Thursday Nov 30, 2006
Prediction 4 Months Ago and Actual Trends Today, by Neural Network
Today, Shanghai Security Index (000001.SS) touched 2100, and, from my previous neural network research on 0000001.SS, about 4 months passed. In that blog, I placed a prediction picture, and now, here it's a verification picture of the actual price trends comparing to the prediction:
The Blue line is the prediction calculated by Neural Network.
The Red/Green line is the actual price trends.
Click on the picture to enlarge it
Posted at 12:55PM Nov 30, 2006 by dcaoyuan in Trading | Comments[7]
Your prediction was for a general trend upwards and this is based on the fact that the previous period was also a general trend upwards. The prediction and actual do not match closely at all.
You'd ahve to be insane to invest on this basis
Posted by technical investor on November 30, 2006 at 05:48 PM PST #
I'm not a day trader, so I don't care if it's exactly close to the actual. I care if the trend will go on, when will it end or turn over. I posted the period about 4 months, and won't post more, I know the neural network used to be disputed. I also don't think anybody other than me will invest based on this picture, as they don't know how this was computed out, and, I'm not a licensed adviser (and won't be one).
But, I do invested on this basis. Till today, I've earned about 22%. And the prediction was based on the last 10 years data, not the previous years
trends.
I'll keep my own research on this. For most traders, AIOTrade will be a usual technical analysis platform.
Posted by Caoyuan on November 30, 2006 at 06:43 PM PST #
Hi,
I was wondering what kind of NN you were using, and how did you train it ? I've heard about people training NN in order to get signal rather than prediction. What do you think ?
Posted by Natural Intelligence on November 30, 2006 at 08:12 PM PST #
It's a regular MLP, I just carefully select the input set.
For me, I choose another machine(SVM) to get signal. SVM can do prediction too. It's interesting that the NN do produced the prediction like this one. Usually NN can only produce short term results, such as several days/weeks/months according to data's frequency. That why I have a fairly confidence on my result, because I know how it was computed.
From the Statistic option, the result is also reasonable, I think the reasonable result should be:
1. produce the same trend as actual;
2. As NN try to minimum std deviation, the result should wave around the actual data.
My result looks like obeyed these two criterias
Posted by Caoyuan on December 01, 2006 at 12:33 AM PST #
Will the program incorporate the hability to do predictions? Of course the prediction might be as false as my technical analysis ... so you don't bear any responsability.
I would enjoy the possibility of trying this stuff on other stocks to see what happen.
It would be specially cool if I could try the nn for a period in the past and see if the prediction is more or less right or wrong.
For example: I train it from june to decemeber 2005 and see the prediction for jan-march 2006.
Of course the nn should not cheat and look at the data in 2006.
what do you think?
Posted by James on December 01, 2006 at 08:57 PM PST #
Do you use Joone for the NN engine?
http://www.jooneworld.com/
Could you describe step by step what data you feed in and how to do train your NN?
Posted by James on December 03, 2006 at 01:49 PM PST #
Hi James,
The NN engine was written by me and was released in the svn trunk, but as the UI parts are still lack, I have not yet integrated into the released version.
As the feature requests concerned mostly about the traditional methods, such as back-testing, technical indicator and charting, the AI features was pending until I have time.
For the picture above, as I mentioned, it was a fairly long term prediction, about 4 months, and without any data fed back to do adjusting. What's the benefit of long term prediction over than short term? Let's do a simple test:
Assume the short term prediction's accuracy is always better than long term, for example, the 3 days prediction's mean error is 2%, it would be useless, because in the future 3 days, the change of the price may be only within a range of 5%, so the error may be not less enough to win.
On the other side, assume the long term prediction's mean error is 10%, but it last about half a year, so, if the predicted trend is remarkable, the price's change range may be more than 30%, it brings the opportunity for winning.
Posted by Caoyuan on December 05, 2006 at 12:20 AM PST #