I want to obtain a prediction model using support vector regression on a time series data set. In literature, I have read that the break up ratio for training/testing/validation of data should be around 60/20/20 percent. But the length of my data set is very large, so if I use 60% of it for training, it takes a lot of time.

Instead of taking the ratio as said above, is there a minimum length of training sequence that if used is less time consuming and still ensures statistical reliability of the proposed model?


You could fit the model with only 50%, 40%, ... (and so on) of the data (the training dataset), as long as the model will prove to be reliable during the validation and the testing phases.

In other words, you need to make sure that will be enough data to assess the model performance in terms of overfitting and accuracy.

Found the below related post; Ben Allison's answer provide some advice in cases where the method is computationally intensive.
Is there a rule-of-thumb for how to divide a dataset into training and validation sets?


Not the answer you're looking for? Browse other questions tagged or ask your own question.