I am new to automated trading as of a few months ago. What I understand is the EA’s need to be optimized to correctly work on different pairs/timeframes. I know I have to be careful of over optimization so I developed a few strategies
The step should not be to narrow. If it is really narrow the chances of over optimization are greater. Therefore the EA I am using I use a step of 10 for the 4 variables rather than 1.
Use forward tests as well as back tests.
So I am using the EA names artificial intelligence. It’s the second one though with a take profit built in. The first attempt I used the one without a TP and found it only lost money on the live demo trading.
Now I am on the second attempt and have set the four variables with a step of 10 to a max of 500 and the SL & TP with a max of 800 and a step of 10. I optimized them all on a 1H chart at once (don’t know if this is right or not) from beginning 2010 to beginning 2013. Which is a period of three years. I then took the results and tried each one of them from beginning 2013 – Beginning April 2014. If the curves continued to look good as they did in the back testing/optimization period, then I threw them onto a demo account. I also took into consideration the total trades as well as profit factor and win %.
I have 20 pairs and gold running currently. So far I am up 4.5 % closed profit in a week and a half. This is good but I want to design better. One of the things I noticed was my SL and TP were too high for some pairs so I want to make round 3 have tighter stops take profits, I was thinking of setting it to 400 or 500 rather than 800. This will also make the total number of trades higher as well.
Basically what I am looking for is a system that runs on at least 20 pairs so the risk is spread around. I want the turnover to be higher rather than lower so the trades over a year got to be at least 20ish. But as i go to optimize again I am wondering if I am using the right time frame? And my biggest question is am I using the right amount of optimization time for forward testing time? Should my back testing/optimization time change as the time frame changes? Is 2010 too far gone for me to be worried about? If I want to go down to a 15 min or 30 min chart, what should be periods be for back testing/optimization? What about a 5 min chart?
In short you want to make sure that if you look at say 100 different configurations of your strategy over the last say 6 months in test. If one config is in the top 10% in test, then when you trade it you expect that it will still be in the top 10% when you live trade it.
The overall strategy should show a good walk forward efficiency if you want to trade it. Its worth seeking out DarwinFx on here as he has a cool tool for doing this with MT4 that you might find useful.
What you might find when you go to lower time frames is that you get randomly stopped out more often. You need to start taking into account things like economic releases etc. Its doable but you have to be very careful with stop losses since its more likely that extreme events will occur on lower timeframes.
I personally think it is a lot better to be trading 5 min charts for a few practical reasons. If you want to test a strategy live then if its trading ~2-3 times per day then you can over a month get a good feel for its live performance. If you are trading once a week, this same live testing period is going to take you 15 times longer.
Okay thanks for your reply on this. Last time i looked into the multistep WFA i got confused and after reading the section in that book you put in the link, i still am.
So it says you have 10 years of data and the WFA breaks it into 30 sections which is 4 months and then does a walk forward test on each which is about 25% so, a month or so. So each of these periods gets optimized separately with the best parameters chosen, then these are taken and walked forward for the next month to see if the results are similar. There is a number for each and you do this 30 times.
What i don’t get is each of these different optimization periods are going to come up with different parameters. All different parameters will all have a varying degree of success on the following walk forward. What is the point of doing this if all parameters will be different in each optimization period? This is where i get lost. At the end of the 30 walk forwards, you still dont have the best parameters to use. All you have is data from a tonne of different optimizations and following walk forwards. The majority of which will likely not work since they will just a computer optimization with an insane profit factor, unless of course you sift through them manually to see which ones still work in the walk forward period (exactly what I am doing now).
It says something about comparing these numbers to check the robustness of the strategy. That`s fine but all strategies will work with the right parameters. I compare this WFA to what I have been doing on my past systems and it looks like a whole lot more testing and work and I just cant figure out what you do with all that info at the end. How do you turn 30 different optimizations into any sense and something that can be taken forward into the future?
I also understand what you said about the lower time frames and testing. However i am using a EA that i know works the way its supposed to, i didn’t create it and its been around for years. Therefore, is any demo testing actually required for a EA like this? Or does the walk forward just need to test out good? Originally I wanted to try and stay out of the smaller time frames because of the fundamental shocks that are more likely to occur.