Tha major reason is because it has sound mathematical assumptionsAccording to Box and highly interpretableJenkins, ARIMA is a linear non-stationary model. It ideas strong stochastic and probabilistic assumptions that majority ofis an imporvement over ARMA.
In my experience, ARIMA might be favored over other methods do not have those properties, stationarity orders, regressions on input and errorbecause of its flexibility.
Given this highly interpretable mod, it produces accurate but not necessarily the bestYou can achieve far better results if you decompose your signal into simpler components and use simple linear models to forecast each time series and then combine them into one forecast.
ItSo, it is highly practicalinitially easier to work with and unlike popular methods in Kaggle, it can be utilized in an adaptive manner, e.g updating coefficients when novel observations comehandle non-stationarity if there is homogeneous nonstationary behavior in the time series.
Also,There are many scenarios where ARIMA model fails. e.g. if there is used as a benchmark asheteroskedasticity in your data, it gives you a deep insight and understanding about the stochastic process under the studycannot be modeled with ARIMA.