Forex Software — Hope joined the Support Vector Machine method in the Indicator.

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-07-07T12:41:02Z

After quite a number of tests I can conclude on my side that SVM cannot predict log returns, next bar direction, next fractal or whether next bar is higher or lower from the previous with reasonable accuracy, it tops out at 50% or fails completely. The reported accuracy of 90% seems like a myth, but I have read some organizations do achieve these kind of rates. But they have serious resources, which I don't. I'll leave this as it is.

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-20T22:45:37Z

hehe I am not too sophisticated.

I have a universe of stocks in a charting program and measure the ups and downs with a couple sorts........ the result of which tells what is going to happen tomorrow.

I use about 7,000 US listed stocks in an old outdated charting program and I have developed numerous sorts, a few of which I have posted in here. This is time consumiong as I have to download and then import the data from Yahoo after 8 pm and do the sorts which takes quite a while.

One day I may get the Worden software and do some neat things, that is very powerful for sorting and charting.

Posting a prediction is suicide so I have never done that.

Some ETF traders have automated systems using RSI 2 that are very accurate on individual instruments.

I guess making one for Forex instruments is quite possible using daily or weekly, however time to develop it is not there.

I am not really into forecasting, I thought it might be great for a while but now cannot see any money in it for me.

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-20T19:08:57Z

Blaiserboy wrote:

I made, once, something that will predict the movement of the major stock indicaes, however that was using a universe of stocks.
It had a great accuracy!
I would do it today but the operation is time consuming.

Speak up. How did you do it? What was the underlying method? Correlation?

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-20T18:00:35Z

I made, once, something that will predict the movement of the major stock indicaes, however that was using a universe of stocks.

It had a great accuracy!

I would do it today but the operation is time consuming.

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-20T01:10:13Z

Thanks for the encouragement, Dave! I returned to precompiled version, beats RM speedwise hands down.

Return prediction turned out be useless, now I tried to train it to predict next bar's direction -- 49% correct. Will play with data later, but SVM really seems more of a hype.

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-13T15:27:08Z

I spent a lot of time with that program and a couple similar..... and made no money.... That is why I abandoned it.

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-13T15:19:49Z

I went through the results again. Apparently that 86% accuracy is rubbish and that 81% is not better either. The actual prediction rate for predicting bars with movement is very low, so low it's useless. Certainly there are parameters for tweaking, secondly a rethink how to build the training data is on the cards.

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-13T12:33:49Z

New update:

I went back to rapidminer, using the same dataset to verify my previous results, here it is.

Not too bad but should be closer. The biggest problem is that I'm not sure whether I can trust these numbers.

I'll do next some tests with directionality in mind, if numbers are above 80%, then it truly is promising. I think VC must have some experience with this, it's a shame he has gone silent.

I must add this is classification (C-SVC).

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-13T12:20:22Z

Ok Dave, it's somewhat working for me now.

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-12T18:10:45Z

I can not offer any assistance, I deleted the software many months ago and have forgotten whatever little I knew.

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-12T15:44:18Z

Additional info: dataset relation is 50%, i.e equal sizes of 10000 bars for training and testing. Timeframe H1. I did use returns (close - open), close values alone are statistically a bad choice, I've been told. Basically I tried to predict bars which have at least 15 pips difference between open and close, and the accuracy of it is 86%. I used RBF kernel, all settings default (because I'm at wits on how to change the parameters...).

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-12T15:33:15Z

Get Yer Ya-Ya's Out! I exercised my brain a bit and I think I sorted out one error I made while formatting the data. Nothing official, but the testing is working much better, see the result below.

I would have one, who else?

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-11T19:29:16Z

I wish there was a clear-cut manual for either of them, learning is the easy part it seems.

Dave, my RM guru, what should I do then to get at least a hint of result?

Logic is: training data->train SVM->SVM train model->testing data->results.

How do I connect those?

P.S. postimage.org doesn't seem to work properly, screenshot is attached.

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-11T15:44:30Z

I played with Rapid Miner for a couple years, actually just deleted it about a year ago. You can do some amazing stuff with it, especially scanning the net for things. (if you have a gazillion terrabytes of hard drive. LOL)

I did not code any strategies that I found but time spoent on it gave me some understanding of things to do with data mining.

From my experience with it, I decided to stick with what I am doing as the learning curve in statistical methods is too steep for me.

Re: Hope joined the Support Vector Machine method in the Indicator.

2015-06-11T15:01:22Z

I tried rapidminer, it's a nightmare experience, ditched it and went for the precompiled windows version run from cmd. Data format is a bit of a hassle but I think it was a success as it worked. Interesting part is that I get 100% accuracy, which can't be correct...

Pic of the HG below.