Quote from slacker:
That is a good application for GAs, called a 'classifier' role. Train bots to identify uptrend, downtrend and chop.
Rosy spoke of "state" and you think in terms of the above. Both sound like making money by extracting the full offer of the market. I'll answer your Q's below in that light.
Problems with GAs, what does the chromosome look like?
Trend is the chromosome. Most do not know what trends look like and what their parts are.
An array of bytes or a b-tree structure?
Neither. I think this is most people's stumbling block. They are not open to the correct realization.
How do you splice 'crossover' the b-tree at evolution time?
Markets are not evolving. BUT you think, and others like you, think they are. Certainly, information flows in structures differently, BUT basically, and in detail, markets have two sides, operators and regulators. To participate, you need money or instruments. Wall Street makes up new instruments and plugs them into kinds of markets; those with capital get handled by the operators and the instrument makers.
How to avoid overfitting with a GA?
Overfitting is a myth and is exclusively associated with applying improper maths. By using the correct math and applying it to the real system wherethe informations originates, then you can get beyond the "state" and its characterizations that are too rough to have statistical significance.
If your chromosome is large enough to solve the problem it is large enough to 'remember' key positive or negative specific in your sample data....
Ah yes ....... the examination is of the trend forming and it has descrete and significant items. They are named and their calss is called "events". There is no thing called "sample". there are only facts from the flow or process of information being outputted in a labelled order. It is, of course, filtered and weighed.
Where can you get enough sample data to completely test insample/outsample?
This is the simplest question. BUT, constraining the thinking to improper and unfounded regions has no significant yield. The stepping stones are uncomplex. As a result there are 10 price items and 11 volume items which make up the "continuation" portion of a trend. Trends have two ends and they collectively have 35 items.
All items are in one class: events.
How do you insure that a system that is 'random number' based (evolution) is bug free?
Systems for the operation of markets are extremely limited. The cause is singular: variable granularity. You are not thinking in terms of the proper mathematics. Change this error of yours. Buy some Legos; they will teach you what Sister Montessori taught with rational numeration systems.
What is the 'fitness' value in a trading system you are going to evolve the system to find?
100% if you do "round" and do not use irrational numbers for variable prices.
Net profit? (too risky, large drawdowns but big profits)
Pragmatically, you take the full offer of the market less premium. Premium means the cost to participate whatever you may have to pay. If you are human, you use 10 to 100millisecond observations. If you extend the human limitations with tools, the full offer is available always in liquid markets.
Sharpe ratio?
On slower fractals the tested annualized result in unsleective universes is 60 plus (third party testing)
Develop your system and then put in on cheap GPU cards for parallel processing in CUDA.
GAs are much more fun than NN as you can see them learn. NN are more for predicting,
This is an error in design, Predicting is a handicap in a rough system.
GA are more about responding to events.
Another error of design. It is never necessary to be this inefficient and ineffective since the Order Of Events is known before the event occurs.
Good trading...
