Should ARIMA models be used to model stock price changes or returns?

I'm simulating price paths using ARIMA models. Is it best practice to model stock prices through their absolute changes (ARIMA model with 1 integration) or to model stock price returns and convert returns to a price time series? I have noticed that if we model price changes, when simulating the ARIMA models forward in time, they do not capture that absolute price changes are greater at higher stock prices and absolute price changes are smaller at lower stock prices. Also prices can go negative. It seems that modeling returns, and translating this into a time series, would correct for both of these.