Neural Networks - page 20

[name]

After "chewing" (superficially though - still need time) it looks like we are never going to be using ANN effectively. We (the small potatoes) do not have enough money for the hardware required to use ANN as it should be used

[hughesfleming]

On my own....why all the negativity?

This is not so complicated. First you have to figure out what it is that you are trying to predict. The next issue is noise. You need to make sure that your model is not fooled by randomness and data mining bias. There are free tools out there that will allow you to develop a proof of concept before you spend a lot of money on hardware.

If you are predicting trend for example, you need see if your model has an edge over a simple average. If you are predicting turning points, first look to see if you can out perform a fourier transform. If the edge is small, then you have to ask yourself if the training time is worth the effort. It does not matter if you use a neural net or one of the various support vector machines, you still face all the problems of curve fitting the past to predict the future. Out of sample these systems tend to fail because any success you might have had in your testing might still have been down to randomness. No amount of money invested in hardware will solve that problem.

Now these tools can be useful but you will need to have your expectations in check. If a model can improve a strategy by a few percentage points then over time and a large number of trades, you will come out ahead.

regards,

Alex

[name]

hughesfleming:
On my own....why all the negativity?

This is not so complicated. First you have to figure out what it is that you are trying to predict. The next issue is noise. You need to make sure that your model is not fooled by randomness and data mining bias. There are free tools out there that will allow you to develop a proof of concept before you spend a lot of money on hardware.

If you are predicting trend for example, you need see if your model has an edge over a simple average. If you are predicting turning points, first look to see if you can out perform a fourier transform. If the edge is small, then you have to ask yourself if the training time is worth the effort. It does not matter if you use a neural net or one of the various support vector machines, you still face all the problems of curve fitting the past to predict the future. Out of sample these systems tend to fail because any success you might have had in your testing might still have been down to randomness. No amount of money invested in hardware will solve that problem.

Now these tools can be useful but you will need to have your expectations in check. If a model can improve a strategy by a few percentage points then over time and a large number of trades, you will come out ahead.

regards,

Alex

Alex

Thanks for the reply

The reason why I told what I told is that, unless we are not well equip ed with hardware, ANN calculation will always be "incomplete". And then it comes to what we are already doing : estimating using our own NN

[hughesfleming]

Ok...got it. If you do feel like experimenting, I did layout one way to use Rapidminer with metatrader here just in case you did not see the thread. https://www.mql5.com/en/forum/181252

[name]

hughesfleming:
Ok...got it. If you do feel like experimenting, I did layout one way to use Rapidminer with metatrader here just in case you did not see the thread. https://www.mql5.com/en/forum/181252

Alex

Thank you for all your help

[nevar]

I have several promising out of sample tests like the one I attached.Finding profitable algos with AI is all about your inputs and that is an art not science .When I was searching different algos for live trading I came a cross Recurrent Neural Network-Restricted Boltzmann Machines in python Modeling and generating sequences of polyphonic music with the RNN-RBM — DeepLearning 0.1 documentation and Bernoulli Restricted Boltzmann Machines feature extractor and LogisticRegression classifier in python Restricted Boltzmann Machine features for digit classification — scikit-learn 0.15.2 documentation .They sound quiet interesting does anyone have any experience with any of these two algorithms for live trading ?

[tampa]

Neuro trend cloned (indicators + template)

neurotrendnncloned.rar

[Joakin]

Useful link : Neural Network Tutorial

[seekers]

Artificial Neural Networks : artificial_neural_networks.pdf

In this note we provide an overview of the key concepts that have led to the emergence of Artificial Neural Networks as a major paradigm for Data Mining applications. Neural nets have gone through two major development periods -the early 60’s and the mid 80’s. They were a key development in the field of machine learning. Artificial Neural Networks were inspired by biological findings relating to the behavior of the brain as a network of units called neurons. The human brain is estimated to have around 10 billion neurons each connected on average to 10,000 other neurons. Each neuron receives signals through synapses that control the effects of the signal on the neuron. These synaptic connections are believed to play a key role in the behavior of the brain. The fundamental building block in an Artificial Neural Network is the mathematical model of a neuron

[seekers]

forecasting_ability_but_no_profitability_-_an_empirical_evaluation_of_genetic_algorithm-optimised_tr.pdf

This paper evaluates the performance of several popular technical trading rules applied to the Australian share market. The optimal trading rule parameter values over the in-sample period of 4/1/82 to 31/12/89 are found using a genetic algorithm. These optimal rules are then evaluated in terms of their forecasting ability and economic profitability during the out-of-sample period from 2/1/90 to the 31/12/97. The results indicate that the optimal rules outperform the benchmark given by a risk-adjusted buy and hold strategy. The rules display some evidence of forecasting ability and profitability over the entire test period. But an examination of the results for the sub-periods indicates that the excess returns decline over time and are negative during the last couple of years. Also, once an adjustment for non-synchronous trading bias is made, the rules display very little, if any, evidence of profitability.