This is my problem: using R. I need to generate data from a time-series without storing the historical data since they could not be available in some cases, so my objective is to find a generic model from which create a dataset similar (considering the statistical properties and the behaviour) to the starting one and if necessary do some transformations in order to change some statistical properties (but this is not a serious problem).
Until now, I explored different paths but the results were not so good:
- ARIMA MODEL - slow but it gave me a model for each of the three time series's components (trend, noise, seasonality), however since the
arima.simcommand takes randomly, points from a user-defined function, using this strategy, I will lose all the relationships between time stamps and data. The results for the trend and seasonality components seem to be extremely bad and far from what I need, while for the noise it works quite well. - BLOCK BOOTSTRAP - quite useful but I need to store the whole dataset in my generator and for the final version it will not be possible.
- ECDF and Histograms - I found something about the possibility to generate a dataset using one of these two objects but still it's not so clear to me how to do that.
Are there other strategies for creating a model useful for my purpose? Is there something about these three strategies that I didn't take in account?
