11-08-2016 02:17 AM
You don't have to log transform your data and in many cases you should not.
But in some cases this can be convenient. In finance eg, you will often assume that prices are log normally distributed (which may or may not be true), which makes log(1 + r_i) normally distributed. This is neat becasue much of classic statistics assume normality.
11-08-2016 04:58 AM
Why: Because a time series has to be stationary (=time-invariant) to be modeled (with ARIMA). Stationarity includes variance stationarity.
Why log: Because it is easy and it is often able to stabilize the variance.
11-08-2016 10:47 AM
Sometimes to reduce the effect of valid extreme values, the log will have smaller range and the transform less likely to have "excessive" influence on a model.
11-08-2016 02:34 PM
Yet another - if you data contains an exponential growth trend then a log transformation will turn this into a straight line trend which can be easier to model.