When you put two models in an ensemble you will surely get a lower trading frequency (because the two models are bound to disagree under some conditions), however this does not in principle guarantee worse results (read about statistics like the ...

Machine Learning with algoTraderJo

Good posts! Nonetheless there are some important conceptual problems with the hypothesis you have made. I think it would be most interesting if you look at the actual results of the ensembles and tell me how these compare to your predictions and ...

Machine Learning with algoTraderJo

I would like to propose an exercise. Here are the results of a neural network that uses the next bar's return as target and a neural network which is equal in every other way except for the fact that it predicts the bullish/bearish excursions within ...

Machine Learning with algoTraderJo

You have made some excellent points. You are right in that ensemble results depend heavily on how you interpret different scenarios, it also depends heavily on how many systems you use. There are many different ways to interpret ensemble results and ...

Machine Learning with algoTraderJo

Depends strongly on how you interpret ensemble results. In this case I did not take a lack of agreement as a close signal but merely as a "don't do anything" signal. Therefore open positions remain open if the ensemble components disagree.

Machine Learning with algoTraderJo

All systems here are equally reacting. They all make one prediction on every daily bar and they always make either a short or long prediction. In this manner all systems generate an equal amount of signals, therefore success or failure of the ...

Machine Learning with algoTraderJo

This is actually a very common misconception. Correlation does not determine whether an ensemble will work or not work. The correlation coefficient between the two models in the ensemble that gives worse results is actually not very high (0.53) ...

Machine Learning with algoTraderJo

I am not taking the past 100 days, I am taking the past 100 examples. Each example uses the past three day returns as inputs and the next day's return as target. I am using an IRpropPlus optimizer to optimize the feed forward neural network to a ...

Machine Learning with algoTraderJo

To continue moving in the exploration of NN I am now going to show you the effect of making variations in the hidden and output functions as well as the number of optimization cycles per training for the NN using 100 examples to train (last three ...

Machine Learning with algoTraderJo

The Asirikuy framework is not coded in R, neither does it require any R coding (you can optionally interface with R if you wish). The framework is coded fully in C/C++. It currently has FANN, Shark and Waffles to do machine learning (all these are ...

Machine Learning with algoTraderJo

As a rule of thumb always prefer several different categorical binary inputs instead of single more complex categorical inputs, they allow you both to test information more specifically and they usually give better results (IMHO). It's great to hear ...

Machine Learning with algoTraderJo

Adding the 4 inputs per bar that you suggest and comparing the results with the SVM standard test I have been using here (145 examples, 20 inputs) gives the results I post below. This means that this model now uses 100 inputs (from the past 20 bars, ...

Machine Learning with algoTraderJo

Change your hidden layer and output neurons to Hyperbolic Tangents (instead of the FastSigmoid/Linear neurons used on the last post) and you get very different results: image Changing functions means changing everything, so you need to be very ...

Machine Learning with algoTraderJo

As training is important when building NN, we can also try an ensemble using rNN models with different levels of training, then act only on a majority vote of the models. This means that we are attempting to extract different levels of ...

Machine Learning with algoTraderJo

Sorry I'm having trouble understanding what you posted. I would need you to be clearer so that I can run these tests. What are the inputs for each test? what are the targets? How many tests? How do you propose to act on the targets (what is the ...

Machine Learning with algoTraderJo

On the latest results I posted I use the rNN output to make a decision based on directionality (positive long, negative short) but you can certainly try more creative approaches. This is the code for the NN Shark function I used for this example: ...

Machine Learning with algoTraderJo

Weighting ensembles is something we have not covered yet, but what you suggest is part of the things I do. There are many ways to decide how to weight the voting of a model ensemble. There are also many ways to interpret the voting. This becomes ...

Machine Learning with algoTraderJo

I probably did not explain myself very clearly before I do not want to know exactly how you trade or what ideas you have for discretionary trading (sadly I don't have the time to read it with the attention it deserves and come up with an ML ...

Machine Learning with algoTraderJo

Since we have both NN (100 examples, 3 inputs, 500 cycles per training) and SVM (145 examples, 20 inputs) models with decent results (the first a regression model and the second a classification model). It's cool to attempt to build an ensemble of ...

Machine Learning with algoTraderJo

An important issue with NN implementations is how much to train them. There are two ways to do this, the first is to train the NN down to a given error limit (for example train until the error is below X threshold) and the second is to train for a ...

Machine Learning with algoTraderJo