机器学习算法预测下一个价值

Question

I am trying to predict the next value of a series. What the best machine learning / algorithm I need to use?

I have for example this matrix:

 [114, 160, 60, 27]
[74, 97, 73, 14]
[119, 157, 112, 23]

and I want to predict this values:

 [114, 160, 60, 27 , **80 , 90**]
[74, 97, 73, 14 , **10 , 15**]
[119, 157, 112, 23 , **50 , 48**]

What is the best way to do it?

Answer 1

If I understand well your question, in your case :

X = [114, 160, 60, 27] and Y = [80,90] [74, 97, 73, 14] [10,15] [119, 157, 112, 23] [50,48]
And you want to fit a machine learning algorithm on this data ?

You could use any supervied learning algo like regression or SVM using X as input and Y as output.

You could also use iterative learning : you learn a predictor f a step ahead with :

X = [114, 160, 60, 27] and Y = [80] [74, 97, 73, 14] [10] [119, 157, 112, 23] [50]

You do the prediction one step ahead :

f(X) = [pred1] [pred2] [pred3]

After that you incorporate the prediction in the input, so now you have :

Xbis = [114, 160, 60, 27, pred1] and Yter = [90] [74, 97, 73, 14,pred2] [15] [119, 157, 112, 23,pred3] [48]

And you train another predictor fbis on Xbis and Ybis.

So at the end, you have two predictors f and fbis , both of them predicting one step ahead. It enables you to do prediction two step ahead. Of course, you will need more data to train a good predictor.

More generally, if you want to do time series prediction, you can use "the window method" which is a general method to create your input and output from the time series to then learn a predictor.

Note also that LSTM are quite use in time series prediction and seems to give pretty good results.

Hope this helps !!

Benoit