3
$\begingroup$

I have come across some papers discussing the use of genetic algorithms as a forecasting tool. I don't however understand how a genetic algorithm, which to my (limited) knowledge is used to solve optimization problems, can be used for forecasting.

I do realize this is a very general question, but a simplified explanation and suggestions for how to get started with such forecasts would be very helpful and I'm sure of interest to others as well.


Bonde, G., & Khaled, R. (2010). Stock price prediction using genetic algorithms and evolution strategies. Institute of Artificial Intelligence, University of Georgia. Athens.

$\endgroup$
  • $\begingroup$ why is it necessary to use feature discretization is this paper ? can we just use genetic algorithms directly on the training data (training the model at each step and counting the number of correct responses as the fitness function) ... or is it just because in a general way: "discrete features improve classifiers results", and thus it is independent to the use of a genetic algorithm here ? (knowing that the genetic algorithms are not used for feature discretization in this paper) $\endgroup$ – guzu92 Jan 19 '17 at 17:18
1
$\begingroup$

By reading a little bit the paper you attached, GA is used for feature selection. It seems to me that they are trying to optimize the weights of a logistic regression model in order to predict an increase or decrease of the stock price that is represented as a binary variable. The weights represent the importance of each feature they have chosen empirically to add to their model.They use training and testing datasets from previous years. So they know if their model is predicting well or not. Once they extract the optimal weights using the GA they can apply their model in the present to predict future increases or decreases of a stock price.

$\endgroup$
  • $\begingroup$ why is it necessary to use feature discretization is this paper ? can we just use genetic algorithms directly on the training data (training the model at each step and counting the number of correct responses as the fitness function) ... or is it just because in a general way: "discrete features improve classifiers results", and thus it is independent to the use of a genetic algorithm here ? (knowing that the genetic algorithms are not used for feature discretization in this paper) $\endgroup$ – guzu92 Jan 19 '17 at 17:16
  • $\begingroup$ I am not really sure what they are doing and if they actually discretized the data or not. Unless they optimized the thresholds too with the GA and they haven't mentioned it. The paper isn't really well written. (Btw, I erroneously wrote that the weights represent the importance of each feature. This is the case in simple regression if you normalize your data. In logistic regression actually you can't easily intepret them). I guess you are right and they might have just done it to improve the results or it maybe be related with the type of their data? Not really sure $\endgroup$ – amanita kiki Jan 28 '17 at 2:34

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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