-1

I am trying to execute neural network using caret for a classification problem. My dataset has 94000 rows and 5 variables (1 binary dependent variable and 4 independent variables). I have cleaned the data and transformed the necessary variables. No missing data present. Data is balanced. I am using nnet using caret for cross validation. Now I am running below code:

model <- train(Output ~ Var1 + Var2 + Var3 + Var4, 
                          method = "nnet", 
                          data = my_data, verbose = FALSE, 
                          trControl=trainControl(method='repeatedcv', 
                                                 number = 10, 
                                                 repeats = 5,
                                                 verboseIter=FALSE), 
                          tuneGrid=expand.grid(.size=c(0,1,2,5,10,15), 
                                               .decay=c(0,0.001,0.01,0.1)))

Can you please tell me how can I improve my model further?

Additional details: 1. The data is financial market data. 2. I will be checking for best threshold value using Youden's index after prediction. Thus, I will select best threshold value only for reporting.

As I am novice to Neural Network, I am seeking suggestions for improving model using R and what kind of tweak or other model I can use.

Zico
  • 185
  • 12

1 Answers1

0

Apparently, your question is quite broad. Here are some ideas and observations that come to my mind when I read the code and the additional details that you have posted.

  1. I think it is pointless to evaluate zero values for the neural network size and weight decay.
  2. Given the above, you can improve your model by augmenting the grid size, mostly for neural network size.

Commenting that you are using financial market data does not add additional information for the reader. Are you using time series? What kind of variables are you considering? Are you addressing the problem as a regression or classification one?

bastian.abaleiv
  • 53
  • 6