Statistics of a anti-grid like system - page 4
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Thanks for the explanation, zzuegg. While it would be nice if you had the first system with only upside, perhaps that is a bit too much to expect!
I have not looked seriously at grids or antigrids until recently. I was highly skeptical of the grid concept when I first heard of it, especially because some people claimed they could make money in a completely random market, which cannot be true. antigrids I literally only heard of this month, and the little I understand makes more sense (as the profitability is based on the tendency of the market to trend. Am I right that you are buying into a trend in stages, winding down in stages if price moves against you, with some sort of profit taking in big moves?
One thought I had which has surely been anticipated is to wind down positions more rapidly than you increase them. So while one unit might be added for each step forward, two units might be removed for each step back (but not past zero). The idea is to keep some of the profit even if you don't hit a profit target. How does this relate to what you are doing?
@zzuegg, am I right in guessing that you have mildly optimised the parameters of the grid system?
I have another thought for a variation on this system. You have a highly variable position size, which is being increased as a trend develops in one direction and decreased as a trend reverses. My theoretical view is that you can't really justify such large changes in leverage, so something to aim for might be to achieve the same idea of having some scaling in and out, but not to the same degree.
Anyhow, it's interesting to compare the statistics to mildly successful stop and reverse systems I have been developing. I think your system may have better performance on the grounds that the drawdown is a smaller fraction of the profit.
Both of these systems are very straightforward trend following systems. They are always long or short with a single position size. The first one uses a single indicator, with its length optimised on 2001-2009 and then run on 2010-2011. The bit at the end of the graph is a bit misleading, as it is only the last 3 months out of 19. Perhaps slightly more frequent reoptimisation was necessary. Note the very low win rate, not uncommon in simple trend-following systems, I believe.
The second is a system which uses 4 indicators of different lengths to determine when to stop and reverse, again optimised on 2001-2009 and run on 2010-2011. The result is a lot less trades and significantly better performance. Some of this may be luck - the bottom lines of the two systems are often closer than this.
You might be a good mathematician but I suspect you have a few things to learn about mt4.
1st: don't use strategy optimizer ... it leads to curve fitting.
2nd: don't test on open prices unless everything in the ea runs at open_prices.
3rd: 30 trades in almost 2 years ... imo you're gonna need some more samples.
Sure all the numbers are good. But allot of people here can duplicate those numbers especially with trend following systems and such few trades. The profit-factor is only high because of the win-size and the draw-down is only low because of the loss-size. However, for 10 months now the system have been losing. The system is clearly out of sync with the markets. You have to really ask yourself, will i still keep trading this system if I enter real-money now and it keeps losing like this for another 10-months.
Death by 10,000 cuts and Death by single blow have the same out-come IMO. I just prefer the single blow.
Ps: when you create a grid like system, one of the last things you wanna do is put it through the strategy optimizer. Generally, you come up with an idea, you code it and test it. It should be profitable on current year, then the next, then the next. Then profitable on the next currency pair, then the next and so on. When it fails, you throw it in the recycle bin.
Hello, here we go:
people claimed they could make money in a completely random market, which cannot be true
Completely random maybe not, but if you define some boundary's theoretically even a standard grid system combined with high bankroll should be able to beat random generated charts.
am I right in guessing that you have mildly optimised the parameters of the grid system?
I surely have tested different parameters during development, but i never used the optimiser on any parameter. This of course is also some kind of optimisation but i think that is the standard procedure during development. The biggest affect in overal performance is of course the globalTraget/startingLot ratio. The gridsize (as long as you don't make it too short) only affects the number of trades and consequently the profit. Drawdown stays nearly the same because with larger grids i am reversing later. (i let the optimizer run over night to see what comes out ;) )
One thought I had which has surely been anticipated is to wind down positions more rapidly than you increase them. So while one unit might be added for each step forward, two units might be removed for each step back (but not past zero). The idea is to keep some of the profit even if you don't hit a profit target. How does this relate to what you are doing?
Interesting, but if i am getting this right i think the result would be a grid like system which performs better on ranging/countertrend phases. I currently use the same scaling faktor for both directions, which of course is the reason for the heavy drawdown. The upside of course is that in breakouts, spikes and gernerally trending markets i reach the profit target quickly. The reason behind this idea is that i simply don't know where the markets move.
I do not believe that the system will be profitable in the long run at it's current stage, there are still a few problems which need's to be solved. (Dynamic gridsizes for avoiding ranges and maybe even using smaller grids when i got the right direction, use a equity trailing based exit instead of fixed one, portfolio style trading to avoid big drawdowns caused by one symbol) this are the ideas i will implement as next. (or try). But i currently believe that such systems can successful exploit the 'fact' that markets cause to trend. One of the problems with such testings is of course that you automatically use symbols that you know they are trending. If i add EURCHF in the portfolio i am quite sure the system will perform way better if you test the last 2-3 years.
Indeed, with a strong trendfollowing system you only need a small number of winning trades. Your data is unfortunately very limited. This can suggest that you are using a lot of filters, or trade only on very specific indicator states. For me personal 32 trades in two years are not enough to even estimate the possible expectation. (But statistics is your field).
The two systems can not really be compared, even if they both try to exploit trend's the technique is quite different. I assumed you would have more interest in statistical trading systems ;)
Here is my trendfollower which is currently running live. Unfortunately not that good in the current aggresive markets. Seems it is not such adaptive as i tough.
Best regards
Michael
When it fails, you throw it in the recycle bin.
I generally try to find the cause why it was failing, searching for a possible fix, and starting the whole testing process again. If the performance shape of the new system is very different, then i delete it. But if the filtration works as expected, the overall shape should be the same, profit faktor should be higher or equal, but the number of trades should not drastically different.
my point of view.
You might be a good mathematician but I suspect you have a few things to learn about mt4.
1st: don't use strategy optimizer ... it leads to curve fitting.
2nd: don't test on open prices unless everything in the ea runs at open_prices.
3rd: 30 trades in almost 2 years ... imo you're gonna need some more samples.
Sure all the numbers are good. But allot of people here can duplicate those numbers especially with trend following systems and such few trades. The profit-factor is only high because of the win-size and the draw-down is only low because of the loss-size. However, for 10 months now the system have been losing. The system is clearly out of sync with the markets. You have to really ask yourself, will i still keep trading this system if I enter real-money now and it keeps losing like this for another 10-months.
Death by 10,000 cuts and Death by single blow have the same out-come IMO. I just prefer the single blow.
Ps: when you create a grid like system, one of the last things you wanna do is put it through the strategy optimizer. Generally, you come up with an idea, you code it and test it. It should be profitable on current year, then the next, then the next. Then profitable on the next currency pair, then the next and so on. When it fails, you throw it in the recycle bin.
@ubzen, thanks for your sermon, but a little less haste might have avoided some mistakes.
1st: While you are free to not use strategy optimiser, your advice is not good to anyone who has enough knowledge to make it useful. Surely you are familiar with the importance of using out of sample data for evaluation? In particular, walk forward optimisation is a procedure that gives realistic rather than "curve-fit" performance, but also provides a quantitative measure of how well optimised parameters work in out of sample testing (walk forward efficiency ratio). With a methodology that has been shown to give good WFRE (actually, I think the isolated stats of the walkforward runs are more important than how they relate to those of the optimisation runs, but that's a personal view), walkforward optimisation provides a simple way of changing a static system with N free parameters into an adaptive one with zero parameters. This can certainly be of value if it is the case that the market has any characteristics that are persistent but which change with time. Unless the statistical characteristics of the market never change, I suppose it must have.
2nd: I understand the inaccuracies that using open prices can introduce into a test, but your advice was of no relevance to my post. The systems used no stops or targets, so the testing was as realistic as a test using full tick data. I will remark that the main error introduced by using open prices only is when it is possible for there to either be an entry and an exit in the same bar or two different possible entries or exits in the same bar. In many cases when the bars are small enough for this not to be the case, open price only simulation is accurate enough.
3rd: More samples would be nice, but I am constrained by reality! As well as the out of sample results I presented above, I did walkforward optimisation using a smaller testing period (1800 days) and a mere 180 days of testing. This produced an acceptable walkforward efficiency of 0.5 and 9 winning periods out of 11 in the out of sample testing of the second system despite the fact that there were only an average of 9 trades in each 6 month period (the 100 trades in the union of the out of sample periods is the sample size that matters). Interestingly, despite the much larger sample sizes, the walk forward performance of the single indicator system over 11 cycles (same 1800 day/180 day regime) was scarcely profitable, further discouraging use of this system.
I didn't claim the numbers were good. They are passable, but hardly ideal for trading real money. The potential performance of a system is determined by the statistical quality and the number of trades. These systems have ok statistics but have a very small frequency of trades, which reduces their value considerably. This can be made more precise mathematically, showing how a larger number of trades of lower quality may be statistically similar to a smaller number of trades of higher quality.
It is the ratio of the profit to the drawdown that I look at. This, not the drawdown itself, gives some hint of performance. In principle, statistics relating to the relationship between the mean return and the standard deviation of the return on the individual trades gives a more precise idea, but the simple drawdown stat captures some of the tendency of a system to produce sequences of losses. I have noticed that in many systems, losses have a tendency to cluster.
You made a statement about some system losing money for the last 10 months. This figure bears no relationship to my post and is not correct. Both systems have lost money during the last 3 months of sideways EURUSD, like most simple trend-following systems. Avoiding this would be nice, but that would make the systems something different.
"Death by 10,000 cuts and Death by single blow have the same out-come IMO. I just prefer the single blow." - Interesting philosophy, but can't see the connection between trading and preferred choice of method of death.
@Elroch: A sermon-eh? You're such a nice guy and I really like you but you're the one preaching to the choir. You're taking the thread off course when you throw a Trend following system's curve and try to suggest that it could be used as a comparison in some fashion. Because I wanna keep my responses short and sweet, I didn't elaborate because I taught the reasons were clear.
1st: your advice is not good to anyone who has enough knowledge to make it useful
I didn't just come to this conclusion overnight. These were advises passed to me by some of the best contributors who walked this forum. But being the person I am, and didn't just follow blind advise, I actually put in time experimenting for myself. And guess what came to the same conclusion. This statement: Perhaps slightly more frequent reoptimisation was necessary ... is where I had to smile and decided to respond to that post. When you re-optimize, its no longer the same system you based everything up to this point upon. You disagree? Ok, why didn't you re-optimize the system every 3-months during the testing? Just food for taught.
I strongly doubt that you'll be investing any money into that system the way its going now. But lets say you re-optimize and now the system appears to be back on track. Are you ready to invest now? There are numerous articles and techniques on this site About Being In Phase, or using system characteristics like Law of Runs and Z-Score to avoid those losing streaks which hunts Trend following systems. The problem is when you put the system in one Phase, the market my be ready to change Phase. Around here we call that a market change.
The above stuff I have some experience with. However when it comes to optimization I don't have experience with. Those would include Self-Optimizing EA's, either by using the Optimizer very frequently automatically or Neural Nets. But the biggest difference with this advance methods IMO is you provide the ingredients and the system tell you how to trade-it. Than opposed to the simpler methods which you provide the ingredients and also tell the system how to trade it, with little changes here and there. In either case, the ingredients you provide will determine what comes out. Garbage-in, garbage out.
Surely, you are familiar with the importance of using out of sample data for evaluation? In particular, walk forward optimisation
Check this out, Here. The walk-forward analysis link I ran into the first week I toke up forex. Also, take a look at my thread Here. It deals mostly with creating trend following systems according to typical convention as I was aware of at the time.
2nd: The systems used no stops or targets, so the testing was as realistic as a test using full tick data.
Even-If it uses a Algorithmic close, (like a ma-cross-over), this could happen several times within a 15-Minutes period. The only way this would not make a difference is if you coded the system to run the Exit Algorithm (which is tight--as far as i can deduce) every 15-Minutes_At Open Price. If in reality, your EA does all calculations every Tick, then hmmm think about it. It's 15-Minutes we're talking about here, I've seen price run over 200 pips up-and-down in 15_Minutes. While your EA in reality will respond to the movement in real time. Your back-test is sleeping, waiting for the next Open_Price. As to the point of when this would never be the case, is when your Stop_Loss logic would Never lie within a 15_Minute time movement and I find it very hard to believe for a Tight_Stop Trend following system.
3rd: More samples would be nice, but I am constrained by reality!
Humm, I don't know about that. Why not take your own advise. Run your system on other currency-pairs and return here and post the worse result. If it's not Death by a thousand Cuts then I'll quit forex today ;). But maybe you'll say, umm.. thats not fair, I optimized it for EURUSD. But what in the world said EURUSD prices could not suddenly start behaving like some other currency pair? So I guess with your limitation, you'll have to forward test this longer. By the time you get enough data, I'll be an old man. And hopefully there'll be no Market-Change by then (strongly doubt it). Another reason I don't know why you drop the curve in this thread.
I didn't claim the numbers were good. In my opinion, those numbers are extraordinary if and only if I could convince myself that I'll see similar results in the future. I mean 25% return in 20 months. What other investment is gonna give me that?
You made a statement about some system losing money for the last 10 months. So I guess it made 1/3 the trades in the first 17 months and 2/3 the last 3 months. My apologies.
"Death by 10,000 cuts and Death by single blow have the same out-come IMO. I just prefer the single blow." - Interesting philosophy, but can't see the connection between trading and preferred choice of method of death
Again do me a favor. Run the system on other pairs and show the worse result. Then you'll understand what I mean by death by 10000 cuts. Lol.