Exponential moving average is easier to implement and understand if appropriate. ARIMA forecasting is more difficult to implement and understand if appropriate. ARIMA forecasting will be applicable in cases where exponential forecasting is not applicable. There is no universally best forecasting method. In any particular case, the answer depends on the time series to be forecast.
When testing different methods the ARIMA forecast is sometimes taken as a benchmark against which other forecasts are evaluated.
Look at Rob J Hyndman and George Athanasopoulos, Forecasting: Principles and Practice (3rd ed). A web version of this book is available at
Exponential smoothing is appropriate when the data has an easily understood pattern (e.g., seasonality in demand). ARIMA would be the better method if the deviations away from the mean are more irregular. Having said this, one shortcoming of the ARIMA method is that it presumes that the conditional variance is constant, which is why I have not used this method in over a decade. Instead, I favor using the ARCH/ARMAX method, which allows one to model the volatility in the data but also uses ARMA terms to model the autocorrelative nature of the data.
By the way, I respectfully do not concur with Professor Frain about ARIMA forecasting being difficult.
Here is a link to a chapter on exponential smoothing.
Chapter Forecasting with exponential smoothing. The state space approach
Kevin F. Forbes With modern software ARIMA or some form of automatic ARIMA is very easy to implement. My problem is that I have seen many economists making an awful mess of an ARIMA forecast.