I received this email today: I recall you made this very insightful remark somewhere that, fitting a standard arima model with too much data, ie. a very long time series, is a bad idea. Can you elaborate why? I can see the issue with noise, which compounds the ML estimation as the series gets too long. But is there anything else? I’m not sure where I made a comment about this, but it is true that ARIMA models don’t work well for very long time …