Abstract

A common way to model nonnegative time series is to apply a log transformation and then use classical ARIMA techniques. We demonstrate using Canadian Fire Weather Index
data that simulating from such models can lead to unrealistic data scenarios.
Minification models 
provide another approach to nonnegative time series, but they can be too restrictive. A random coefficient version of these processes has more flexibility than the fixed coefficient version of the process, but the dependence structure is still not accurate. A Markov chain extension is proposed in this presentation. Through simulations, we demonstrate how it behaves, and we show that maximum likelihood estimation for the model parameters is possible. the Fire Weather Index series.