# Considerations for forecasting

I am trying to make a generic forecaster for many short (~14 points) to long term (~365 points) time series data assuming the seasonal period to be weekly. The predictions are going to be made for a streaming input so I will have to automate most parts. The way I am approaching now is to loop through a bunch of time series models and pick the best one based on AIC.

1. Probably, that is not the best idea but I am starting out with this. Could you please suggest what other considerations/approaches should I take?

2. Some of the data is just too random without much seasonality or trends. Should I skip forecasting them? If so, how should I decide?

3. What would be the best way to evaluate the predictions considering I have pretty large number(~100000) of datasets I am working with.

• And then I find this. So I am a bit lost now. – Sidhha Apr 15 '15 at 9:52
• I'd suggest comparing relative probabilities exp((AICmin − AICi)/2) instead of plain AICs -- see AIC. Could you post some of these rel probs for your data ? – denis Jun 4 '15 at 16:42