RemixAutoML (version 0.11.0)

ResidualOutliers: ResidualOutliers is an automated time series outlier detection function

Description

ResidualOutliers is an automated time series outlier detection function that utilizes tsoutliers and auto.arima. It looks for five types of outliers: "AO" Additive outliter - a singular extreme outlier that surrounding values aren't affected by; "IO" Innovational outlier - Initial outlier with subsequent anomalous values; "LS" Level shift - An initial outlier with subsequent observations being shifted by some constant on average; "TC" Transient change - initial outlier with lingering effects that dissapate exponentially over time; "SLS" Seasonal level shift - similar to level shift but on a seasonal scale.

Usage

ResidualOutliers(data, DateColName = "DateTime",
  TargetColName = "Target", PredictedColName = NULL,
  TimeUnit = "day", maxN = 5, tstat = 2)

Arguments

data

the source residuals data.table

DateColName

The name of your data column to use in reference to the target variable

TargetColName

The name of your target variable column

PredictedColName

The name of your predicted value column. If you supply this, you will run anomaly detection of the difference between the target variable and your predicted value. If you leave PredictedColName NULL then you will run anomaly detection over the target variable.

TimeUnit

The time unit of your date column: hour, day, week, month, quarter, year

maxN

the largest lag or moving average (seasonal too) values for the arima fit

tstat

the t-stat value for tsoutliers

Value

A named list containing FullData = original data.table with outliers data and ARIMA_MODEL = the arima model.

See Also

Other Unsupervised Learning: AutoKMeans, GenTSAnomVars, ProblematicRecords

Examples

# NOT RUN {
data <- data.table::data.table(DateTime = as.Date(Sys.time()),
                               Target = as.numeric(stats::filter(rnorm(1000,
                                                                       mean = 50,
                                                                       sd = 20),
                                                                 filter=rep(1,10),
                                                                 circular=TRUE)))
data[, temp := seq(1:1000)][, DateTime := DateTime - temp][, temp := NULL]
data <- data[order(DateTime)]
data[, Predicted := as.numeric(stats::filter(rnorm(1000,
                                                   mean = 50,
                                                   sd = 20),
                                             filter=rep(1,10),
                                             circular=TRUE))]
stuff <- ResidualOutliers(data = data,
                          DateColName = "DateTime",
                          TargetColName = "Target",
                          PredictedColName = NULL,
                          TimeUnit = "day",
                          maxN = 5,
                          tstat = 4)
data     <- stuff[[1]]
model    <- stuff[[2]]
outliers <- data[type != "<NA>"]
# }