Neural Programming for Profitable trading | Page 2 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Neural Network Trading, Serious people only! 127 replies

Neural Network trading using Matlab and Metatrader 8 replies

Interested in Neural Network Assisted Trading 4 replies

Growing Neural Gas Algo for FX trading 2 replies

Neural Networks in Trading (Java, Algodeal, Technical Analysis) 3 replies

Neural Programming for Profitable trading

  • #21
  • Jul 30, 2006 11:00pm Jul 30, 2006 11:00pm
Quoting Bemac
How do you compensate for a random set that catches an anomaly Spike that leapfrogs it to the top of the heap for Profit?
That's exactly the problem. You can't really effectively do it (at least I haven't been able to). The NNs that I end up with catch that random spike, and thus look great from a profit perspective, but when you go back and actually look at the trades the NN makes, it nails that one spike dead on, but the other 80-90% of the trades are slight losers, or just barely make a profit.

I've tried training the NNs with different objectives (such as maximizing: (#winners - #losers) * profit, avg profit per trade, etc) with no luck.
  • #22
  • Jul 30, 2006 11:16pm Jul 30, 2006 11:16pm
Quoting Loki
That's exactly the problem. You can't really effectively do it (at least I haven't been able to). The NNs that I end up with catch that random spike, and thus look great from a profit perspective, but when you go back and actually look at the trades the NN makes, it nails that one spike dead on, but the other 80-90% of the trades are slight losers, or just barely make a profit.

I've tried training the NNs with different objectives (such as maximizing: (#winners - #losers) * profit, avg profit per trade, etc) with no luck.
While using the Standard Iteration Optimisation process in the past I have Charted the profit [Basic Smoothed Line Chart] in the Order that they were run.

Then it is a simple matter of looking for the Highest Smooth Bell Curve on the Graph to locate the set of Variables that produce not neccessarily the Highest NetProfit, but more reassuredly, the Most Dependable Set of Inputs.

If you sorted your Random Sets, would you then be able to look at it that way?
  • #23
  • Last Post: Jul 30, 2006 11:55pm Jul 30, 2006 11:55pm
Quoting Bemac
While using the Standard Iteration Optimisation process in the past I have Charted the profit [Basic Smoothed Line Chart] in the Order that they were run.

Then it is a simple matter of looking for the Highest Smooth Bell Curve on the Graph to locate the set of Variables that produce not neccessarily the Highest NetProfit, but more reassuredly, the Most Dependable Set of Inputs.

If you sorted your Random Sets, would you then be able to look at it that way?
At least with my software program, you can't look at any of the disgarded NNs. Basically the program says, "this is the one that is the best according to what parameter you want maximized" and throws everything else away. There is something kind of like what you are talking about. The parameter is "Return on Account * Equity Curve Correlation". The equity curve correlation is basically how close the equity curve (profits) is to a line. This parameter returns the NN that returns the most profit whilst steadily linearly upward (meaning no sharp drawdowns or huge gains). The problem is that in order to have a high equity curve correlation, you need to have a a nearly constant return on investment. The NN has to constantly make trades and profits, even if it doesn't really feel like they are strong trades.

I'd rather have my NN only make a trade if it feels it is good, and not have to worry about how smooth and linear the equity curve is.
0 traders viewing now