Hitting a wall with a seemingly unforecastable time series: how to deal with it?

Question

I have a time series for which I can't seem to get any good results. I've tried multiple ARIMA models and exponential smoothing models and none of them give good results.

I have 104 weeks of weekly data, and I'm trying to forecast over the next 7~8 weeks. The best I can get is a MAPE of 35%. I want to get it to 20% or better. (I've not tried Croston's because sparsity/intemittency isn't an issue).

I' wondering whether there is any benefit to trying more advanced models (and the only more advanced models I can think of are NNets) or not.

If ARIMA doesn't give good results, would a more complex algorithm give any better results? Or is the series just not forecastable ?

What are the approaches for dealing with such a time series?

Is there a way of knowing whether 35% is the best accuracy possible with the series?

(I realize that the question is broad. Although I am trying to solve an immediate problem, I am also look for more general answers, as text books don't seem to address this problem).

Too vague to be answered. We don't know the domain, the forecast horizon, your expectations for how good a forecast should be, whether the issue is outliers or .... lots of things. — zbicyclist, Jan 01 '18 at 23:58
@zbicyclist I've added some detail. But I'm looking for broad answers. — Skander H., Jan 02 '18 at 00:20
Why do you think there could be a model that will get you MAPE of 20%? — Michael R. Chernick, Jan 02 '18 at 02:48
Closely related: https://stats.stackexchange.com/questions/222179 — Juho Kokkala, Jan 03 '18 at 07:31

IrishStat · Answer 1 · 2018-01-02T00:40:59.153

Often time (read nearly always) untreated level shifts and one-time anomalies create havoc with simple arima model identification. As @AdamO Interrupted Time Series Analysis - ARIMAX for High Frequency Biological Data? elegantly wrote the correlogram should be calculated from residuals using a model that controls for intervention administration, otherwise the intervention effects are taken to be Gaussian noise, underestimating the actual autoregressive effect. If you wish I will try and help further by inspecting the actual data. Please post it a csv column format.

Hitting a wall with a seemingly unforecastable time series: how to deal with it?

1 Answers1