How to improve the result of "Time Series Forecasting Using Deep Learning" ?

I am working on "Time Series Forecasting Using Deep Learning." (https://www.mathworks.com/help/deeplearning/examples/time-series-forecasting-using-deep-learning.html?searchHighlight=predictAndUpdateState&s_tid=doc_srchtitle)

The result of the prediction is not satisfactory compared to what I expected.

How can I improve the result of prediction?

For instance, what options can I change?

It may improve if I use more data, but it is limited.

I changed epoc number, initial learning step size, training data number, etc; nonetheless, the result is not satisfactory

Please let me know if there are any ways to improve the result for prediction. Thanks

Answer 1

[Kritika Bansal]

Hi,

You can try tuning the parameters like ‘MiniBatchSize’, ‘MaxEpochs’ and ‘Solver’ to train the network well. Also try to tune the parameters within a particular ‘Solver’ like tuning the value of ‘Momentum’ for ‘sgdm’. Refer to the link below to explore more such options:

https://in.mathworks.com/help/deeplearning/ref/trainingoptions.html

Answer 2

[Jaechan Lim]

I changed solverName from "Adam" to "rmsprop" and somehow it worked better.

I also needed to adjust the values of "InitialLearnRate".

The tuning process is not easy, but thanks, anyway.