It is a big topic but here is a simplistic recipe! The starting point would be to check the distribution of the historical returns. Histogram would give an idea of how the shifts are distributed. Have a look at the tails, if the tails are fat or don’t ‘tail-off’ then that would be indicative of jumps or non constant volatility.
If you decide that a simple arithmetic or geometric Brownian motion is sufficient, then you can analyse the volatility by the level of underlying. If the volatility happened to be proportional to the level of the underlying then geometric brownian would be more appropriate. And if you are worried about fat tails then you can add
Jump to the dynamics or introduce non-constant volatility.
You can also have a look at the historical price series to see if the series is mean reverting, meaning if the price frequently reverts back to some mean which could be non constant, then it would be better to model it as Ornstein Uhlenbeck process.