Should forecasting with neural networks only be treated as a supervised learning (regression) problem?

Question

I have recently made a work about the application of neural networks to time series forecasting, and I treated this as a supervised learning (regression) problem. I have come across the suggestion of treating this problem as an unsupervised, semi-supervised, or reinforcement learning problem. The ones that made this suggestion didn't know how to explain this approach and I haven't found any paper of this. So I found myself now trying to figure it out without any success. To my understanding:

Unsupervised learning problems (clustering and segmentation reduction) and semi-supervised learning problems (semi-supervised clustering and semi-supervised classification) can be used to decompose the time series but not forecast it.

Reinforcement learning problems (model-based and non-model-based on/off-policy) is to decision taken problems, not to forecast.

It is possible to treat forecasting time series with neural networks as an unsupervised, semi-supervised, or reinforcement learning problem? How it is done?

Answer 1

I think the choice of technique strongly depends on how fine-grained your forecast-predictions need to be.

When it comes to forecasting by Reinforcement Learning (RL), one prominent example is the stock-trading RL agent. The agent must decide which stock to buy or sell, thereby drawing upon predictions concerting the expected future development of some stock. Given this approach, you would not necessarily let the RL agent explicitly generate estimates of how stock prices are gonna develop at any point, but instead you would only observe the predicted decision concerning whether to buy or sell etc.

But if you think hard enough, I am certain that you could come up with setups of RL agents that would allow you to explicitly generate future estimates of values to be forecast. In this case, the final buy/sell decision would have to depend on the explicit future stock price predictions to enforce accurate predictions.

Concerning unsupervised learning, you could cluster data points (training samples) with respect to how some value(s) of interest changed $t$ time steps in the future (after having observed the training sample). Then, you could associate clusters with rough forecast-estimates. After all, you would treat the forecast value as a label associated with data points. Afterwards, you could use some kind of nearest neighbor approach to determine which cluster is closest to some novel data sample. Then, you take as a prediction for the new data sample the forecast prediction (i.e. label) that is associated with the closest cluster/prototype etc. But strictly speaking, as soon as you start turning forecast values (which were previously part of some unlabeled time-series dataset) into labels, you turn the training procedure of course into a supervised technique again.

How well especially the latter training approach would work, I can't tell since I have never heard anyone using this method. But if training data is too scarce to employ some deep learning method, why not giving it at least a try if accuracy doesn't have to be too precise?

After all, it's just a matter of creativity and testing which method works best given your specific machine learning problem at hand.