best way of univariate prediction for sparse data I have a client who has sparse hourly data (by sparse I mean there are too many hours with 0 calls). I used TBATS in R to forecast hourly data for them. Regardless of the point forecast, the actual values are always in the 80% prediction interval. I wonder if there is any specific method/package in R that is specifically used for uni variate forecasting of sparse data.
Thanks
 A: Try a zero-inflated negative binomial regression model with several lag effects. If you use the mboost or gamboostLSS packages, you can get good regularization along with a zero inflated model. And in the gamboostLSS framework, you can build separate models for the conditional distributions of each of the zero inflated negative binomial parameters. An added bonus of using the gamboostLSS package is that you have the full conditional distributions. With that in hand, you are able to do simulations, and with simulations you can create a probabilistic forecast with an arguably more theoretically justified model of errors than with TBATs or other stuff available in the forecast package.
You could also compare the results of your zero inflated model to just a conditional negative binomial.
Fair warning though. This method will be fairly computationally expensive.
A: this is probably way too late but I was in a similar situation. But what you can try is fabricating data. Take the data you have, get a best fit function, then depending on what you want to do you can either add noise and then populate or just populate on the function. 
I would definitely recommend adding noise. Adding data on the function only is useful in cases where you are working with hard cutoff limits. Example I was doing engine failure prediction for a mining company.
