I am not used to time series forecasting, so I feel sorry that my question might be stupid.

Now i'm dealing with real world time series data, which is very short. I want to know what method I should try that i can successfully predict this dataset.

First I'll explain about my dataset.

  1. High value(about 10^8~10^9) and high variance.
  2. Univariate
  3. Data length : 60
  4. Do not have seasonality and trend.

Below is seasonal decompose, acf and pcaf after I logged my dataset.

Seasonal Decompose



I tried some stat methods such as ARIMA, ETS, but the result was poor. I also tried on LSTM, but it had poor result. Also I think it's not an good method due to lack of dataset and it's univariate.

In my case, what method I should use?

  • $\begingroup$ You have each plot twice in your post, right? Do you want to delete the duplicates? Also, can you edit your post to include your actual data? Also, you note that you have high values of around $10^8$, but your series shows a $y$ axis only up to 22. Are these logged data? If so, be aware you need to be careful in back-transforming forecasts. Best to let your software deal with any transformations. $\endgroup$ Commented Oct 10, 2022 at 7:36
  • $\begingroup$ I'm sorry that I cannot post actual data cause it's company data. I logged my data using np.log1p and then made plots. I also transformed back after forecasting. Thank you for your answer $\endgroup$
    – Muroa
    Commented Oct 10, 2022 at 8:06
  • $\begingroup$ "I tried some stat methods such as ARIMA, ETS, but the result was poor. I also tried on LSTM, but it had poor result" Your question does not provide details why this is the case. What results? In what sense are they poor? "due to lack of dataset and it's univariate." What does this mean? $\endgroup$ Commented Oct 10, 2022 at 11:56
  • $\begingroup$ Sorry for my poor question. The poor result means that it seems failed to predict. In ARIMA and ETS, predicted values were same value(horizontal line in graph). In LSTM, predicted values were also almost same value. "due to lack of dataset and it's univariate' means that it's not appropriate to lstm. In my shallow knowledge, enough dataset and features are needed to make predict by LSTM $\endgroup$
    – Muroa
    Commented Oct 10, 2022 at 16:15

1 Answer 1


There is very little structure in your data. Your ACF/PACF plots do not indicate any kind of autoregressive behavior. In such a case, an extremely simple forecast, like the overall historical average, may be the best you can do. This may well be what your automatic ARIMA or ETS method yields.

Alternatively, you could try to search for drivers of historical dynamics, like promotions or similar activities. If you can forecast those, you can use them to improve your forecast of your focal time series. You may want to take a look at How to know that your machine learning problem is hopeless?

Incidentally, if by "I transformed back after forecasting" you mean that you exponentiated the point forecast that was calculated on logged data, I hope you know that this will not give an expectation forecast, but a median forecast. The difference is often minor, but in particular for high volatility series, it can be crucial. See the "Bias adjustments" section here. (I recommend the entire online textbook.) It's usually better to rely on a proven piece of software to determine whether Box-Cox transformations are indicated, and do the transformations and back-transformations. The gold standard are still the forecast and fable packages for R (both written by the author of the textbook I linked).

  • $\begingroup$ Thanks for your kind answer. It gives me lots of knowledge, which I didn't think about it. $\endgroup$
    – Muroa
    Commented Oct 10, 2022 at 16:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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