My use case is I have a dataset consisting of stock returns of S&P 500 Index stocks over a period and I also have the Index returns over the same period. The problem is a regression problem where I am using a rolling window technique using 52 weeks of returns as in-sample and 12 weeks as out-of-sample. I am using various machine learning techniques on the in-sample data to forecast the out-of-sample and compare the results. As a second experiment I'm using Bayesian hyperparameter optimization to try and obtain a more accurate forecast. In a third experiment I would like to use a data augmentation technique from the distribution of the in-sample data together with optimized machine learning models to see whether the forecast performance improves.
Hi,
Have you tried Synthetic Minority Oversampling yet? You could use this method as one way to oversample minority instances of your data (if that is your goal). We have, for instance, used SMOTE previously in an anomaly prediction taskArticle Deep learning with knowledge transfer for explainable anomal...
and found substantial improvement in learning performance by oversampling the original dataset and then training the learning model. For testing purposes, you can hold out a section of the original data to avoid bias. Their are excellent libraries for SMOTE in R/Python etc. An alternative could be ADASYN though my experience here has been very positive with SMOTE. Hope this helps.Best Wishes,
Joyjit
Whilst SMOTE is generally used for imbalanced classification problems, it seems that there is very little literature on using oversampling techniques for regression problems. My objective in the 3rd experiment is to use an oversampling (or more than 1) method to augment a small univariate set of data. The reason I'm restraining my dataset is because I want to use the same data and method as some benchmark papers I have chosen. If you could point me to some links and papers that would be highly appreciated.
Patrick Bezzina Hi, have you tried SMOGN! While I have never used this myself, I know there are papers out there that have successfully applied SMOGN to regression problems. Here is a paper describing the approach for imbalanced regression tasks http://proceedings.mlr.press/v74/branco17a/branco17a.pdf and the corresponding Python library https://pypi.org/project/smogn/. I have never used this library myself, but since this is a variant of the conventional SMOTE, it does appear promising to me. Hope this helps!
Best Wishes,
Joyjit
- Badges
- Science topic