Quote:
Originally Posted by aroonraj
Hi Stever,
The initial idea that I shared in this thread was just for discussion and as a start to find better models. I agree that one need to consider as inputs other forces too for prediction. Many forex brokers do not have option to trade other comodities so its not possible to get the data for other commodities.
I currently use data from different currency pairs (inter-market) to predict the strength of EURUSD. Its still in research so I cannot mention any results, but I agree that predicting based on the history data of a single currency pair and its derived indicators is not very reliable. Your winning % of 60 is good as a start. if you can achieve the same with your model in forex, you can create reliable trading strategies.
Also prediction on daily basis will show good winning ratio in training and testing NN, but poor results practically on accounts with small amounts. I am not a full-time trader and am not an investor so I use smaller timeframes (M1-M15) for forecating signals. You can think my trading style as a NN scalper.
I am open for new ideas and discussion either in this thread or in a new thread on building trading systems based on NN.
|
Hello
I am studying NN too. Once I had reached same colclusion as you Arrone (i.e. predicting based on the history data of a single currency pair and its derived indicators is not very reliable.)
I still have not achieved reliable accuracy using NN especially using self data.
But including myself, some people can predict over 70% using only self data as input using some statistical methods in university text book (for example, None Neural Network method).
From this fact, my believe is that it should be possible to predict over 70% of accuracy using neural network too only using self data or using derived data from self data.
I think the problem comes from neural network itself.
and My view about NN is as followings:
Firstly, it has too much parameters while we only have limited number of inputs for training.
Secondly, it is too slow.
Thirdly, it often gets to wrong path to the solution while training (if you use many different training methods including genetic algorithm, steepest descent, conjugate gradient and so on you will see and feel this.)
Anyway, just written few comments as I think I am on where you are too in terms of NN wise.
good luck !!
