Systematic trading | Page 69 | 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

  • #1,361
  • May 5, 2011 8:32am May 5, 2011 8:32am
Quoting foracy
The key problem I see in your approach that it isn't about being able to predict the future price. What is more important is the win/loss distribution and that is not covered in what you have shown. If you have large loss clusters it will kill you no matter how good your price prediction may be.
Yes, I agree. I will need to improve the robustness of the prediction, I will need really smart and defensive position sizing method in order to survive losing streaks. We can also assume that prediction will suck at the time when important news are being released - which can be filtered, because many of news release times are in the calendar. Anyway as I said, I am on the start with this, for now it is just an experimental framework to play with....
  • #1,362
  • May 5, 2011 8:47am May 5, 2011 8:47am
Quoting sals3r0
I can but it won't be very obvious what it does mean, as I don't really play with trading strategy now, I am just focusing on evaluating a prediction performance of the system.

I usually have 6 training sets and 6 validation sets. I train the SOM+SVM on those training sets and evaluate performance on validation sets.
So...
Unless you already are using it, I would recommend adding an additional layer on top of your predictor, one that measures at what kind of market state the result is accurate. Market has some states that can be modeled and NN tends to work in one state but not in other. Volatility clusters as they say.

By the way, I personally don't think you can predict the next X bars accurately unless you have some extra rules as volatility tends to cluster at levels and before the bar (especially as high as 4 hour) you can't know if it reaches some level and if it does, there might be results because the price touched that level, not because some previous bars moved as they did.
  • #1,363
  • May 5, 2011 3:22pm May 5, 2011 3:22pm
Quoting mikkom
Unless you already are using it, I would recommend adding an additional layer on top of your predictor, one that measures at what kind of market state the result is accurate. Market has some states that can be modeled and NN tends to work in one state but not in other. Volatility clusters as they say.

By the way, I personally don't think you can predict the next X bars accurately unless you have some extra rules as volatility tends to cluster at levels and before the bar (especially as high as 4 hour) you can't know if it reaches some level and...
I don't do that already but it is something I have already on my mind, because simple pattern recognition is not enough here. The question is how to do it - I am thinking about using the SOM for this purpose. I already use it to pre-cluster training cases, so maybe there could be some smart way how to use SOM for clustering "market states" too by feeding some interesting inputs to it - like hour of the day, absolute price, current volatility, etc....but this all need experiments (SOM is quite new technology for me anyway)...maybe I will let the GA to do the dirty work of selecting proper inputs to SOM, but then it will not be easy to define the cost function for the GA

I agree with you, the price 5 bars to future cannot be predicted with high accuracy, you are also right about different market states and different predictability - I hope a bit that with smart clustering method and with calendar news release filtering I could filter as much as possible those "unpredictable states" out in order to try to minimize losing streaks in the strategy.

At least I got nice framework to play with for now Thanks for good ideas mikkom
  • #1,364
  • Edited 5:52pm May 7, 2011 3:39pm | Edited 5:52pm
I have started to play this baby
This graph was produced using tick data over a 3 year period, 492 trades. This is one strategy on EUR/USD and trades are intraday. I am amazed how difficult it is to find systems like this and now I am hoping it will continue like this for another 3 years. (-:
Anyone else like to share any testing results.
Attached Image (click to enlarge)
Click to Enlarge Name: strategyX.png Size: 23 KB
  • #1,365
  • May 8, 2011 9:00am May 8, 2011 9:00am
foracy, what's the position size management in the simulation?
  • #1,366
  • Edited 4:04pm May 8, 2011 3:36pm | Edited 4:04pm
Quoting usalchemist
foracy, what's the position size management in the simulation?
Hi the moneymanagement is very conservative (no martingale,just increasing slowly). What makes this result so good is the loss/win distribution (max conc loss =4).

I would love to know if other people have similar results. I have made billions of calculations and the combination of about 500 trades in three years(intraday) and holding up with such a low drawdown is so unique.(in my case anyway) Even then it takes a long time to make some serious money and that is only if it lasts.
  • #1,367
  • May 9, 2011 1:01am May 9, 2011 1:01am
Quoting foracy
I have started to play this baby
This graph was produced using tick data over a 3 year period, 492 trades. This is one strategy on EUR/USD and trades are intraday. I am amazed how difficult it is to find systems like this and now I am hoping it will continue like this for another 3 years. (-:
Anyone else like to share any testing results.
Is this in- or out-sample?
  • #1,368
  • May 9, 2011 2:42am May 9, 2011 2:42am
foracy, it's not possible to compare training result of mine against yours since you haven't specified your position size management strategy.
  • #1,369
  • Edited 6:00am May 9, 2011 5:45am | Edited 6:00am
Mikkom,

These results are in-sample but I have been running this system live since begin february (with a small amount) to measure slippage. The results have been within the set parameters but as you can see from the graph it takes a long time before it becomes (really) profitable.

I do also believe that if you can produce high enough frequency within a time range and some of the figures really stand out, that out of sample testing becomes less of an importance. The idea behind this that at one point even the most powerful computer can't curve fit it any more and there for it must be a market behaviour. (also I am not doing GA)
Based on my testing this strategy stands out in terms of frequency (500 in 3 years), drawdown and profitability. This is why I would like to know if anyone else has had similar or better results than this (intraday and tick data).
  • #1,370
  • May 9, 2011 4:30pm May 9, 2011 4:30pm
Hi foracy,Mikkom and everybody,

I know eveyone has their own idea about how to measure the kind of market state. Can we put these as inputs for SOM to measure the market state ) zigzag with pip value b) slope of trendline or even slope of channel c)Trade that particular time frame if Hurst exponent is > 0.6 (I-VAR indicator as an alternative that is popularized in another topic?) d)variance or ATR

My other question is lets say I want to build a system based around MACD.I want to find out the best signal range(O-H-L-C bar and slope) of MACD ,how do I find that out ,with Monte carlo simulation?Genetic programming?Thanks.
  • #1,371
  • Edited 7:07pm May 9, 2011 6:49pm | Edited 7:07pm
Purley, do you have any experience with programming?

Why can't you try to generate some kind of hypothesis about your expected behavior and then run some tests to verify it.

For instance, what exactly do you expect from a series with a hurst greater than 0.6 ( I would assume you would expect it to be a low noise, trending, and persistent environment)? More importantly, you want to understand if the characteristic itself is persistent enough to capture something exploitable. For example, does the hurst value have some kind of serial correlation associated with it? If not, then what use might you expect from it?

The way I tend to approach a lot of GA is to utilize it as a tool to 'augment' my ideas (or possibly discover new relationships, which is another application). If you were searching for the best hurst behavior to exploit whatever it is you are seeking, GA might be able to help you enumerate over unusually large search spaces. Rather than utilize the 'blender' approach, It can be a great help here.

You have thrown out several viable ideas, but should go back and try to test them and find out why or why not they might work. I think it's very rare that someone who has done extensive research in each of these broad topics is going to post detailed results, but if you share it with the board, you might just get contacted by others who wish to continue the research, or be more forthcoming with personal research in these areas.

One other point, is you need to be very explicit about what you define as the 'market state.' It might sound trivial, but it's very important to identify first.
  • #1,372
  • May 10, 2011 12:42am May 10, 2011 12:42am
Quoting foracy
I do also believe that if you can produce high enough frequency within a time range and some of the figures really stand out, that out of sample testing becomes less of an importance. The idea behind this that at one point even the most powerful computer can't curve fit it any more and there for it must be a market behaviour. (also I am not doing GA)
Based on my testing this strategy stands out in terms of frequency (500 in 3 years), drawdown and profitability. This is why I would like to know if anyone else has had similar or better results than...
The point of out-sample is that you don't have to believe anything, you can see how your system would have fared in the future. If you are finding systems yourself it's unfortunately very hard to use out-sample correctly.

About curve fitting, there are different kinds of curve fits, one is the right type of fitting (that's what we traders basically do) one and other one is incorrect (ie. unprofitable on anything but tested period). The main point of out-sample is to make sure you are finding strategies that work on future too.
  • #1,373
  • May 10, 2011 1:22am May 10, 2011 1:22am
Quoting foracy
This is why I would like to know if anyone else has had similar or better results than this (intraday and tick data).
Yes, I have trained similar or better result based on setting every trade S/L risk at 1-2% of equity. That's why I asked you about your position size management strategy in order for us to compare apple to apple.
  • #1,374
  • May 10, 2011 1:42am May 10, 2011 1:42am
Quoting usalchemist
Yes, I have trained similar or better result based on setting every trade S/L risk at 1-2% of equity. That's why I asked you about your position size management strategy in order for us to compare apple to apple.
I send you a private message with exact details. Did you use tick data taking the spread into account?
  • #1,375
  • Edited 4:07am May 10, 2011 3:59am | Edited 4:07am
Replied. Pls check.
No, we don't use tick data for we don't do HFT. We assume big spread (5 pips for EURUSD).
I was a huge fan of tick data and I convinced the programmers to convert all Dukascopy tick data into backtestable MT4 data when we started the project. Long story short, it's unnecessarily too big a data set for our intended strategy. We decided to save it for later HFT effort.
What's your experience with the tick data set? Did it make a huge impact on your backtesting effort? How was that so?
  • #1,376
  • May 12, 2011 2:13am May 12, 2011 2:13am
I've come to realize my misunderstanding and confusion about the usage of out-of-sample data and I've philosophized a satisfactory understanding of its usage: backtesting assumes that history repeats itself under the same set of circumstances. If you think this statement is true, the thrust of the truism must lie in the definition of circumstances. If a set of circumstances is described correlatively, future happening of the same set of circumstances must yield similar results.
Cambridge American English Dictionary defines circumstance as an event or condition connected with what is happening or has happened.
What's sad about the definition of circumstances is that time is the most important condition in system trading that's unrepeatable. There can't be a repeated singular time in history unless we're in Bill Murray's Groundhog Day. Therefore, when we assume history repeats itself, we have no choice but to look for other condition. In essence, we're saying that we hope certain circumstances other than time that repeat in the future can provide insight into the forthcoming price action.
When we try to define a set of circumstances, numerical variables invariably come into play. When we're talking about optimization, we're talking about the optimizing of the value of these variables. They're not strategy. They're just part of the strategy.
When an out-of-sample test fails, the strategy fails. It's wrong to think that the optimized value fails. It's important to know the difference between a strategy and its numerical variables. I used to doubt the usefulness of out-of-sample test. I incorrectly assumed that other out-of-sample test practitioners were going back to re-optimized the value when post-optimization result went bad. In fact, they don't.
They abandon the trading strategy if out-of-sample test fails because the strategy is likely to have contained an uncorrelated set of circumstances that produced the historical price action.

http://books.google.com/ebooks?id=FX...C&pg=GBS.PA252
Attached Image (click to enlarge)
Click to Enlarge Name: 5-12-2011 2-44-31 PM.png Size: 80 KB
  • #1,377
  • Edited 4:15am May 13, 2011 3:36am | Edited 4:15am
Quoting usalchemist
They abandon the trading strategy if out-of-sample test fails because the strategy is likely to have contained an uncorrelated set of circumstances that produced the historical price action.
If you do this, your out-sample becomes in-sample. By rejecting something based on out-sample you are using it ("the future") as part of your fitness calculation. If you do that, you just have two in-sample periods with different fitness algos. One that is used for combination and optimization and other that is used for validation.

Out-sample should be used for verification of your overall strategy (fitness, combination, genration etc), you CAN NOT use it as part of your system.
  • #1,378
  • Edited 1:27pm May 13, 2011 6:40am | Edited 1:27pm
Quoting mikkom
If you do this, your out-sample becomes in-sample. By rejecting something based on out-sample you are using it ("the future") as part of your fitness calculation. If you do that, you just have two in-sample periods with different fitness algos. One that is used for combination and optimization and other that is used for validation.

Out-sample should be used for verification of your overall strategy (fitness, combination, genration etc), you CAN NOT use it as part of your system.
At one point you check if it works outsample (in your case the final step). I guess what you mean is that you do not use the out of sample period to construct the fitness algo, right?
  • #1,379
  • May 13, 2011 9:50am May 13, 2011 9:50am
Quoting foracy
At one point you check if it works outsample (in your case the final step). I guess what you mean is that you do not use the out of sample period to construct the fitness algo, right?
Yes, at the final step I reset and repeat and see if the outsample is similar on all runs. I might post some pics later. I only do this to see if the outsample (the future) is similar to in-sample (the past/training period). It never is as good but with some tuning it might get quite similar. If you can repeat the result 5-10 times and out-sample is always similar to in-sample then you will know that you have quite efficient fitness algorithm.

(Usually that is not the case - out-sample is almost always much worse and that tells about your parameters, there is too much correlation, curve fitting etc. Good fitness algo and some additional filters will help to get rid of these and then one bad algo here and there won't matter)
  • #1,380
  • Edited 10:48am May 14, 2011 10:31am | Edited 10:48am
Here are two examples, I can repeat both my changing some aspects of fitness algorithm and running the program for few hours (the strategies that are combined to form these curves are the same on both examples). The curves will not be totally the same but out-sample is similar. In both of these, the out-sample is not used in any way except for verification (verification visually as is shown below - you know instantly which fitness algo is better).

I'm not going to tell what I changed between these but the changes were just parametrical, ie. I allowed some values to exceed some levels on other and disallowed some aspects on the better one.

Please don't ask me further about this, I will not go further than this to these exact curves/rules but feel free to discuss about this if you want to (one note is that I don't weight results by years, in fact all points in time are used without weightings, first year is as important as last).

Again, this is why out-sample is important, I can now see how my fitness works and how it would have worked for the next 1.5 years. Without real out-sample I wouldn't have a clue.

I hope this clarifies at least a bit.

Also, this is worth watching - nothing to do with trading however
Attached Image(s) (click to enlarge)
Click to Enlarge Name: good.png Size: 50 KB
Click to Enlarge Name: bad.png Size: 46 KB
2 traders viewing now