OK, sorry. Then the fileds are: n-1 = yesterdays adjusted close n-2 = the day befoe that etc ? And out is tomorrows adjusted close?
yes. well, actually out is today's adjusted close. However, it is also the training and prediction variable, so once applied to an out of sample/validation set, it becomes tomorrow's predictive close (with today's close being the n-1 variable in the new dataset).
dt, not an expert here, but without even using Jerry's suggestion of using technical indicators as attributes, maybe normalizing the #s by taking the % change would lead to a more predictive model? just a suggestion to see what that model would do vs the model you first built and how much this type of normalization contributes. jw
That is certainly another way to do it, and an often used method. However, my goal here was to try to get an RM project that everyone could actually work "hands on," rather than observe. So the priority should be to have everyone who wants to learn it, try to emulate and add their wisdom to the progressive RM model. This is one reason I didn't want to jump to having several people running several different idealized trading systems with a large number of input variables. If we are learning RM as a group, we should at least be able to get a simple simulation to work properly first. Once everyone agrees, we can start to progress to more fancy stuff.
I am in the process of teaching myself RM as well. I'll play around with your files so we are working with the same data. Won't be until after the weekend as I'll be out of town, but should be a good learning experience.
didn't see this until after I posted, but that sounds good. we'll use your excel data sheet and go from there.
Sounds good. One other point I'll throw out, since eventually someone will get to it is that there is no way to add offset bias for the hidden layer comparators (according to the developers of RM). However, I think if the input data is scaled properly and proper thresholding functions are used, it should be ok. This is one reason I was concerned about using differences as inputs rather than absolutes (as the input scale would then be balanced with positive and negative values about an approximate zero mean, it might cause the hidden layer comparators to saturate). Have to think about it more.
good news. I manually rescaled the inputs (which means I guess RM isn't rescaling automatically, if so, then why does it offer choice for input layers?). Also added a manual bias to the input, and we have rudimentary tracking for a start. Also, changed to hyperbolic tangents for comparators for now, to increase the range of the outputs. Still need to work on it more. But, the moral is rescale the input manually and add a bias for now.
I would like to propose a different twist on the idea of developing a Rapid Miner project that may be easier to develop and more beneficial right away for everyone. The idea would be to develop a Rapid Miner testing/evaluation method to identify trends and/or common themes in past trade results. Each person could then use this to immediately improve their performance while developing more complex financial models. Some potential ideas to evaluate: 1) Which market(s) provided the highest profit and/or largest losses 2) What time of day/day of week is the most/least profitable. 3) What commen technical indicators are present (i.e. with trend, against trend) at the time of the trade that result in both positive and negative trades I think this method may yield results quicker than developing a full ATS using data mining techniques Any thoughts on this?
Excellent idea. Prior to step 1 you'd need a way to define proit or loss. A trade strategy that enters at the Open will have a very different result than one that is stop dependant in the same market on the same day. In decoding my own NN models there is a major shift in result based on the postion of the open in realtion to the range of the last bar. Instead of Profit or Loss one may want to focus on MFE and MAE then add the trade strategy later. 2) Time of day varies by market. In the Russell Index it's entirely different than Forex. Jerry030