Redes Neuronales y similares

De una conversacion de Linked In:

Neuroevolutionary approaches to currency trading.

I’ve been using Neuroevolutionary approaches in forex trading, and even created indirect encoded neural network systems that used not the sliding window of prices, but the actual geometric representation, in hopes of improving generalization. A short peer reviewed paper on this approach is available at ACM (and a long preprint on arxiv) and also included in my book on neuroevolution [1,2]. The results were actually better than simply using a direct encoded NN sliding windows with regards to generalization. But still, the approach did not seem to produce the results I hoped for.

Has anyone explored other interesting machine learning systems, and what was your experience? It would be great to build up a list of machine learning approaches attempted, and general results acquired, training and generalization behavioral patterns. For example, has anyone tried SVM? or boltzmann machine, or genetic programming, or trees, or perhaps baysiann networks? and if so, what was your experience with them?

I’ve tried different ML and Evolutionary Computation methods such as NN, SVM, GA, PSO and etc. According to my experience, you need to combine several methods to achieve acceptable results.

In my experience, sometimes SVM generalizes better then NN, sometimes worse. As they have different inductive biases, you could experiment with combination of these (and also consider other ML systems with different biases).

If this is a propietary question, then my apologies but I’ve never understood the following nor have I ever tried to use these approaches to predict the market. ( I use an econometric approach ). My comment/question is the following: Whether one uses an NN or an SVM or whatever, the defined output is critical also. ( input is also but people talk more about input question than output question ). By this I mean, do you define the return over a particular time horizon ? Do you code it as +1 buy, 0, neutral, -1 short or is the output actually the numeric return over the given specified horizon ? Given that the return is over a horizon, to me it seems like NN and SVMs are not so well built for prediction of returns because the returns are over horizons that need to specified apriori ? Thanks for any enlightenment on this issue.

Has anyone experience with trading with the Puzzle Based Learning systems http://cs.adelaide.edu.au/~zbyszek/Papers/PBL-IEEE.pdf

Wow! there’s a lot to to this algo trading stuff. I was taught trading by one of the best traders in the world, Mike Baghdady. His approach to trading is simple and very effective. By watching price action and support and resistance levels when you enter or exit a trade the highest probability of profitability is likely. There simply is no better trader/mentor than Mike Baghdady.

I think that using the actual geometric representation can’t give you an edge over sliding windows. I have been using the Singular Spectrum Analysis in order to smooth the time series and use them as an input, and that gives some improvement compared to the use of raw price series.

However the best results I have achieved was to combine Neuroevolutionary approaches with chaos theory and then to apply this on a market that matches on a fundamental point of view.

John, regardless of whether you think or not of how effective geometric representations are in neural analysis, benchmarks and experimentation show that the method is more effective than the standard NN approach. It is not the utilization of geometric patterns with standard Neural Networks, but with indirect encoded neural networks that are specifically made for geometric analysis.

Neuroevolutionary approaches to currency trading.

http://www.erlang-factory.com/upload/pr … n_2012.pdf

I develop and test with my team in the last 2 y NN and SVM.. we spend >10.000 h in coding ..conclusion ? i am now adding new inputs to the NN and genetic algorithm for reinforce learning and optimizations algorithm where the target for GA is reduce DD /time/and > profit .
Best results we have base on NN Correlations /Cointegrations and Garch .Basically we have 40-60% a year profit .. with a 15% drown down on portfolio (60-100 trades /day).

Guys I was reading some samples from the book Handbook of Neuroevolution Through Erlang. I think that this is a master piece on the subject. The handbook is not hiding behind copy/paste mathematical formulas. It is written with deep understanding and special emphasis to make the subject user-friendly.

I’ve reread what I wrote, and I realized that what I wrote about 2d and 3d data representation might sound crackpotish if one is not familiar with substrate/hypercube encoded Neural Networks. This particular indirect encoding approach was popularized just a few years ago in the system called HyperNEAT. What you have is basically a multidimensional substrate, meaning you have a group of neurodes, each with a coordinate in this multidimensional space, connected in some topological manner. The synaptic weights and connections between these neurodes is defined by a neural network that you are evolving. You feed to this evolving neural network the coordinates of the neurodes in the substrate, and the output of the NN is the synaptic strength, connectivity expression, and other needed parameters. Once the NN has produced the synaptic weights for all the connections in the substrate, then this subsrate can begin interacting with the environment, or in this case the financial data.

To present the data to the substrate, it has to itself be represeneted as an input substrate with coordinates. Hence the NN will define the synaptic weights between the neurodes and this data you are presenting to it, since this data has to itself be represented as an input substrate as well, so that it has coordinates. Thus we end up with 1, 2, 3 and other hyper dimensional input substrates. So an actual candlestick chart is a 2d graph, we can parse it, define a resolution we want it to have with regards to input substrate presentation, give each pixel in this graph a coordinate, and it would represent a 2d plane fed to the neurodes in the substrate. Thus, geometrical information is natively retained in the candlestick chart, and the substrate can analyze it, a substrate encoded system can much better (from published research papers) deal with geometrical patterns. Now if we want to feed multiple candlestick charts, the input substrate that these 2d planes will represent, is a 3d system. We simply stack the 2d charts on the z axis (given that the 2d charts are on the x,y axis). Thus the input is a 3d substrate. And in this manner we can feed multiple charts to this substrate encoded system. This provides significant flexibility in data presentation and NN compression, and interesting opportunities. A more detailed elaboration on this indirect encoding approach in general can be found on the net, if you search for the following terms: HyperNEAT, Hypercube encoding, and Substrate encoding, and of course in my book

From one hand those guys are hired by the HF quantshops where an edge is found in the speed of execution. Then the edge needs deep infrastructure understanding and market structure understanding that are completely different fields of knowledge. Only by combining this meta knowledge those quant shops succeed.

Other possible way is to look towards econometrics and deep market knowledge. Then you may be able to find a good definition of the problem and apply correctly the model. The economist need to define the problem the AI specialist has to apply correctly the model.

Those are two main models I know that lead to an edge. There may be other ways in order to beat the markets but they are fringe like entropy analysis, market states clustering and chaos theory etc.

Gene Sher • Correct, but also in my book we build a state of the art memetic algorithm based neuroevolutionary system, distributed and concurrent, called DXNN. Based on benchmarks, it’s leading in the neuroevolutionary field. It’s also much more general, and has features not available elsewhere. You can set DXNN to behave as a genetic algorithm rather than memetic algorithm based system by setting the local tuning to 1 step (this will disable the 2 phase approach).