State Space Models

Has anyone here read about/implemented state-space models successfully, a-la Durbin Koopman? I am wondering if these time-series methods have any useful predictive power – I know ARIMA etc. have been shown in most empirical avenues (at least for market data) to have zero forecasting power and do terribly out of sample.

Is there any merit in state-space time-series models? Or is it time to throw traditional time-series analysis out the window and all-in on neural nets trained with petabytes of alternative data?