2

I have a timeseries with a sample every 5 minutes. I want to forecast the timeseries multiple step ahead (e.g., 60 minutes, which is 12 samples ahead) using its past values. Unfortunately my model does not seem to capture well the dynamics of the timeseries. Maybe it comes from my lack of data.

My question is then the following: can upsampling (e.g., going from one sample every 5 minutes to one sample every minute, and then filling the gaps through interpolation) can be an effective means to augment my dataset and thus help my model capturing the dynamics of the timeseries?

What do you think? I did not find any paper covering the idea. I have seen the following cross-validated post covering data augmentation in timeserie: Data Augmentation strategies for Time Series Forecasting. While quite interesting, it is not dealing with the simple idea I am asking about.

Ferdi
  • 4,882
  • 7
  • 42
  • 62
Xema
  • 195
  • 7

1 Answers1

3

"My question is then the following: can upsampling (e.g., going from one sample every 5 minutes to one sample every minute, and then filling the gaps through interpolation) can be an effective means to augment my dataset and thus help my model capturing the dynamics of the timeseries?"

Definitely not as you are injecting structure into the your new data stream.

Predicting 12 periods out OR the sum of the next 12 periods out is not a problem at all.

Your problem could be data or it could be software or limited experience. If you wish to post your data or a facsimile please do so and I may be able to comment further.

IrishStat
  • 27,906
  • 5
  • 29
  • 55
  • I was doing some experiments and had largely better results when upsampling to 1min, thus my question. However, it appeared that when resampling my data to 5 min, I did a big mistake during the aggregation of the timeseries (when time conflict arises). When there was a time conflict, I summed the value instead of taking the mean of them. Now that is corrected, I have equal results whichever the sampling rate. Thank you a lot for your answer :). – Xema Mar 28 '19 at 15:06