1

I wish to predict whether the difference in Value and Growth returns is positive or negative for the next month.

To do this, I have collected data of a few features(to be specific, Macroeconomic Indicators), and the difference in Value and Growth returns for about 200 previous months. I wish the prediction to depend on the order of the previous months. Decision trees and others don't care about the order so I looked upon the net to find about LSTM. Is it correct to use them here?

Valentin Calomme
  • 6,026
  • 3
  • 21
  • 52
Vinayaka Srivardhan
  • 11
  • 1

3 Answers3

1

About what I have understood, you need to use time series forecasting algorithms. As your data is time series, you can look at these algorithms based on Deep Learning. You can also use classical algorithms for time series forecasting. You can use LSTM also which is good, details are given in link. Look here

shahid hamdam
  • 151
  • 2
  • Most of the time series forecasting algorithms (which I have seen) seem to predict on the basis of previous values of the same series, which is not the case here. I wish to predict a series on basis of features, more like a Titanic live or die? problem. – Vinayaka Srivardhan Jun 05 '20 at 14:16
0

I think the onus of model selection is on the features/predictors you have in the dataset rather than your decision to model the inputs as fixed-length or of variable length (if you are working with speech data, then variable-length sequences make a whole lot of sense).

I am working on a similar problem wherein my predictor space includes a single continuous variable measured every hour, along with 4 categorical variables, i.e. hour of the day, day of the week, month, and year. In an LSTM model, creating fixed-length sequences of my continuous variable makes sense, but I don't find it convincing to create the fixed length of days, months, or years to feed in the LSTM cell.

pandi20
  • 1
  • 1
0

Your problem is a classification problem. Given some input, you want to provide a binary output (positive or negative). Which means that you can use any classifier to tackle it.

Decision trees and others don't care about the order

Not necessarily. Let's say that you decide to use the last 12 months the predict the next month. Each month could be a feature. And in this case, the classifier will "care" about the order as the order is implicitly represented by the different features.

The main advantage of using an LSTM over something like a decision-tree is that an LSTM's input can be sequence of arbitrary length whereas a decision tree requires a fixed-length input.

Valentin Calomme
  • 6,026
  • 3
  • 21
  • 52
  • Thanks for the reply. So if I introduce a new feature which is basically the serial number(i.e. month number), and try to classify using Decision trees, then there will be an order imposed due to the new feature? (Actually I am new to ML, and I find Decision trees easier to visualize) – Vinayaka Srivardhan Jun 05 '20 at 14:14
  • Let's take a simplistic example. If you have 2 features, your model will try to learn relationships between them. For instance, when feature0 is smaller than feature1, the class is positive, otherwise, negative. Now, if your features happen to be ordered or not, it doesn't matter. Your mode will look for relationships. Think of features as "columns" if that helps – Valentin Calomme Jun 05 '20 at 14:50