Systematic trading | Page 26 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Trade another day through Systematic Trading 203 replies

Nonparametric Association Measures in Systematic Trading 57 replies

Anyone use Systematic from Alpari? 1 reply

A Systematic Approach to Markets 68 replies

Systematic Trading vs. Discretionary Trading 14 replies

Attachments: Systematic trading
Exit Attachments

Systematic trading

  • #501
  • Dec 17, 2009 2:09am Dec 17, 2009 2:09am
Do the models you have evolved for 2009, work well in 2009?
The breaking of a wave cannot explain the whole sea.
  • #502
  • Dec 17, 2009 2:22am Dec 17, 2009 2:22am
Quoting Craig
Do the models you have evolved for 2009, work well in 2009?
Good question. Yes they do but I can never know if they work only because of raw curve-fitting (one year period is just too short to know) which is a very probable reason.

That said, I have made same test for multiple other years and for other years, models combined seem to be working okay on almost all the 10 year scale.
  • #503
  • Dec 17, 2009 2:26am Dec 17, 2009 2:26am
Quoting mikkom
Good question. Yes they do but I can never know if they work only because of raw curve-fitting (one year period is just too short to know) which is a very probable reason.
Do you only optimize on large time-frames, or have you tried smaller time-frames?
The breaking of a wave cannot explain the whole sea.
  • #504
  • Dec 17, 2009 2:29am Dec 17, 2009 2:29am
Quoting Craig
Do you only optimize on large time-frames, or have you tried smaller time-frames?
Mostly larger although I now have some smaller time "components".

edit: This is one of the reasons I would like to hear how other strategies have worked, especially those straetgies that have been in use or can be backtested before 2009/8.
  • #505
  • Dec 17, 2009 2:32am Dec 17, 2009 2:32am
Quoting mikkom
Mostly larger although I now have some smaller time "components".
I guess the reason I asked was because most longtime traders seem to agree this has been a tough year, RTM isn't working and the volatility has been dying away. This all got me thinking about GA again and if it was capable of finding something in these markets.
The breaking of a wave cannot explain the whole sea.
  • #506
  • Dec 17, 2009 2:43am Dec 17, 2009 2:43am
I would even go so far that I would claim that 2009 has been the hardest year for any FX strategy. I have been pondering a lot of moving again back to some futures and stock trading (not moving away from FX but distributing the risk) but if I do that it would mean expanding the framework to those markets and that is going to be a lot of work (not least because I have to rethink the memory/disk aspect - at the moment I have all the data in memory).
  • #507
  • Dec 17, 2009 7:20am Dec 17, 2009 7:20am
I think there is a ideological shift in the financial markets. During the beginning of the year the fear of the USD was termendous, to the point that everyone you see on TV and read says how the USD is going down quicker than the titanic. I would think that fear would shift a whole world market just as the greed of the US equity market shifted all the world's perspectives during the '98-99 bull market run.

It's to the point now that my friends who know nothing about any of this are talking to me about shorting the dollar and buying gold. Makes me think a new change is about to occur.

Just my thoughts and two cents.

Happy Holidays to all, how ever you may practice them!
A woman drove me to drink, and I forgot to thank her
  • #508
  • Edited 8:07am Dec 17, 2009 7:29am | Edited 8:07am
Well, last September we saw some extreme volatility everywhere, and as the market was correcting this effect in 2009, we had a rather unusual year. Also, as time goes by, more and more money is traded electronically/algorithmically.

So my gut feeling is that we will never have exactly the same situation as the period 2000-2008, but starting from 2010 the markets will most likely start reverting partly to what it was in the period before 2009. So I guess your models will do better next year compared to 2009, if this happens, but probably not as good as in the backtests. But I don't have hard facts about that, just my estimation.

PS: the best thing of course would be if your models where self adaptive, but this is easier said than done
  • #509
  • Dec 17, 2009 8:30am Dec 17, 2009 8:30am
Quoting challenger78

PS: the best thing of course would be if your models where self adaptive, but this is easier said than done
Self adaptive I like the sound of that. I have been working on this. Progress very slow though, I need to get sacked.
  • #510
  • Dec 17, 2009 8:58am Dec 17, 2009 8:58am
Quoting challenger78

PS: the best thing of course would be if your models where self adaptive, but this is easier said than done
The biggest question is: self-adaptive to what? I could very easily set the framework to generate models that have worked very well for the past 2 weeks ,[and in longer periods in the past] every day - this is the ultimate in "adaptivity" imho. The real question is, what to adapt to, do the last X weeks for example really tell how the market currently is?
  • #511
  • Dec 17, 2009 9:17am Dec 17, 2009 9:17am
Quoting mikkom
The biggest question is: self-adaptive to what? I could very easily set the framework to generate models that have worked very well for the past 2 weeks ,[and in longer periods in the past] every day - this is the ultimate in "adaptivity" imho. The real question is, what to adapt to, do the last X weeks for example really tell how the market currently is?

Curve fitting is our friend. (trying to proof this is the hard part)
  • #512
  • Dec 17, 2009 9:23am Dec 17, 2009 9:23am
Quoting spekkiefx
Curve fitting is our friend. (trying to proof this is the hard part)
In some ways curve-fitting is what every system developer is doing. The hard part is to somehow make sure you curve-fit to the correct things.
  • #513
  • Dec 17, 2009 9:33am Dec 17, 2009 9:33am
Quoting mikkom
In some ways curve-fitting is what every system developer is doing. The hard part is to somehow make sure you curve-fit to the correct things.

Curve fit for the last 2 weeks and you role this on for each day looking back 2 weeks what would the end results be? Compare 2 weeks with 3 and 4 weeks. This is what interests me.
As mentioned maybe the market is changing because of all the robots. This type of robot could take advantage of the rapidly changing market. I have some early results that are interesting but still very early stage.
  • #514
  • Dec 17, 2009 9:48am Dec 17, 2009 9:48am
Quoting spekkiefx
Curve fit for the last 2 weeks and you role this on for each day looking back 2 weeks what would the end results be? Compare 2 weeks with 3 and 4 weeks. This is what interests me.
As mentioned maybe the market is changing because of all the robots. This type of robot could take advantage of the rapidly changing market. I have some early results that are interesting but still very early stage.
I actually don't think that would work (too little data points to make meaningful predictions) but maybe I'll try it in the future. One possibility that I have been pondering is to alter risk based on how well the algo/model has worked in the past X period (period should be selected so that it represents meaningful length for the given algo)

Meanwhile I have been testing some other models that try to cut the risk as soon as possible and it seems these kinds of models work on 2009 as well as they did in the past so it seems I have to implement a new, all algos on all timeframes system too to test this further (at this point the risk cutting is made very, very simply and I want to explore this further on smaller timeframes).

It's really pleasing to have this framework so I can test ideas in a kind of monte carlo way (lots and lots of in some way random algos with some similar attribute or bias to certain ways of managing/opening/closing trade) to so I can see what makes difference and what does not.
  • #515
  • Dec 17, 2009 10:18am Dec 17, 2009 10:18am
Quoting mikkom
I actually don't think that would work (too little data points to make meaningful predictions) but maybe I'll try it in the future. One possibility that I have been pondering is to alter risk based on how well the algo/model has worked in the past X period (period should be selected so that it represents meaningful length for the given algo)

Meanwhile I have been testing some other models that try to cut the risk as soon as possible and it seems these kinds of models work on 2009 as well as they did in the past so it seems I have to implement...
Your quite right, in this form it won't work....but I am merely highlighting the principle. I am focused on the hourly time frame and look at closing the trade within the day.
  • #516
  • Dec 17, 2009 10:53am Dec 17, 2009 10:53am
Quoting mikkom
I would even go so far that I would claim that 2009 has been the hardest year for any FX strategy. I have been pondering a lot of moving again back to some futures and stock trading (not moving away from FX but distributing the risk) but if I do that it would mean expanding the framework to those markets and that is going to be a lot of work (not least because I have to rethink the memory/disk aspect - at the moment I have all the data in memory).
I'm doing exactly that now, the automated stock trading starts in the new year. I did a lot of optimization to make sure the memory footprint is light, I can also distribute across multiple processes and/or servers.
The breaking of a wave cannot explain the whole sea.
  • #517
  • Dec 17, 2009 4:29pm Dec 17, 2009 4:29pm
Quoting challenger78
PS: the best thing of course would be if your models where self adaptive, but this is easier said than done
...adaptive systems that are based on flexible, not fixed amount of indicators
  • #518
  • Dec 18, 2009 12:10pm Dec 18, 2009 12:10pm
Quoting mikkom
The biggest question is: self-adaptive to what? I could very easily set the framework to generate models that have worked very well for the past 2 weeks ,[and in longer periods in the past] every day - this is the ultimate in "adaptivity" imho. The real question is, what to adapt to, do the last X weeks for example really tell how the market currently is?
How about walk forward optimization to find the optimum re optimization period without overfitting?
  • #519
  • Dec 18, 2009 12:52pm Dec 18, 2009 12:52pm
Quoting challenger78
How about walk forward optimization to find the optimum re optimization period without overfitting?
I suggested this some time ago, I know another trader who uses GA successfully and does this very thing, almost like a 'meta-optimizer' or a optimizer for the optimizer.
The breaking of a wave cannot explain the whole sea.
  • #520
  • Edited 3:19pm Dec 18, 2009 3:01pm | Edited 3:19pm
Quoting Craig
I suggested this some time ago, I know another trader who uses GA successfully and does this very thing, almost like a 'meta-optimizer' or a optimizer for the optimizer.
I have had that thought in my mind, however it would require me to build a quite different result measuring system than I currently have (and strategy combinator) so I just haven't had time to test this yet.. I tested one variation already where I weighted risk based on the past trade DD and it smoothed curve quite a bit.

I haven't forgotten the actual walk forward version either.

The big question however is, does the last X period really represent what the market is right now [markets change in a second so it might be a better idea to use a strategy that loses small at bad periods and wins at good periods] and if it does, then the best method for me would be to generate strategies that have worked in the past X period, not just optimize older strategies.

By the way craig, have you taken look at cuda (I know you use C)? I'm thinking of using it in my framework in the near future to get some extra processing power, the java bindings look very c'ish (malloc, pointers and all).

(For those people who don't know what cuda is, here is a nice video about what you can do with few graphics cards and cuda http://fastra2.ua.ac.be/)
0 traders viewing now