[UrmaBlume]

Originally Posted by Ranger »

Hi UB Thanks for sharing the Haar ELD, that was nice of you and thanks for your valuable insight. We locked our Haar file to constrain our strategy. In our posting, it's clear that we want other contibutors to join our team with the objective of producing good trading strategies because without a good trading strategy, traders are doomed to failure.We have shared a couple strategies that can be easily modified; optimized and further used in conjunction with the trader's other methods or for educational purposes. The code is open source and simple. That's our contribution - strategies. Can you help the forum by sharing a strategy that you developed?Regards, RANGER

Ranger,

Thank you for the kind words.

While we do indeed do a lot of work with strategies, automated systems and intelligent agents and I am happy to help with some of the basics of these protocols and technologies - certain commitments keep me from posting code.

From what I can read here I would suggest that much more attention be paid to input selection and input preprocessing.

Also neural networks by themselves, for us, don't make a complete strategy - you need to test NN output against regression splines such as MARS and then when you have evaluated which is best the outputs need to be optimized via genetic survival of the fittest routines and then actual buy/sell signals developed using decision trees, bayesian networks and or rules generators.

The tease of NN's is that all you have to do is throw some data at them and they will learn enough to provide answers - far from true and the training and operation of a NN is, in our experience, less than 5% of what it takes to build a functioning, profitable, automated, intelligent trading strategy.

cheers

UB

[Ranger]

Hi UB

It’s always an honor to have you comment on my post.

After reading your posts, it’s clear that our modeling techniques are different than yours, respectfully note:

We do pay particular attention to our input selection; perhaps different from your technique, our preprocessing consists of selecting representative data for our model(s). We do filter data, hence our use of the Haar Wavelet but we consistently strive for selecting a balanced diet for our children.

We think it’s unnecessary to compare the model outputs to any form of linear regression using our modeling techniques. We understand your technical basis but because our strategy formation and are optimization techniques are different, we see this as unnecessary.

In connection with optimization and strategy development, note:

Sometimes, we build a model using several or many inputs. We then re-run this model with varying initialization values because we want to generalize it, then we:


a) form a model committee; each committee member has the right to vote on Entry/Exit - Consensus polling or;

b) we average the optimized output values and then optimize for Entry/Exit and;

c) Sometimes, we make model contests – i.e. if longs are even with shorts we remain flat, go consensus voting;

d) We also create ideal input groups based on factors consistent with their strengths, then form model committees; contests.

We agree that the task of modeling can be complex but the task of understand strength and weakness is equally challenging and it’s well within the technical grasps of our forum members.

Thank for contributing to this posting UB.

RANGER

[Jeremytang]

Hi UB,

I'm curious, if I may ask... What is the output that your NN generates?

Jeremy