Imputation (replacement) of missing values in univariate time series. Offers several imputation functions and missing data plots. Available imputation algorithms include: 'Mean', 'LOCF', 'Interpolation', 'Moving Average', 'Seasonal Decomposition', 'Kalman Smoothing on Structural Time Series models', 'Kalman Smoothing on ARIMA models'.
The imputeTS package specializes on (univariate) time series imputation. It offers several different imputation algorithm implementations. Beyond the imputation algorithms the package also provides plotting and printing functions of time series missing data statistics. Additionally three time series datasets for imputation experiments are included.
The imputeTS package can be found on CRAN. For installation execute in R:
If you want to install the latest version from GitHub (can be unstable) run:
To impute (fill all missing values) in a time series x, run the following command:
Output is the time series x with all NA's replaced by reasonable values.
In this case interpolation was the algorithm of choice for calculating the NA replacements. There are several other algorithms (see also under caption "Imputation Algorithms"). All imputation functions are named alike starting with na. followed by a algorithm label e.g. na.mean, na.kalman, ...
To plot missing data statistics for a time series x, run the following command:
This is also just one example for a plot. Overall there are four different types of missing data plots. (see also under caption "Missing Data Plots").
To print statistics about the missing data in a time series x, run the following command:
To load the 'heating' time series (with missing values) into a variable y and the 'heating' time series (without missing values) into a variable z, run:
y <- tsHeating z <- tsHeatingComplete
There are three datasets provided with the package, the 'tsHeating', the 'tsAirgap' and the 'tsNH4' time series. (see also under caption "Datasets").
Here is a table with available algorithms to choose from:
|na.interpolation||Missing Value Imputation by Interpolation|
|na.kalman||Missing Value Imputation by Kalman Smoothing|
|na.locf||Missing Value Imputation by Last Observation Carried Forward|
|na.ma||Missing Value Imputation by Weighted Moving Average|
|na.mean||Missing Value Imputation by Mean Value|
|na.random||Missing Value Imputation by Random Sample|
|na.remove||Remove Missing Values|
|na.replace||Replace Missing Values by a Defined Value|
|na.seadec||Seasonally Decomposed Missing Value Imputation|
|na.seasplit||Seasonally Splitted Missing Value Imputation|
This is a rather broad overview. The functions itself mostly offer more than just one algorithm. For example na.interpolation can be set to linear or spline interpolation.
More detailed information about the algorithms and their options can be found in the imputeTS reference manual.
Here is a table with available plots to choose from:
|plotNA.distribution||Visualize Distribution of Missing Values|
|plotNA.distributionBar||Visualize Distribution of Missing Values (Barplot)|
|plotNA.gapsize||Visualize Distribution of NA gapsizes|
|plotNA.imputations||Visualize Imputed Values|
More detailed information about the plots can be found in the imputeTS reference manual.
There are two datasets (each in two versions) available:
|tsAirgap||Time series of monthly airline passengers (with NAs)|
|tsAirgapComplete||Time series of monthly airline passengers (complete)|
|tsHeating||Time series of a heating systems supply temperature (with NAs)|
|tsHeatingComplete||Time series of a heating systems supply temperature (complete)|
|tsNH4||Time series of NH4 concentration in a wastewater system (with NAs)|
|tsNH4Complete||Time series of NH4 concentration in a wastewater system (complete)|
The tsAirgap, tsHeating and tsNH4 time series are with NAs. Their complete versions are without NAs. Except the missing values their versions are identical. The NAs for the time series were artifically inserted by simulating the missing data pattern observed in similar non-complete time series from the same domain. Having a complete and incomplete version of the same dataset is useful for conducting experiments of imputation functions.
More detailed information about the datasets can be found in the imputeTS reference manual.
You can cite imputeTS the following:
Moritz, Steffen, and Thomas Bartz-Beielstein. "imputeTS: Time Series Missing Value Imputation in R." R Journal 9.1 (2017).
If you found a bug or have suggestions, feel free to get in contact via steffen.moritz10 at gmail.com
All feedback is welcome
Updated Description: Orcid Id added, packages required for unit test add as "Suggested"
Small correction in README.md, small update to citation file
Replaced NEWS with NEWS.md for better formatting
Updated citation file
Minor changes to vignette
Small speed improvments for na.kalman
Improved input check for all functions
Bugfix for unit tests
Changes to unit test (because of zoo update)
Bugfix for na.kalman with integer input
Improved error messages for na.seasplit and na.seadec
Minor vignette changes
Fixed for problems with Solaris/Sparc
Fixes for problems with vignette on osx
Bugfix for plots without missing data
Increased performance for na.locf
Minor bugfixes for specific data.frame inputs
Minor bugfixes for specific xts object inputs
Improved Code Documentation
Added new software tests
Computation time improvments for na.locf (up to 10000 times faster)
Computation time improvments for na.interpolation (up to 10000 times faster)
Computation time improvments for na.kalman (only slightly faster, under 10%)
Fixed unnecessary warning message with some na.kalman options
Adjusted default parameters for plotNA.distributionBar (using nclass.Sturges for breaks parameter)
Fixed issue with too sensitive input checking
Enabled usage of multivariate input (data.frame, mts, matrix,...) for all imputation functions except na.remove. This means users do not have to loop through all columns by themselfes anymore if they want to use the package with multivariate data. The imputation itself is still performend in univariate manner (column after column).
Improved compatibility with different advanced time series objects like zoo and xts. Using the imputation functions with these time series objects should be possible now. These series will not be explicitly named as possible input in the user documentation. Absence of errors can not be guaranteed. However, there are no known issues yet.
Added several things for unit tests with pkg 'testthat'
Added unit tests for every function
Adjusted error messages
Internal Coding style improvement: replaced all T with TRUE and all F with FALSE
Adjustment tsHeating / tsHeatingComplete datasets (set 1440 as frequency parameter)
Adjustment tsNH4 / tsNH4Complete datasets (set 144 as frequency parameter)
Fixes for grammar, spelling and citations in the whole documentation
Revised examples in the documentation for all functions
Restricted output of na.remove to vector only (issue with incorrect time information otherwise)
Added better x-axes labels for plotNA.distribution
Added github links to description file
Added citation file
Updated Readme (badges for travis ci and cran status)
Fix in documentation for na.interpolation (due to outdated descriptions)
Fix in documentation plotNA.distribution / plotNA.distributionBar (due to interchanged descriptions)
Added references to used packages in na.kalman and na.interpolation documentation
Allows now also numeric vectors as input
Removed na.identifier parameter for all functions (too error prone, better handled individually by the user)
Minor changes in na.interpolation with option = "stine"
Added na.ma imputation function
Replaced "data" in all function parameters with the more common "x"
Improvement of all code examples
Renamed heating/heatingComplete dataset to tsHeating/tsHeatingComplete
Renamed nh4/nh4Complete dataset to tsNH4/tsNH4Complete
Added tsAirgap / tsAirgapComplete datasets
Improved imputeTS-package documentation
Added na.kalman imputation function
Added README.md function
Added statsNA function
Added plotNA.gapsize function
Renamed vis.imputations to plotNA.imputations
Renamed vis.barMissing to plotNA.distributionBar
Renamed vis.missing to plotNA.distribution
Fixed issues with parameter pass through and legend for all plotting functions
Improved dataset documentation
Update of vis.differences (better looking plot now)
Added vis.missing to visualize the distribution of missing data in a time series
Added vis.barMissing, which is especially suited to visualize missing data in very huge time series
Update na.interpolate (added Stineman interpolation and enabled ... parameter for all interpolation algorithms to pass through parameters to the underlying functions)
Added two datasets of sensor data
vis.differences for plotting differences between real and imputed values
Removed internal functions from visible package documentation
Added additional algorithms: na.seasplit and na.seadec
internal function for algorithm selection
Created initial version of imputeTS package for univariate time series imputation
added the simple imputation functions: na.locf, na.mean, na.random, na.interpolation, na.replace
added na.remove function for removing all NAs from a time series