Model_v2 is the code used to train a time series forecasting model. Actually, the training accuracy is quite good (i.e. 95%) and the testing accuracy is quite bad (i.e. 35). There's clearly the presence of overfitting. The purpose here is to improve the testing accuracy. I tried many things, but none of them worked well.
I tried to increase the dropout ratio, increase the learning rate, increase the number of epochs, changed the number of input features. Many people think it is going to work with the right recipe.
Suggestions :
1- Use weight decay (L2 regularization on the weights)
2- Try enforcing that all batches have the same number of -1, 0, 1 (i.e. classes)
3- "Setting a confidence threshold before accepting a prediction...so, unless buy or sell > 0.8, then do nothing"
I have been working on machine learning models for the past one year. I caan help you with his using various techniques.
Looking forward to have a chat with you
Hi.. I have completed a number of AI timeseries problem.. I can finetune your script with a couple of more parameters to increase the testing performance..
Hi!
I'm your statistician. Such a difference might be because the train set is not comprehensive and misses many values.
I will find the reason why the difference and do my best to improve your test accuracy.
Let me know when to start.
your overfit problem can be resolved in these ways: if the model is regularized, increase the dimensions of regularization parameters, for non-regularized models trying reducing the number of characteristics by their selection, or their extraction. Sometime happens that the model is fitting to much on the training data making it harder to generalize, and sometimes the problem can be resolved by adding more training data but maybe this is not the case. And otherwise trying optimizate the hyperparameters of the model.
If you consider my proposal I can show you the accuracies I get.
Good day,
I am interested in working with you on this project you posted. I have developed my own system in analyzing Forex tick data that predicts whether or not the next price will be higher or lower than the current closing price. At the moment as it stands, it has a training accuracy of 99% and a test accuracy of 70% based on back testing EUR/USD tick data from all the way to 2010. If such a system seems to be something that can work for you, send me a message and we can discuss it in more detail.
Kind regards,
Justin van Zyl
Hello. The problem you met is common but dont have the conventional solution. I met many of such situation and belive i can help you.
Note the estimated time may vary due to size of you dataset and your architecture.
I am mainly working with DL4j but also familiar with Keras.
I'm trained as a financial engineer and I've worked extensively on time series methods. I've also worked with Recurrent Neural Networks based with Keras package based on tensorflow. I believe I can help you.