performanceEstimation (version 1.1.0)

timeseriesWF: A function implementing sliding and growing window standard workflows for time series forecasting tasks

Description

This function implements sliding and growing window workflows for the prediction time series. The sliding window workflow consists of: (i) starting by learning a prediction model based on the given training set, (ii) use this model to obtain predictions for a pre-defined number of future time steps of the test set; (iii) then slide the training window forward this pre-defined number of steps and obtain a new model with this new training set; (iv) use this new model for obtaining another set of predictions; and (v) keep repeting this sliding process until predictions are obtained for all test set period.

The growing window workflow is similar but instead of sliding the training window, we grow this window, so each new set of predictions is obtained with a model learned with all data since the beginning of the training set till the current time step.

Usage

timeseriesWF(form,train,test, learner,learner.pars=NULL, type='slide',relearn.step=1, predictor='predict',predictor.pars=NULL, pre=NULL,pre.pars=NULL, post=NULL,post.pars=NULL, .fullOutput=FALSE,verbose=FALSE)

Arguments

form

A formula specifying the predictive task.

train

A data frame containing the data set to be used for obtaining the first prediction model. In case we are using the sliding window approach, the size of this training set will determine the size of all future training sets after each slide step.

test

A data frame containing the data set for which we want predictions.

learner

A character string with the name of a function that is to be used to obtain the prediction models.

learner.pars

A list of parameter values to be passed to the learner (defaults to NULL).

type

A character string specifying if we are using a sliding (value 'slide') or a growing (value 'grow') window workflow (defaults to 'slide').

relearn.step

The number of time steps (translated into number of rows in the test set) after which a new model is re-learned (either by sliding or growing the training window) (defaults to 1, i.e. each new row).

predictor

A character string with the name of a function that is to be used to obtain the predictions for the test set using the obtained model (defaults to 'predict').

predictor.pars

A list of parameter values to be passed to the predictor (defaults to NULL).

pre

A vector of function names that will be applied in sequence to the train and test data frames, generating new versions, i.e. a sequence of data pre-processing functions.

pre.pars

A named list of parameter values to be passed to the pre-processing functions.

post

A vector of function names that will be applied in sequence to the predictions of the model, generating a new version, i.e. a sequence of data post-processing functions.

post.pars

A named list of parameter values to be passed to the post-processing functions.

.fullOutput

A boolean that if set to TRUE will make the function return more information in the list that results from its execution like for instance the models (defaults to FALSE).

verbose

A Boolean indicating whether a "*" character should be printed every time the window slides (defaults to FALSE).

Value

A list with several components containing the result of runing the workflow.

Details

The main goal of this function is to facilitate the task of the users of the experimental comparison infra-structure provided by function performanceEstimation for time series problems where the target variable can be numeric or nominal. Frequently, users just want to compare existing algorithms or variants of these algorithms on a set of forecasting tasks, using some standard error metrics. The goal of the timeseriesWF function is to facilitate this task by providing a standard workflow for time series tasks.

The function works, and has almost the same parameters, as function standardWF. The help page of this latter function describes most of the parameters used in the current function and thus we will not repeat the description here. The main difference to the standardWF function is on the existance of two extra parameters that control the sliding and growing window approaches to time series forecasting. These are parameters type and relearn.step. We have considered two typical workflow approaches for time series tasks where the user wants predictions for a certain future time period. Both are based on the assumption that after "some" time the model that we have obtained with the given training period data may have become out-dated, and thus a new model should be obtained with the most recent data. The idea is that as we move in the testing period and get predictions for the successive rows of the test set, it is like if a clock is advancing. Previous rows for which we already made a prediction are "past" as we assume that the successive rows in both the train and test data frames are ordered by time (they are time series). In this context, as we move forward in the test period we can regard the rows for which we already made a prediction as past data, and thus potentially useful to be added to our initial training set with the goal of obtaining a fresh new model with more recent data. This type of reasoning only makes sense if we suspect that there is some kind of concept drift on our data. For stationary data this makes no sense and we would be better off using the workflow provided by function standardWF. Still, the current function implements two workflows following this model-updating reasoning: (i) sliding window; and (ii) growing window. Both use the value of the parameter (relearn.step) to decide the number of time periods after which we re-learn the model using fresh new data. The difference between the two strategies lies on how they treat the oldest data (the initial rows of the provided training set). Sliding window, as the name suggests, after each relearn step slides the training set forward thus forgetting the oldest rows of the previous training set whilst incorporating the most recent observations. With this approach all models are obtained with a training set with the same amount of data (the number of rows of the initially given training set). Growing window does not remove older rows and thus the training sets keep growing in size after each relearn step.

References

Torgo, L. (2014) An Infra-Structure for Performance Estimation and Experimental Comparison of Predictive Models in R. arXiv:1412.0436 [cs.MS] http://arxiv.org/abs/1412.0436

See Also

standardWF, performanceEstimation, getIterationsInfo, getIterationsPreds, standardPRE, standardPOST

Examples

## The following is a small illustrative example using the quotes of the
## SP500 index. This example compares two random forests with 500
## regression trees, one applyed in a standard way, and the other using
## a sliding window with a relearn step of every 10 days. The experiment
## uses 10 repetitions of a train+test cycle using 50% of the available
## data for training and 25% for testing.
## Not run: 
# library(quantmod)
# library(randomForest)
# getSymbols('^GSPC',from='2008-01-01',to='2012-12-31')
# data.model <- specifyModel(
#   Next(100*Delt(Ad(GSPC))) ~ Delt(Ad(GSPC),k=1:10)+Delt(Vo(GSPC),k=1:3))
# data <- as.data.frame(modelData(data.model))
# colnames(data)[1] <- 'PercVarClose'
# spExp <- performanceEstimation(
#   PredTask(PercVarClose ~ .,data,'SP500_2012'),
#   c(Workflow(wf='standardWF',wfID="standRF",
#              learner='randomForest',
#              learner.pars=list(ntree=500)),
#     Workflow(wf='timeseriesWF',wfID="slideRF",
#              learner='randomForest',
#              learner.pars=list(ntree=500),
#              type="slide",
#              relearn.step=10)
#    ),
#   EstimationTask(
#      metrics=c("mse","theil"),
#      method=MonteCarlo(nReps=5,szTrain=0.5,szTest=0.25)
#      )
# )
# ## End(Not run)