a subset of times vector.

  • seasonalStructure – describes seasonal topology (which can have complex structure) and seasonal knots.The seasonal topology is described by a list of segments and seasonal knots, which are positioned inside the segments, on borders of the segments or, when they are on on borders, they can connect two or more segments.
    seasonalStructure is a list of two elements:

    • segments – a list of vectors representing segments. Each vector must contain two ordered real values which represent left and right borders of a segment. Segments should not intersect (inside same predictor).

    • sKnots – a list of real values (vectors of length one) or vectors of lengths two or greater (seasonal knots) defining seasons of the knots (every knot is represented by time and season). All real values must belong (be inside or on border of) segments listed in segments. If a few values represent a single seasonal knot then all these values must be on borders of some segments (or a single segment). In this case they represent a seasonal knot which connects a few segments (or both sides of one segment).

  • lambdas – a vector with three values representing lambda (smoothing) parameters (time-time, season-season, time-season flexibility parameters) for this predictor.

confidence

A vector of percentiles giving the coverage of confidence intervals. It must be greater than 0 and less than 1. If NULL, no confidence intervals are produced.

lambdas

An optional parameter. A structure which replaces lambda parameters provided with predictors. It is used as either a starting point for the optimisation of parameters or as the exact model parameters.

pattern

An optional parameter which has the same structure as lambdas although with a different meaning. All zero values correspond to lambda (smoothing) parameters which will not be estimated.

nFold

An optional parameter setting the number of folds for cross validation.

reltol

An optional parameter which is passed directly to optim() when optimising the parameters of the model.

gapCV

An optional parameter defining the length of the sequence of skipped values in the cross validation procedure.

solver

A vector with two string values. The only supported combinations are: c("Matrix", "cholesky") (default), and c("Matrix", "qr"). The parameter is used to specify a particular library and method to solve the minimisation problem during STR decompositon.

trace

When TRUE, tracing is turned on.

ratioGap

Ratio to define hyperparameter bounds for one-dimensional search.

relCV

Minimum improvement required after all predictors tried. It is used to exit heuristic serach of lambda parameters.

Value

A structure containing input and output data. It is an S3 class STR, which is a list with the following components:

Author(s)

Alexander Dokumentov

References

Dokumentov, A., and Hyndman, R.J. (2022) STR: Seasonal-Trend decomposition using Regression, INFORMS Journal on Data Science, 1(1), 50-62. https://robjhyndman.com/publications/str/

See Also

STR STRmodel AutoSTR

Examples


TrendSeasonalStructure <- list(
  segments = list(c(0, 1)),
  sKnots = list(c(1, 0))
)
WDSeasonalStructure <- list(
  segments = list(c(0, 48), c(100, 148)),
  sKnots = c(as.list(c(1:47, 101:147)), list(c(0, 48, 100, 148)))
)

TrendSeasons <- rep(1, nrow(electricity))
WDSeasons <- as.vector(electricity[, "WorkingDaySeasonality"])

Data <- as.vector(electricity[, "Consumption"])
Times <- as.vector(electricity[, "Time"])
TempM <- as.vector(electricity[, "Temperature"])
TempM2 <- TempM^2

TrendTimeKnots <- seq(from = head(Times, 1), to = tail(Times, 1), length.out = 116)
SeasonTimeKnots <- seq(from = head(Times, 1), to = tail(Times, 1), length.out = 24)

TrendData <- rep(1, length(Times))
SeasonData <- rep(1, length(Times))

Trend <- list(
  name = "Trend",
  data = TrendData,
  times = Times,
  seasons = TrendSeasons,
  timeKnots = TrendTimeKnots,
  seasonalStructure = TrendSeasonalStructure,
  lambdas = c(1500, 0, 0)
)
WDSeason <- list(
  name = "Dayly seas",
  data = SeasonData,
  times = Times,
  seasons = WDSeasons,
  timeKnots = SeasonTimeKnots,
  seasonalStructure = WDSeasonalStructure,
  lambdas = c(0.003, 0, 240)
)
StaticTempM <- list(
  name = "Temp Mel",
  data = TempM,
  times = Times,
  seasons = NULL,
  timeKnots = NULL,
  seasonalStructure = NULL,
  lambdas = c(0, 0, 0)
)
StaticTempM2 <- list(
  name = "Temp Mel^2",
  data = TempM2,
  times = Times,
  seasons = NULL,
  timeKnots = NULL,
  seasonalStructure = NULL,
  lambdas = c(0, 0, 0)
)
Predictors <- list(Trend, WDSeason, StaticTempM, StaticTempM2)

elec.fit <- heuristicSTR(
  data = Data,
  predictors = Predictors,
  gapCV = 48 * 7
)

plot(elec.fit,
  xTime = as.Date("2000-01-11") + ((Times - 1) / 48 - 10),
  forecastPanels = NULL
)

########################################

TrendSeasonalStructure <- list(
  segments = list(c(0, 1)),
  sKnots = list(c(1, 0))
)
DailySeasonalStructure <- list(
  segments = list(c(0, 48)),
  sKnots = c(as.list(1:47), list(c(48, 0)))
)
WeeklySeasonalStructure <- list(
  segments = list(c(0, 336)),
  sKnots = c(as.list(seq(4, 332, 4)), list(c(336, 0)))
)
WDSeasonalStructure <- list(
  segments = list(c(0, 48), c(100, 148)),
  sKnots = c(as.list(c(1:47, 101:147)), list(c(0, 48, 100, 148)))
)

TrendSeasons <- rep(1, nrow(electricity))
DailySeasons <- as.vector(electricity[, "DailySeasonality"])
WeeklySeasons <- as.vector(electricity[, "WeeklySeasonality"])
WDSeasons <- as.vector(electricity[, "WorkingDaySeasonality"])

Data <- as.vector(electricity[, "Consumption"])
Times <- as.vector(electricity[, "Time"])
TempM <- as.vector(electricity[, "Temperature"])
TempM2 <- TempM^2

TrendTimeKnots <- seq(from = head(Times, 1), to = tail(Times, 1), length.out = 116)
SeasonTimeKnots <- seq(from = head(Times, 1), to = tail(Times, 1), length.out = 24)
SeasonTimeKnots2 <- seq(from = head(Times, 1), to = tail(Times, 1), length.out = 12)

TrendData <- rep(1, length(Times))
SeasonData <- rep(1, length(Times))

Trend <- list(
  name = "Trend",
  data = TrendData,
  times = Times,
  seasons = TrendSeasons,
  timeKnots = TrendTimeKnots,
  seasonalStructure = TrendSeasonalStructure,
  lambdas = c(1500, 0, 0)
)
WSeason <- list(
  name = "Weekly seas",
  data = SeasonData,
  times = Times,
  seasons = WeeklySeasons,
  timeKnots = SeasonTimeKnots2,
  seasonalStructure = WeeklySeasonalStructure,
  lambdas = c(0.8, 0.6, 100)
)
WDSeason <- list(
  name = "Dayly seas",
  data = SeasonData,
  times = Times,
  seasons = WDSeasons,
  timeKnots = SeasonTimeKnots,
  seasonalStructure = WDSeasonalStructure,
  lambdas = c(0.003, 0, 240)
)
TrendTempM <- list(
  name = "Trend temp Mel",
  data = TempM,
  times = Times,
  seasons = TrendSeasons,
  timeKnots = TrendTimeKnots,
  seasonalStructure = TrendSeasonalStructure,
  lambdas = c(1e7, 0, 0)
)
TrendTempM2 <- list(
  name = "Trend temp Mel^2",
  data = TempM2,
  times = Times,
  seasons = TrendSeasons,
  timeKnots = TrendTimeKnots,
  seasonalStructure = TrendSeasonalStructure,
  lambdas = c(0.01, 0, 0)
) # Starting parameter is too far from the optimal value
Predictors <- list(Trend, WSeason, WDSeason, TrendTempM, TrendTempM2)

elec.fit <- heuristicSTR(
  data = Data,
  predictors = Predictors,
  gapCV = 48 * 7
)

plot(elec.fit,
  xTime = as.Date("2000-01-11") + ((Times - 1) / 48 - 10),
  forecastPanels = NULL
)

plotBeta(elec.fit, predictorN = 4)
plotBeta(elec.fit, predictorN = 5)

Common argument

Description

Common argument

Arguments

lambdas

An optional parameter. A structure which replaces lambda parameters provided with predictors. It is used as either a starting point for the optimisation of parameters or as the exact model parameters.

Common argument

Description

Common argument

Arguments

nFold

An optional parameter setting the number of folds for cross validation.

Common argument

Description

Common argument

Arguments

nMCIter

Number of Monte Carlo iterations used to estimate confidence intervals for Robust STR decomposition.

Common argument

Description

Common argument

Arguments

pattern

An optional parameter which has the same structure as lambdas although with a different meaning. All zero values correspond to lambda (smoothing) parameters which will not be estimated.

Plots the results of decomposition

Description

plot.STR plots results of STR decomposition.

Usage

## S3 method for class 'STR'
plot(
  x,
  xTime = NULL,
  dataPanels = 1,
  predictorPanels = as.list(seq_along(x$output$predictors)),
  randomPanels = length(x$output$predictors) + 1,
  forecastPanels = length(x$output$predictors) + 2,
  dataColor = "black",
  predictorColors = rep("red", length(x$output$predictors)),
  randomColor = "red",
  forecastColor = "blue",
  vLines = NULL,
  xlab = "Time",
  main = ifelse(x$method %in% c("STR", "STRmodel"), "STR decomposition",
    "Robust STR decomposition"),
  showLegend = TRUE,
  ...
)

Arguments

x

Result of STR decomposition.

xTime

Times for data to plot.

dataPanels

Vector of panel numbers in which to plot the original data. Set to NULL to not show data.

predictorPanels

A list of vectors of numbers where every such vector describes which panels should be used for plotting the corresponding predictor.

randomPanels

Vector of panel numbers in which to plot the residuals. Set to NULL to not show residuals.

forecastPanels

Vector of panel numbers in which to plot the fit/forecast. Set to NULL to not show forecasts.

dataColor

Color to plot data.

predictorColors

Vector of colors to plot components corresponding to the predictors.

randomColor

Color to plot the residuals.

forecastColor

Color to plot the fit/forecast.

vLines

Vector of times where vertical lines will be plotted.

xlab

Label for horizontal axis.

main

Main heading for plot.

showLegend

When TRUE (default) legend is shown at top of plot.

...

Other parameters to be passed directly to plot and lines functions in the implementation.

Author(s)

Alexander Dokumentov

See Also

STRmodel, RSTRmodel, STR, AutoSTR

Examples


fit <- AutoSTR(log(grocery))
plot(fit, forecastPanels = 0, randomColor = "DarkGreen", vLines = 2000:2010, lwd = 2)

Plots the varying beta coefficients of decomposition

Description

plotBeta plots the varying beta coefficients of STR decomposition. It plots coefficients only only for independent seasons (one less season than defined).

Usage

plotBeta(
  x,
  xTime = NULL,
  predictorN = 1,
  dim = c(1, 2, 3),
  type = "o",
  pch = 20,
  palette = function(n) rainbow(n, start = 0, end = 0.7)
)

Arguments

x

Result of STR decomposition.

xTime

Times for data to plot.

predictorN

Predictor number in the decomposition to plot the corresponding beta coefficiets.

dim

Dimensions to use to plot the beta coefficients. When 1, the standard charts are used. When 2, graphics:::filled.contour function is used. When 3, rgl:::persp3d is used. The default value is 1.

type

Type of the graph for one dimensional plots.

pch

Symbol code to plot points in 1-dimensional charts. Default value is 20.

palette

Color palette for 2 - and 3 - dimentional plots.

Author(s)

Alexander Dokumentov

See Also

plot.STR

Examples


fit <- AutoSTR(log(grocery))
for (i in 1:2) plotBeta(fit, predictorN = i, dim = 2)

########################################

TrendSeasonalStructure <- list(
  segments = list(c(0, 1)),
  sKnots = list(c(1, 0))
)
DailySeasonalStructure <- list(
  segments = list(c(0, 48)),
  sKnots = c(as.list(1:47), list(c(48, 0)))
)
WeeklySeasonalStructure <- list(
  segments = list(c(0, 336)),
  sKnots = c(as.list(seq(4, 332, 4)), list(c(336, 0)))
)
WDSeasonalStructure <- list(
  segments = list(c(0, 48), c(100, 148)),
  sKnots = c(as.list(c(1:47, 101:147)), list(c(0, 48, 100, 148)))
)

TrendSeasons <- rep(1, nrow(electricity))
DailySeasons <- as.vector(electricity[, "DailySeasonality"])
WeeklySeasons <- as.vector(electricity[, "WeeklySeasonality"])
WDSeasons <- as.vector(electricity[, "WorkingDaySeasonality"])

Data <- as.vector(electricity[, "Consumption"])
Times <- as.vector(electricity[, "Time"])
TempM <- as.vector(electricity[, "Temperature"])
TempM2 <- TempM^2

TrendTimeKnots <- seq(from = head(Times, 1), to = tail(Times, 1), length.out = 116)
SeasonTimeKnots <- seq(from = head(Times, 1), to = tail(Times, 1), length.out = 24)
SeasonTimeKnots2 <- seq(from = head(Times, 1), to = tail(Times, 1), length.out = 12)

TrendData <- rep(1, length(Times))
SeasonData <- rep(1, length(Times))

Trend <- list(
  name = "Trend",
  data = TrendData,
  times = Times,
  seasons = TrendSeasons,
  timeKnots = TrendTimeKnots,
  seasonalStructure = TrendSeasonalStructure,
  lambdas = c(1500, 0, 0)
)
WSeason <- list(
  name = "Weekly seas",
  data = SeasonData,
  times = Times,
  seasons = WeeklySeasons,
  timeKnots = SeasonTimeKnots2,
  seasonalStructure = WeeklySeasonalStructure,
  lambdas = c(0.8, 0.6, 100)
)
WDSeason <- list(
  name = "Dayly seas",
  data = SeasonData,
  times = Times,
  seasons = WDSeasons,
  timeKnots = SeasonTimeKnots,
  seasonalStructure = WDSeasonalStructure,
  lambdas = c(0.003, 0, 240)
)
TrendTempM <- list(
  name = "Trend temp Mel",
  data = TempM,
  times = Times,
  seasons = TrendSeasons,
  timeKnots = TrendTimeKnots,
  seasonalStructure = TrendSeasonalStructure,
  lambdas = c(1e7, 0, 0)
)
TrendTempM2 <- list(
  name = "Trend temp Mel^2",
  data = TempM2,
  times = Times,
  seasons = TrendSeasons,
  timeKnots = TrendTimeKnots,
  seasonalStructure = TrendSeasonalStructure,
  lambdas = c(0.01, 0, 0)
) # Starting parameter is too far from the optimal value
Predictors <- list(Trend, WSeason, WDSeason, TrendTempM, TrendTempM2)

elec.fit <- STR(
  data = Data,
  predictors = Predictors,
  gapCV = 48 * 7
)

plot(elec.fit,
  xTime = as.Date("2000-01-11") + ((Times - 1) / 48 - 10),
  forecastPanels = NULL
)

plotBeta(elec.fit, predictorN = 4)
plotBeta(elec.fit, predictorN = 5) # Beta coefficients are too "wiggly"

Common argument

Description

Common argument

Arguments

predictors

List of predictors.
According to the paradigm of this implementation, the trend, the seasonal components, the flexible predictors and the seasonal predictors are all presented in the same form (as predictors) and must be described in this list.
Every predictor is a list of the following structures:

  • data – vector of length L (length of input data, see above). For trend or for a seasonal component it is a vector of ones. For a flexible or a seasonal predictor it is a vector of the predictor's data.

  • times – vector of length L of times of observations.

  • seasons – vector of length L. It is a vector of ones for a trend or a flexible predictor. It is vector assigning seasons to every observation (for a seasonal component or a seasonal predictor). Seasons can be fractional for observations in between seasons.

  • timeKnots – vector of times (time knots) where knots are positioned (for a seasonal component or a seasonal predictor a few knots have the same time; every knot is represented by time and season). Usually this vector coincides with times vector described above, or timeKnots is a subset of times vector.

  • seasonalStructure – describes seasonal topology (which can have complex structure) and seasonal knots.The seasonal topology is described by a list of segments and seasonal knots, which are positioned inside the segments, on borders of the segments or, when they are on on borders, they can connect two or more segments.
    seasonalStructure is a list of two elements:

    • segments – a list of vectors representing segments. Each vector must contain two ordered real values which represent left and right borders of a segment. Segments should not intersect (inside same predictor).

    • sKnots – a list of real values (vectors of length one) or vectors of lengths two or greater (seasonal knots) defining seasons of the knots (every knot is represented by time and season). All real values must belong (be inside or on border of) segments listed in segments. If a few values represent a single seasonal knot then all these values must be on borders of some segments (or a single segment). In this case they represent a seasonal knot which connects a few segments (or both sides of one segment).

  • lambdas – a vector with three values representing lambda (smoothing) parameters (time-time, season-season, time-season flexibility parameters) for this predictor.

Objects exported from other packages

Description

These objects are imported from other packages. Follow the links below to see their documentation.

forecast

seasadj

Common argument

Description

Common argument

Arguments

reltol

An optional parameter which is passed directly to optim() when optimising the parameters of the model.

Common argument

Description

Common argument

Arguments

reportDimensionsOnly

A boolean parameter. When TRUE the method constructs the design matrix and reports its dimensions without proceeding further. It is mostly used for debugging.

Common argument

Description

Common argument

Arguments

robust

When TRUE, Robust STR decomposition is used. Default is FALSE.

Seasonal adjustment based on STR

Description

seasadj.STR extracts seasonally adjusted data by removing the seasonal components from the result of STR decomposition.

Usage

## S3 method for class 'STR'
seasadj(object, include = c("Trend", "Random"), ...)

Arguments

object

Result of STR decomposition.

include

Vector of component names to include in the result. The default is c("Trend", "Random").

...

Other arguments not currently used.

Author(s)

Alexander Dokumentov

See Also

STRmodel, RSTRmodel, STR, AutoSTR

Examples


fit <- AutoSTR(log(grocery))
plot(seasadj(fit))

Common argument

Description

Common argument

Arguments

solver

A vector with two string values. The only supported combinations are: c("Matrix", "cholesky") (default), and c("Matrix", "qr"). The parameter is used to specify a particular library and method to solve the minimisation problem during STR decompositon.

Common argument

Description

Common argument

Arguments

strDesign

An optional parameter used to create the design matrix. It is used internally in the library to improve performance when the design matrix does not require full recalculation.

Common argument

Description

Common argument

Arguments

trace

When TRUE, tracing is turned on.