I am writing my master's project proposal and really want to work on deep learning for time series forecasting. L.S.T.M has been suggested by most of the answers online. The data I will be working with is the sales data of the products on an E-commerce store.
However, I also saw some papers suggesting L.S.T.M do not really work well for real-life time series data. And it has the many problems including difficult tuning process, slow training extra. I could not find useful paper providing convincing benchmark either.
So, my question is, according to your experience, how well does L.S.T.M perform on time series forecasting tasks in comparison with traditional methods like A.R.I.M.A models and regression trees?
@jacob I certainly have seen LSTMs used extensively in time series forecasting. For example, in the recent M4 time series competition (which I understand is akin to ImageNet for forecasting), the winner from Uber technologies leveraged RNNs coupled with other novel approaches. The paper here covers the competition, results, and trends in the field: https://www.sciencedirect.com/science/article/pii/S0169207019301128
As another example, the paper here (https://arxiv.org/pdf/1704.04110.pdf) describes Amazon's DeepAR approach that utilizes an LSTM algorithm. It's worth noting that DataRobot incorporates this approach (along with many others) as part of its auto time-series (autoTS) capabilities.
Hope that helps!
@jacob it's a great question! As @duncanrenfrow mentioned, they have been leveraged very successfully for some time series problems. You may want to think about when deep learning is appropriate in general, not just specifically for time series. For example:
1. Deep learning only becomes strong when there is lots of data to learn from. This not only means when you have many observations, but also lots of features. They are fantastic for image problems because a single 224x224 image has 224*224*3 colour values (RGB) = 150,528 features!
2. Deep learning models take longer to train and generate predictions from, so they may not be suitable if you need real-time predictions from your model
what is about classification for multivariate time series, especially with mixture of categorical and continues values
can you share some such a dataset (train and test ) with performance of your code
then it will be possible to test your performance vs other packages
in any case seems to be you do not prove your capabilities by this way
why you are not open?
do you afraid to show how bad you abilities are?