backtesting shows that
Traing data 3 days,
then predicate right maybe först 2 hours after the traing data, then any data further than 2 hours will be wrongly predicated. This is my guess, Neuro network can only predicate next 30 minuetes move after the traing period.

ok, forward testing.
This EA has sell all the way down and sell all the way up as well.

This EA only show me how to organize the fopen and fwrite, rest we must put some more code there.

Attached Images

pnn.gif (17.5 KB, 532 views)

The code change for PNN is not ok. Now the PNN only does the sell operation and no one single buy operation anymore.

it does sell and buy operations for me, but after closing a sell position it usually opens another sell position. same for buy positions.

i only tested it for a few hours yesterday night put it was not bad.
at 00:30 (GMT+1) it opened a buy position for example.

My one, no single sell. (I put some Money management there to closs the losers, the PNN immediate resume a sell position again. No single buy OP! This was trist. Need help.

Attached Images

pnn.gif (18.3 KB, 438 views)

oki, now it does buy operation. See if later it remember to do the sell as well.
---
above is that I married the uni scalper with PNN. maybe it is my faulty in the coding. Now I back to the pure PNN code, and it just made a buy now. I have another question.

The author aimed to run the EA on M15 time frame, the trainer collects the past 60 period, that's totalt 15 hours. Why 15 hours? That's is not a who day, shall we skip the asian market time for EUR/USD?

Sorry, I made lots of noisy post in this shread. It is appreciated if the administrator delete my previous posts.

SO, here is what I've learnt.
1) the PNN changed the code to else if (class != 0) { then the system does both buy and sell position.
2) pay attention, if you train the trainer in the test, then the file pnn.date is stored in the /tester/files/ while the one called pnn.date is empty in the /experts/files/, so when you run the EA, it takes the data from /experts/files/pnn.date. if there is empty data, then only buy or only sell situation will happen.
3) when you do the trainer, optimization, do it and run it on EUR/USD M15, use the option "every ticket". use previous 3 trading days data.
4)when you run the EA, put only the EA on the chart, not the trainer. The trainer is best to be trained during the asian market, and do not trading on the asian market.
5) next step, it is maybe ok to test to put the MACD into the array, I guess.

This papper maybe intressting: Robust Novelty Detection with Single Class MPM- in PDF format. It is the whole article.

Metric Embedding for Kernel Classification Rules here also discussed about the PNN.
Gert Lankriet's NN research publication is here

Hi,

I read and read this thread and still don't know what are the statistics and prediction error of this EA and NN ?? For example. My best NN prediction has following characteristics out of sample.

PREDICTION
m1_TS.cht
EURUSD (EURUSD)
Trading
Close 2 1 min bars into the future
2009.01.22 2:13:11 PM

GENERAL ANALYSIS
Statistic
Input Start Date 2009.01.21 12:18:00 PM
Input End Date 2009.01.22 12:15:00 AM
Output Start Date 2009.01.21 12:20:00 PM
Output End Date 2009.01.22 12:17:00 AM
Number of Bars 718

Average Error 0.000551
Correlation (r) 0.993
R-squared 0.985
Mean Squared Error 5.57E-07
% Correct Sign 99.72


So error 5.5pips for 2 min ahead. for 2 bars 5min it's 11 pips. Are you able to show those numbers related to your EA ??

Krzysztof

I am just begin my journey on the PNN. Here I found Gert Lanckriet's papper is intressting, and put here.
Van Gestel, T., Suykens, J.A.K., Baestaens, D.-E., Lambrechts, A., Lanckriet, G.R.G., Vandaele, B., De Moor, B., Vandewalle J. (2001). Financial Time Series Prediction using Least Squares Support Vector Machines within the Evidence Framework , IEEE Transactions on Neural Networks ( Special Issue on Financial Engineering) , vol. 12(4), pp. 809-821.

hello, the papper here is not corrected converted from ps to pdf. can the administrator help me to delete it, as I have attached another correct converted papper in the thread now.

Attached Files

papper.pdf (1.28 MB, 54 views)

I think this pdf is not complete i.e. does not have the most important part -
conclusion fugures.

Krzysztof