ts_levels: Detect levels in time series

In FredHasselman/casnet: A Toolbox for Studying Complex Adaptive Systems and Networks

Detect levels in time series

Description

Use recursive partitioning function rpart::rpart() to perform a 'classification' of relatively stable levels in a timeseries.

Usage

ts_levels(
  y,
  minDataSplit = NROW(y)/4,
  minLevelDuration = round(minDataSplit/3),
  changeSensitivity = 0.01,
  maxLevels = floor(NROW(y)/minLevelDuration),
  method = c("anova", "poisson", "class", "exp")[1],
  minChange = NA,
  returnTrends = FALSE,
  minTrendDuration = max(minLevelDuration, 2),
  maxTrends = maxLevels,
  removeTrendsLarger = c(-0.1, 0.1),
  removeTrendsShorter = minTrendDuration,
  doLevelPlot = FALSE,
  doTreePlot = FALSE
)

Arguments

y	A time series of numeric vector
minDataSplit	An integer indicating how many datapoints should be in a segment before it will be analysed for presence of a level or trend change (default = 12)
minLevelDuration	Minimum duration (number of datapoint) of a level (default = round(minDataSplit/3))
changeSensitivity	A number indicating a criterion of change that must occur before declaring a new level. Higher numbers indicate higher levels of change must occur before a new level is considered. For example, if method = "anova", the overall R^2 after a level is introduced must increase by the value of changeSensitivity, see the cp parameter in rpart::rpart.control(). (default = 0.01)
maxLevels	Maximum number of levels in one series (default = floor(max(NROW(y), na.rm = TRUE)/minLevelDuration))
method	The partitioning method to use, see the manual pages of rpart for details.
minChange	After the call to rpart, adjust detected level changes to a minimum absolute change in y. If a level change is smaller than minChange, the previous level will be continued. Note that this is an iterative process starting at the beginning of the series and 'correcting' towards the end. The results are stored in p_adj. Set to NA to skip, which means p_adj will be identical to p (default = sd(y, na.rm = TRUE))
returnTrends	Should stationary trends also be estimated and returned? Arguments method and changeSensitivity will be the same as for level detection and argument minChange will be ignored. (default = FALSE)
minTrendDuration	Required if returnTrends = TRUE, the minimum duration for a trend. (default = max(minLevelDuration, 2))
maxTrends	Maximum number of trends in one series. (default = maxLevels)
removeTrendsLarger	Pass NA to ignore. If there are trends in the data, remove them before estimating stationary levels to improve accuracy. Provide a 2 element vector representing a range of slope values that can reasonably be considered a weakly stationary level. If removeTrendsLarger is not NA it takes precedence over removeTrendShorter. (default = c(-0.1, 0.1))
doLevelPlot	Should a plot with the original series and the levels and/or trends be produced? (default = FALSE)
doTreePlot	Should a plot of the decision tree be produced. If returnTrends = TRUE there will be 2 trees. This requires package partykit (default = FALSE)
removeTrendShorter	Pass NA to ignore. If there are trends in the data, remove them before estimating stationary levels. Provide a trend duration in terms of data points that can reasonably be considered a period of transient behaviour before a new stationary level is reached. If removeTrendsLarger is not NA it takes precedence over removeTrendShorter. (default = 5)

Value

A list object with fields tree and pred. The latter is a data frame with columns x (time), y (the variable of interest) and p the predicted levels in y and p_adj, the levels in p but adjusted for the value passed to minChange.

Author(s)

Fred Hasselman

See Also

Other Time series operations: ts_center(), ts_changeindex(), ts_checkfix(), ts_detrend(), ts_diff(), ts_discrete(), ts_duration(), ts_embed(), ts_integrate(), ts_peaks(), ts_permtest_block(), ts_permtest_transmat(), ts_rasterize(), ts_sd(), ts_slice(), ts_slopes(), ts_standardise(), ts_sumorder(), ts_symbolic(), ts_trimfill(), ts_windower()

Examples

# Levels in white noise?

set.seed(4321)
y <- rnorm(100)
wn <- ts_levels(y)
plot(wn$pred$x,wn$pred$y, type = "l")
lines(wn$pred$p, col = "red3", lwd = 2)

# This is due to the default changeSensitivity of 0.01

wn2 <- ts_levels(y,changeSensitivity = .1)
lines(wn2$pred$p, col = "steelblue", lwd = 2)

FredHasselman/casnet documentation built on May 5, 2024, 9:38 p.m.