I was just looking at a machine learning package for Python/R called 'Prophet' by Facebook which is making a bit of noise in the machine learning/data science world due to its simplicity, especially in Python. Here's a summary:

• Time series algorithm, automatically standardises time and predicted variable

• Automatically de-trending on three components -- long term (over the entire data?), monthly/weekly/days of the week, and holidays (comes with holidays data for different regions)

• By default, it fits 25 linear or logistic models over the first 80% time of the data

• By default, assumes Laplace distribution for the covariates

Prior to reading this, it always bothered me when papers don't mention what type of distribution each of the covariates are, unless of course, graphically shown individually (so stuff like Poisson distribution is easier to tell). Now that I'm reading above linked article, it took me by surprise that the package assumes Laplace distribution for all the covariates by default.

So then my question is two-fold. Is it ok to assume normal distribution if the paper doesn't mention anything for each of the covariates (some papers mention them for few to many covariates, and I think some of them may be obvious so no need to mention, though maybe not obvious to all)? On top of this, why would one choose to assume Laplace instead of normal distribution? What would be the advantages/disadvantages of such decision, and what would be the result on the estimates/errors/bias?