So far, we have assumed that the variable we are modeling behaves like a real number, taking any possible value. This is reflected in the fact that we assume that the current value of the series is equal to the previous value, plus some Gaussian noise. But this is not very useful when we modeling count data, such as the number of clients, or the number of insurance claims, and so on. When these numbers are large, the discreteness of the data is not a huge problem, but when we're modeling events that occur on a small scale, the consequences of ignoring the discreteness are much worse. 

The tscount package allows us to model count time series, if the data follows a Poisson or negative binomial distribution. The framework is rooted in generalized linear models (GLMs), using previous values of the series to predict the current ones.