Hi guys, hi from Andrea Unger, Walk Forward Analysis!
Good? Bad? What is it?
I’m often asked what I think about it and if I use it?
Well, the short answer is I think is good but I don’t use it.
I’ll explain why.
So, first of all, I also must admit that I found a certain level of confusion sometimes, talking about Walk Forward Optimization or Analysis.
In some cases, there are people who simply think that this is not a simple validation of the system, forward is just to see how the system behaves once you stop developing it and you look what it does in real time.
This is an out-of-sample analysis, which is fine, that is not the Walk Forward Analysis that I’m asked about.
The Walk Forward Analysis is an approach where you develop a system with parameters, then you optimize these parameters over a certain period of time.
Let’s imagine 4 weeks, once you find in these 4 weeks which parameters performed the best or led to the best performance of a system, considering that you might believe that what worked well for the last 4 weeks is supposed to continue working decently well also in the coming week, you put this combination of parameters at work in the coming week.
So 4 weeks you test what happened, one week you use the optimum of these 4 weeks.
Once this week is over, you get back and you take again the last 4 weeks, so you forget about the first week now, now you take the last four, new optimization, new parameters, maybe the same as ideal case, seldom anyway, you find the new parameters and you use them in that sixth week in this chronological description and then you go on like this.
This is one method, another way is when you keep on optimizing on the whole history, so you always add one week and you use in the coming week the optimum of the complete history you have at your disposal.
It is not my favorite approach, ideal approach, because I said I don’t use it, because over when you come to the end of story of I don’t know, 200 weeks you might have some trouble in finding really what is working best in the last period which is the driver of the whole idea.
What has been working well now, should work well also for the next couple of days or week.
This is the whole story and I think it’s good and the real benefit is that in this way of developing systems, you create a large out-of-sample history because you have many weeks where you actually test in out of sample the results.
So you also have an indicator of how good your system is, or how robust your system is, in keeping doing well once it goes live, which is good.
You have a wide period and it is not the classic last two years or first two years, it’s across the whole historical data you have, which is very very good.
This is everything fine, so why don’t I use it?
I don’t use it because the way I develop systems, creates difficulties in identifying what parameters to optimize.
If you develop a system using indicators, you have plenty of parameters to choose from and obviously you can do a good job, but you know if you watched other contents on this page I don’t use indicators in my development.
I use chart patterns and so on, these patterns are listed in some function to use in the development phase, but the patterns are an idea of how the market moves.
I have a pattern that tells me if it has been indecision and now decision, this is a yes/no thing, 50/50.
If I have a parameter here which is 0.5 it’s fine, but it would make no sense to optimize it from let’s say 0.45 to 0.55 with steps of 0.01.
It would be pure overfitting, no longer seeking for the best parameter.
Just because the idea I have is yes it’s an indecision phase, no is a strong decision phase, that’s it.
So how can I really use this?
And also you might say that I could play with a different list of all the patterns, use these two to look for the best pattern.
Yes but the pattern itself is a go/no-go thing, pattern true so you can trade, pattern false don’t trade.
Again, to look for the best pattern it’s probably a cherry-picking the best trades but how can I know that this pattern will occur again in the week that we analyze afterward?
It’s really different as a structure the way I develop systems so that in this structure of mine it’s very hard to apply the Walk Forward Analysis as it is conceived in the original form.
That’s the reason why I don’t use it, I’ve also talked to Kevin Davey who is an expert on Walk Forward Analysis, he knows my approach, he knows my development techniques and he also came to my same conclusion that it’s very hard to apply the Walk Forward Analysis concepts to my way of developing systems.
This, again, does not mean that Walk Forward Analysis is wrong, I appreciate it, I think it’s a good tool.
If in your developing way it works, perfect, use it!
If you know how to make it work for my approach, let me know, I will be interested in it and I will be listening to you carefully.
If you don’t want to use it because you don’t believe in it it’s fine as well, provided you make some money with your trading systems.
Numbers to make decisions on, Walk Forward Analysis is a further application of using numbers which is algo trading, which is what I recommend to everybody.
That’s it guys, no particular secret in this.
Let’s keep in touch,write here to let me know your experience and see you next time.
Ciao from Andrea Unger.