1
$\begingroup$

I am doing some time series forecasting task up to 4-week ahead (in weekly scale).

However, I think that even if I fit a ARIMA model (with some optimisation on selecting parameters) to the time series, the RMSE is worse than a simple baseline model

In the baseline model, I just use the last observation t, as a predictor of t+1 to t+4.

I wonder what is wrong on my ARIMA model? is there any thing I missed to do for the time series input before forecasting?

Edit: some details are added:

  1. the frequency is weekly. And there are about 15 year of data
  2. I am talking about the within-sample prediction, not out-of-sample prediction
$\endgroup$

1 Answer 1

1
$\begingroup$

You didn't give us many details about your data, but if you don't have many observations then the simple methods like predicting the previous value work remarkably well. What is the frequency of your data? If you have weekly data, predicting four weeks ahead is not a big horizon, but for daily data, it may be long. Maybe there are no patterns in your data? Time-series models detect patterns like cycles (seasonality) or trends. If your data is white noise, those algorithms won't work (and may overfit), so predicting the global mean or last value can work better. So yes, it can be the case that the simple benchmark is hard to beat.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.