Machine learning in trading: theory, models, practice and algo-trading - page 48
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
I also experimented a bit with RNeat, and came to the conclusion that it cannot be handled the same way as conventional neural networks.
Thank you for sharing your findings, very interesting...
I took your example with irises and reworked it for my data and got an error, what's wrong I don't understand. Please tell me exactly tried to substitute your data besides iris, what were the results compared to conventional network?
I experimented with RNeat a little more, and came to the conclusion that you can't work with it the same way as with conventional neural networks.
1) unlike conventional models, RNeat training does not use raw data. The model is sort of randomly generated, improved, and only at the end is tested on the raw data. Usually models use raw data, build their logic on it, and then use validation sampling to see if the model's logic is correct, or if it has just memorized the raw examples. Unlike the others, RNeat isn't capable of memorizing the raw data at all, for it doesn't know about it, all the model knows is the desired results and how close it is to them.
2) crossvalidation will not help improve fronttest results. We all seem to agree that it is possible to train a regular model, run a couple of crossvalidations, and somehow handle the errors on all samples for the final fitness score of the model. Since RNeat does not know the raw data, it does not matter for it what sample the data is in - training or test, it will adjust its logic to the desired result in any case. Theoretically it will "learn" (more precisely mutate :) ) on all samples that are used in the fitness function. All you can do is train the model to the desired accuracy, and hope that it is not over-trained, quite a risky approach for forex. Crossvalidation can be used only after training as a final estimate of possible result in fronttest, and in no case this validation sample should be used inside fitness function.
Please tell me exactly have you tried to substitute your own data besides irises, what were the results compared to the usual network?
I have added data from forex (eurusd d1 for a year), the model is learning for 24 hours already and shows its own best estimate = 0.017 (the more the better, for iris it was 0.7). So far I have not got any results, the model shows error 50% even for training data and I have to let it learn for a couple more days.
On crossvalidation the training parameters are picked up. Does this NS have some parameters? There have to be.
Model is getting more and more complicated and is adding more and more neurons. There is a parameter defining the maximal number of neurons, to limit their growth. And a second parameter defining the number of individuals in genetics. Usually they both the more the better, the limit is introduced rather to save memory and time, it's better to use the maximum value allowed for yourself, rather than pick up.
The model, judging by descriptions, has a price for each neuron and connections between them, so it itself seeks to use as few neurons as possible. If the training is successful, the model will reach the required accuracy before it reaches the maximum number of neurons.
I have added data from forex (eurusd d1 for a year), the model is learning for a day already and shows its own best estimate = 0.017 (the more the better, for iris it was 0.7). The model shows error 50% even on the training data and I have to let it learn for a few more days.
Interesting article on evaluating classification models.
I tried 3 different metrics to estimate model on subset of predictors: Accuracy as in this article; R^2 model; and just average model error mean(abs(y-x)/(max(y)-min(y)), where X is vector of prediction results, and Y is required results).
Accuracy for fitness functions was quite suitable, but I had a problem with not many examples in the sample, the accuracy comes out very staggered. Two models, close in quality, just gave the same result. But if there are a lot of training and validation examples, it will be a good metric. I can't say anything about the rest of the metrics in the article.
R^2 behaved strangely, it could even be negative, even though more than 50% of the results were correct, in general I didn't like it.
The average error suited me best. Stepwise as in accuracy is not present, because I substitute in the formula the result of regression, even before rounding to classes 0 or 1. Even if the two models give the same result for the classes, I can still determine the best one, in terms of how much closer the regression result is "closer" to the desired results.
Are there a lot of predictors?
About 400
Are there any guys here who understands Fourier spectral analysis at least more or less. Need some help.
And what is not clear? We decompose a periodic function into harmonics and get a spectrum.
That's true, but my problem is different...
I need to find similarities in BP those areas that are close to each other, proximity can be measured by correlation, Euclidean metrics, etc... According to my experiments, I understood that such metrics are not suitable for my task, I think that very well suit the search for proximity through the amplitude, phase, frequency... But how to do this I have no idea, so I ask who knows how and what, what is right or wrong, how it should be in general.
P.S. you can in private, so as not to off-point, in principle, the conversation is not on the subject of this branch and is unlikely to anyone interesting
What is unclear there? We decompose a periodic function into harmonics and get a spectrum.
We can also decompose a single pulse, not periodic at all)