Genetic, “directed genetic”, CRISPR, fuzzy?

I am relatively new to the concept of algorithmic trading so there is a great likelihood that I am missing some crucial points.  

The workflows discussed in the manuals and distilled from different threads on the forum is more or less as follows:

 

look at the price chart > try to spot some regularities >  try to formalize the observation > develop the system > test, etc.  

OR 

add indicator(s) > try to spot regularities, and so on as above..  

 

Another approach is brut force — it may take some, possibly very long, time to stumble on something worthwhile. 

Another is to implement genetic algorithm. Depending on the conditions chosen and on the underlying engine, the result will be better or worse. 

 

All of the above strive to combine a few principles, primarily that the  resulting formalization (i.e., the rule) has to be generalized description of the curve formed by prices in order to avoid overfilling , and the superimposed money and risk management principles.

 

Those tasks are difficult.  Additionally, it seems to me that it is even more difficult to aim for the rules describing trends spanning days, while it is  a bit easier to come up with some solution for shorter time fluctuations. 

 

 

Is it feasible to reverse (or modify) that workflow?  Basically it would look as follows.

 

Look at the chart > identify (mark) the regions of interest, regions of potential entries and exits.   Having the set of parameters  (absolute and relative prices, ranges, indicators, etc, etc), and after editing money and other principal constrains, task the computer to find a set of rules that describes the price movement between marked regions?  In order to avoid overfitting and instead to obtain desired generalized  description of the curve (price movement), introduce variable fuzziness into parameters (either their absolute or relative values).  

 

Is what I described just a particular instance of genetic algorithm, perhaps “directed evolution” ?   Instead of hit or miss I would like to  nudge the program in the desired direction….. 

 

CRISPR is a true acronym taken from molecular biology and serves here as a red herring.  

 

Look at the chart > identify (mark) the regions of interest, regions of potential entries and exits.   Having the set of parameters  (absolute and relative prices, ranges, indicators, etc, etc), and after editing money and other principal constrains, task the computer to find a set of rules that describes the price movement between marked regions?  In order to avoid overfitting and instead to obtain desired generalized  description of the curve (price movement), introduce variable fuzziness into parameters (either their absolute or relative values).

 

I have been looking at this outside of SQ using Machine Learning (SVM, Decision Trees, Bayes).  Trying to predict major turning points using external data sources.

 

It is very time consuming gathering, normalising and aligning the data to the timeseries you are trying to predict.

 

However if you can predict the overall market direction bias, and feed that into SQ (I have simulated this with external indicator data), the results are about 10x more profitable with same drawdown than SQ alone.

(Removed in retrospect because it was an embarrassing beginner question and didn’t add any value, haha)

Welcome mikeyc write some guide or something like that. 

Haven’t got time to write in detail, and haven’t completed my work on this, but here’s a quick summary of what I looked at:

 

I used zig zag indicator to output turning points (price is in uptrand or downtrend) on hourly and daily timeframe using historical data.  This was to identify the most perfect trend following.  Since it was historical data, there was no repainting issue, it merely showed the perfect time to be trading long, or trading short.

 

I aligned these signales to 5 minute data.  I have a C# framework for all this sort of stuff.

 

I imported this into SQ3, as an indicator.  After many hours SQ3 started picking up strategies using this perfect indicator, and the profits and drawdowns as would be expected were superb. 

 

Then I used this perfect turning point data into Azure machine learning studio. https://studio.azureml.net/

 

As predictors I uploaded lots of external data, including retails trader sentiment I have, other indicators, etc.

 

Then tried to find the best way to predict the zig zag turning points using these values.

 

I haven’t concluded this work yet, it is ongoing.  I plan to add economic events from the calendar, currency strength data and correlation data to the models.

Mikeyc

interesting approach !  Very time consuming and ……………. risky is the sense of outcome.  But indeed you answered my question about what I called “direct” evolution.  Do you know if Azure will attempt to crate a “crisp” model, i.e.  over-fitted? If so, do you think it is possible, using any tool, be it Azure or SQ or similar,  to insert a “fuzzy” coefficient into the model?