ARIMA(p,d,q) uses p lags of the series, d rounds of differencing to achieve stationarity, and q lags of past errors. On series with real structure, such as volatility, volume, or seasonal macro data, it can work well.
On return series it usually cannot. If returns were predictable by a linear function of their own recent history, that pattern would be arbitraged quickly, and fitted ARIMA coefficients on returns are typically tiny and unstable across samples.
Where it earns its place is as a benchmark. Before claiming a machine-learning model finds structure in a series, fit an ARIMA and see whether it captures the same thing more simply. If it does, the simpler model wins on robustness.
Related: garch, stationarity, autocorrelation, underfitting