you could use ARIMA type models. note that there is typically strong correlations between lagged values of the variable

You can explore Recurrent Neural Networks and its variants such as Long Short Term Memory and Gated Recurrent Unit for time series data.

Hi Brandon Lee

Here are some important considerations while choosing an algorithm I would say.

So you may need to test few models to explore as how it works on your dataset...

Regards

Priynka

Vikas Ramachandra Priynka Sharma Ritika Lohiya Thank you for your kind answers.

Not sure if ARIMA type models could be used with categorical variables? I would have a dependent variable as a categorical variable.

For LSTM and GRU typed models, it might work. However, there might be correction issues that I might potentially encounter. Cause for LSTM and GRU models, it will look at the trends and predict accordingly. However, I want to work on a classification problem. Not sure if it is applicable.

Hi Brandon Lee

I would futher add that ARIMA models are univariate models (one time series only). You cannot fit categorical features as a result because how are you supposed to define categories across a single time series? ARIMA models can handle time related exogenous variables for instance..

Therefore, maybe if you want to incorporate categorical features, you need to consider fitting a single model to multiple time series that differ by some category.

Maybe give a thought again..

Best Wishes...

Priynka

You may choose regression models for your panel data. See e.g., https://www.nber.org/system/files/working_papers/t0008/t0008.pdf

It is not clear to me from your scenario whether you want to classify:

If the problem is #1, this seems fairly straightforward to do with (a) multiclass logistic regression or (b) a dense neural network with a final softmax layer, in either case with a feature for each time step.

If the problem is #2, some kind of recurrent network might be needed, but more information on the problem would likely be needed to choose a reasonable approach.

I would like to suggest random forest

The random forest consisting of many decition trees. It uses bagging and feature randomness when building each individual tree to create an uncorretated forest of trees whose prediction by committe is more accurate than that of any individual tree.

I would suggest to go ahead with ARIMA model

For this scenario, there are many many solutions. From Bayes classifiers, Hidden Markov models (HMM) to Deep Learning models. Since the rates of increase are actually categorized, it can be said that they are related to what HMM models identify as 'states' in statistical terms. From that perspective, the nature of the data and the fact that its time series variable, HMM model may be appropriate...