Queries Regarding Time Series

[CatchSandeepVaid]

For Time Series using HoltWinters method & WekaForecaster:
a) When we call forecaster.buildForecaster(data, System.out), does it train forecaster or underlying classifier (using classifier.buildClassifier(data))?


b) For saving model, should i save the trained forecaster or classifier (do i have to call classifier.buildClassifier(data) before saving) using
weka.core.SerializationHelper.write()?

i) If i save classifier and use it later using weka.core.SerializationHelper.read(), when i call forecaster.primeForecaster(data); i get the error as:
Exception in thread "main" java.lang.NullPointerException
at weka.filters.supervised.attribute.TSLagMaker.clearLagHistories(TSLagMaker.java:2836)
at weka.classifiers.timeseries.WekaForecaster.primeForecaster(WekaForecaster.java:1250)

ii) If i save forecaster & use it later for forecasting, it seems it doesn't use the pre-saved trained classifier.

[Mark]

If you are forecasting programatically, and don't require the evaluation routines provided by the forecasting environment, then you can use HoltWinters directly rather than via WekaForecaster. Note that HoltWinters uses three seasonal cycles worth of data to estimate parameters and smooth over. There should be no issues serializing the model, but the forecasts that it generates will follow directly from the end of the training data. If you re-load the model and want to apply it to make forecasts following data from a later period in time then it will be necessary to re-train the model using the most recently available three seasons worth of data.

Cheers,
Mark.