Trading methods: from Complexity to Simplicity and Back Again

In our trading methods and experiments here on Forex-TSD and Forex Factrory that way basically to move from complexity to simplicity.

What does it mean in practice? We used some advanced methods in order to preprocess the time series. There are different methods: as for me I was interested in the practical implementations of the jurik type ofdigital filters (the most popular among traders) (Thanks to Kositsin) and the SSA (Singulat Spectrum Analysis Thanks to Klot).

The idea came as I saw an isolated project in the Russian language Alpari forums for using digital filters in order to smooth the time series and then putting it into more complex trading strategies. This was the first digital Brain Trend.

After that this was popullarized on Forex - TSD, and many with my friends here we were together (Camaron, Jaguar , Francis, Albert etc.) into this. On Forex Factory we experimented with SSA and feeding it into more complex strategies like Trend magic, TDI (Traders dynamic index), ASCTrend become the Digital ASCTrend.

Later on, I was surprised to learn that this had an effect, on the mql5 database, the MT5 version of the digital Brain Trend become one of the most rated systems (the second and the third place).

We turned into the complex SSA indicators as the ultimate form of sophistication for an oscillator. The oscillator idea in its best.

In fact in the SSA approach (end-pointed and Noxa) they use complex matrix calculations: eigen decomposition of a matrix.

SSA uses singular value decomposition and extracts the trend, the cycle (seasonal components ) and noises.

Then the Noxa CSSA for example is using a genetic optimizer in order to find the best parameters of the

NOXA CSSA for buy and sell. However that does not work so well in practice and is away from the expected money mochine I have believed to be.

Yes it works but you need to carefully examine the market state using a discretionary approach (I did not find yet a way to make an algorythm for market state detection).

Here are example of scripts using metatrader for Support Vector Machine. There is also a matrix indicator, one of a kind.

So the classical implementation of the SSA is using a method to decompose and simplify the

indormation. And from then to use a direct approach to derive by genetic algorythm or other ai learning way trading rules.

However ithas been argued that Generalizing a problem can make the solution simpler or more complicated, and it’s often hard to predict which beforehand.

So it may sound as a paradox but it may appear that using a complex algorythm like SSA, Independant Component Analysis, Wavelets etc. may actually leave you with a bigger problem.

In fact those preprocessing tools are used because the direct approach of feeding the price into a Neural networks makes it close to impossible to make a good prediction because of the specifics of the neural nets that make a lot of local optima.

However with the rise of the kernel machines, they use a complete different paradigm. They say, yes it exist a simple solution, but in order to find that solutions we may need to map it into high dimensional space and they (the kernel does exactly that), we may find a linear solution into this space.

And for that you do not need to map all the points into this multidimensional mathematical space (that is the kernel trick, brilliant).

So here I make a loop to the beginning, if we have a light matrix indicator we may feed it within the kernel. With SSA ep we cannot practically do that because it is very demanding of computational ressources to make a lot of passes.