0

I have a AUTO ARIMA fitted on weekly. It gives a forecast as flat line as seen below

enter image description here

I tried adding seasonality and stuff, but still the forecast is a flat line. Clearly there are some fluctuations, but why isnt my output showing any?

Shravan K
  • 23
  • We have a number of similar questions which may be helpful. Note that a flat forecast may indeed be the best possible. Also, your time series looks like count data with intermittency - in terms of the data generating process, that can't be ARIMA, and although ARIMA may be a possible model, it may not be the best one. – Stephan Kolassa Jun 23 '21 at 06:30
  • i saw familiar questions, but it was not for data which had no trend. can you tell me if ARIMA be useful for data without trend and very less chance of seasonality? if not which model should be considered for such data? – Shravan K Jun 23 '21 at 06:34
  • I personally am rather skeptical about ARIMA (it seems to me it is only taught because it uses impressive mathematics, not because it's good for forecasting). Theoretically, yes, you can use it if there is little trend and seasonality. Count data forecasting is an under-researched (and -taught) field. Do you have predictors that explain your peaks? If so, something like Poisson regression or Negative Binomial regression would be possible (perhaps using harmonic predictors to model any seasonality). Otherwise, you could look into the intermittent time series forecasting literature. – Stephan Kolassa Jun 23 '21 at 06:40
  • Take a look at the tsintermittent package for R. Even if you don't use R, the references there might be useful. – Stephan Kolassa Jun 23 '21 at 06:41
  • Thank you. As per my knowledge there wont be trend because it is a consumption count every week. so it will be grounded. I was strictly told to start with ARIMA then move on to other models in future. I guess better not to expect good result out of ARIMA – Shravan K Jun 23 '21 at 06:46
  • No. i dont have any predictors which explain this peak. It is univariate data – Shravan K Jun 23 '21 at 06:47
  • ARIMA is a good first benchmark. You could also look at Exponential Smoothing, which is again not really appropriate for count data, but it's simple and fast to fit. Also other simple methods, which are often surprisingly hard to beat and should always be considered as benchmarks. Also, be careful about your error measures, especially if you have many low volume count series - optimizing the MAE will give you biased forecasts (Kolassa, 2020). – Stephan Kolassa Jun 23 '21 at 06:49
  • Thank you for your time Stephan. I will look at other simple methods. But can u say in one sentence why ARIMA didnt work on my data? is it because of lack of trend? or since my data looks random without any pattern – Shravan K Jun 23 '21 at 06:52
  • I wouldn't say that ARIMA "did not work", the forecast may be fine. But your series may indeed have no dynamics over time (so the best forecast would be the historical average), but ARIMA mistakenly thinks it does. Any automatic method can pick a wrong model. – Stephan Kolassa Jun 23 '21 at 07:01
  • what do you mean by dynamics over time? – Shravan K Jun 23 '21 at 07:08
  • Autoregressive (AR) or moving average (MA) behavior, which ARIMA models. Even if your history is generated by a specific ARIMA(p,d,q) process, if you apply Auto-ARIMA to it, you will very often get a different model than the one that actually generated the data. (Experimenting here is very enlightening.) And if there are no dynamics in the data generating process, that is an ARIMA(0,0,0) process, or simply IID white noise, so if Auto-ARIMA fits a nonzero p, d or q, it's wrong. (It may still yield usable forecasts.) – Stephan Kolassa Jun 23 '21 at 07:11
  • Thank you for your answers Stephan. I have better idea about my model now :) – Shravan K Jun 23 '21 at 07:12

0 Answers0