0

I think it would be fair for me to explain a little bit on a background into what I am doing so that my question would make more sense.

I am currently working at a company where I need to develop a time series forecasting model. I'm given a dataset which contains features A, B, C, D, and E. And we would like to use these features to forecast feature E into some hours in the future.

I'm running into a problem where the data that I have is not enough at all to get any good performance for the model. I have discussed this with the lead engineer working on the project and I was informed that there is no way for more data collection to take place. So, I am a little stuck here with the little data that I have.

I know of two approaches which are similar to what I need:

  • In the computer vision people like to apply data augmentation on their images in order to produce more images and to reduce overfitting.
  • In classification problems people sometimes ally SMOTE or ADASYN to oversample the minority classes and ensure equal representation of classes.

I'm wondering, does there exist any approach in the field of time series forecasting which could help me generate more data by using some sort of data augmentation or smothing like SMOTE? My main goal here is to generate more time series data so that I can have enough to ensure that my model learns well.

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
Joseph Anderson
  • 31
  • 1
  • 1
    In a similar position, I was able to make use of symmetries (in time and space) and e.g. mirror some data. – Ggjj11 Oct 24 '21 at 20:50
  • Does this answer your question? [Data Augmentation strategies for Time Series Forecasting](https://stats.stackexchange.com/questions/320952/data-augmentation-strategies-for-time-series-forecasting) – Marjolein Fokkema Oct 24 '21 at 22:11

0 Answers0