Should I use preprocessing techniques (such as cross validation or clustering data mining) to prepare my stock price time series data before I apply the GA. What is the minimum number of observations that should be used to test the GA. Is it possible to use the general method (90% training and 10% testing)?
What do you mean by "crossover validation" ? If you mean making k-folds of your data for training and testing this is called "cross validation". While "crossover" is an essential operator in the GA and GP algorithms which swaps parts of some selected individuals. These two terms are totally different.
Yes. That was by mistake. Thank you Hossam. So Is it helpful to prepare and split the time series data with "cross validation" technique before applying the GA.
Dear Abdelrazzaq;
if your fitness (i.e. objective) function is well formulated, then there is no need for pre-processing stage;
however, i want to pay your attention to the work of Prof. Tsang (like the attached one) you may find Genetic Programming is much more powerful when dealing with time series;
regards;
Article Forecasting stock prices using Genetic Programming and Chanc...
You need to apply cross-validation. 90% training and 10% testing is acceptable if the dataset is very large, otherwise a leave-one-out is preferable. In any case you should run your GA at least 30 times to cope with its stochastic nature and provide st. deviation, average, median, min and max results obtained in the 30 runs.
Dear Abdelrazzaq,
If you decide to adopt GP instead of GA, I think you should take a look at this paper, in addition to the paper suggested by Amer Bakhach:
Wagner, Neal, et al. "Time series forecasting for dynamic environments: the DyFor genetic program model." IEEE transactions on evolutionary computation 11.4 (2007): 433-452.
Best regards.
- Badges
- Science topic