Systematic trading | Page 66 | 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

  • #1,301
  • Feb 17, 2011 1:40am Feb 17, 2011 1:40am
Quoting wguan
I have been testing out strategies in my backtesting program and the results has been showing some drawdowns which I am uncomfortable with.

From anyones experience, is there some ways to decrease DD while increase the annualized return (CAGR%) and also boosting the MAR ratio? Is there a process or approach since I don't want to curve fit my system. Any inputs would be great.
All that traders do is curve fit our systems / behaviour to past data.

The actual question is: How to find those events that repeat themselves so you curve fit to the correct parts of the curve.

One answer is to use so much data that the system that actually works always works better than curve fitted system. How to do that is much harder to find than it sounds.

About the dd question, the best way to decrease DD while increasing profit is to use multiple negatively/ low correlated systems. Also remember that DD changes directly when you adjust your trade risk so you might want to compare how different weightings (assuming you have multiple systems) affect your curve.
  • #1,302
  • Feb 17, 2011 1:00pm Feb 17, 2011 1:00pm
M:


I also came along this thing called walk forward analysis in Pardo's book. From what my understanding, you separate your data into 2 segments. Test and optimize the initial segment while use the second segment to do the final testing. I am not sure if that makes sense but it seems correct...

My TF system is a longer term system and I am in the process of trying to quantify and code my second momentum system which I found to compliment the TF very well. I have to agree with you on the matter of uncorrelated strategies that will help to reduce DD. I think incorporating uncorrelated assets will also help reduce DD and volatility. (cuz you are not double betting on any single market)

From what i have been reading, certain exit strategies also help to reduce DD. (ie, volatility TS)

What kind of system do you trade?
  • #1,303
  • Feb 17, 2011 1:52pm Feb 17, 2011 1:52pm
I you're interested in examples of what high frequency traders (including yours truly) might be looking at when designing trading strategies, yous should take a look at tr8drs last couple of posts:

http://tr8dr.wordpress.com/
  • #1,304
  • Feb 17, 2011 4:02pm Feb 17, 2011 4:02pm
That blog is indeed fantastic, the guys grasp of many different fields is inspiring.
The breaking of a wave cannot explain the whole sea.
  • #1,305
  • Feb 19, 2011 10:51pm Feb 19, 2011 10:51pm
I don't think my systems are high frequency. They operate on the daily timeframe.
  • #1,306
  • Feb 21, 2011 8:35am Feb 21, 2011 8:35am
Quoting wguan
From what my understanding, you separate your data into 2 segments. Test and optimize the initial segment while use the second segment to do the final testing. I am not sure if that makes sense but it seems correct...
Of course you need out-sample verification. All in-sample data results should be considered as false (IMHO).
  • #1,307
  • Edited 5:18pm Feb 21, 2011 3:01pm | Edited 5:18pm
Quoting wguan
M:
I also came along this thing called walk forward analysis in Pardo's book. From what my understanding, you separate your data into 2 segments. Test and optimize the initial segment while use the second segment to do the final testing. I am not sure if that makes sense but it seems correct...
Yes, it is wise to test on out-of-sample data, or cross-validate, as some disciplines call it.
However, allow me to warn you to be careful with that book. It starts out very promising, but at some point you realize he really does not know what he's talking about.
That is, he supposedly presents a comprehensive framework to develop, evaluate, and optimize trading strategies. But he does not know the first thing about optimization.
Beware.
  • #1,308
  • Feb 21, 2011 11:41pm Feb 21, 2011 11:41pm
Also wanted to know more about position sizing algorithms.

There seems to be only the following category of position sizing algorithms.
1. Fixed-amount model
2. Equal leverage model
3. Percent risk model
4. Percent Volatility model

Is there more? And the best seems to be 4% one. I am getting a sense that there isn't much ideas around position sizing. (im prolly worng)
  • #1,309
  • Feb 21, 2011 11:41pm Feb 21, 2011 11:41pm
Quoting 6pack
Yes, it is wise to test on out-of-sample data, or cross-validate, as some disciplines call it.
However, allow me to warn you to be careful with that book. It starts out very promising, but at some point you realize he really does not know what he's talking about.
That is, he supposedly presents a comprehensive framework to develop, evaluate, and optimize trading strategies. But he does not know the first thing about optimization.
Beware.
is there any books you recommend?
  • #1,310
  • Feb 22, 2011 3:15am Feb 22, 2011 3:15am
Quoting wguan
Also wanted to know more about position sizing algorithms.

There seems to be only the following category of position sizing algorithms.
1. Fixed-amount model
2. Equal leverage model
3. Percent risk model
4. Percent Volatility model

Is there more? And the best seems to be 4% one. I am getting a sense that there isn't much ideas around position sizing. (im prolly worng)
The question is what you are looking for in a position sizing algorithm - if you trade multiple systems/positions you might want to limit risk to certain level for example or you might want to limit directional/correlated risk to some predefined level.

Also 4% is quite high risk level if you ask me.
  • #1,311
  • Feb 22, 2011 12:12pm Feb 22, 2011 12:12pm
oh oops. i typed it wrong. I meant the number 4 volatility based algorithm is the most perferred one. I only risk 2% max on any trade. I have two systems but then I am not sure if i should risk the same amount for each one. They complement each other well so risking less on each one may got.

My TF system also pyramids, i was also thinking of adjusting the size of each additional units up. Also move stops at certain level would help me reduce risk.

My current problem is coding a proper trailing stop into my backtesting software...

anyways thanks so much for the input!

Btw. Mikkom: what do you trade? do you trend follow?
  • #1,312
  • Feb 23, 2011 2:59am Feb 23, 2011 2:59am
Quoting wguan
Btw. Mikkom: what do you trade? do you trend follow?
I trade anything that works
  • #1,313
  • Feb 26, 2011 3:07pm Feb 26, 2011 3:07pm
anyone got ideas about other position sizing algorithms?

-kelly criterion
-volatility based

these seems to be the only one i know.

in terms of kelly, the equation is:
Kelly = ((WP * W/L) - (1 - WP)) / W/L

where:
WP = Winning percentage
W/L = Avg Win/Avg Loss

isn't this backward looking? should i calculate WP AND W/L every time I initiate a position or shall i use a predefined constant based on a backtested period?
  • #1,314
  • Feb 27, 2011 4:26am Feb 27, 2011 4:26am
IMHO kelly is far too aggressive (you can always use percentage of kelly if you want to but you will realize that people, including you, tend to want to optimize to something else than aggressive profits) and the usual percentage risk actually is volatility based as the stop is (usually) moved further away as the volatility rises.
  • #1,315
  • Edited Mar 3, 2011 1:18am Mar 2, 2011 5:38pm | Edited Mar 3, 2011 1:18am
I have been testing a new fitness algo for a few days and I have to say I'm a bit astounded.. I thought there is a bug or something but after trying different methods (mar etc) and comparing results, this seems to be the real deal.. (other fitness methods give ok results out-sample but nothing this consistent)

The best thing? This one doesn't curve fit. Almost every other method peaks at some point and after that in-sample fitness continues going up while out-sample goes down.

Anyway.. here is one curve (about day or something of calculating srtrategy pool and about 30 mins calculating combinations with one core and 7 cores calculating more stuff) purple line divides in-sample (left) and out-sample (right).

Feel free to ask questions if you want to, there are wome things I might reveal (?). Maybe the first one should be: You should know what kind of results you are aiming at.
Attached Image (click to enlarge)
Click to Enlarge Name: Screenshot-Cf visualizer-4.png Size: 44 KB
  • #1,316
  • Mar 2, 2011 6:14pm Mar 2, 2011 6:14pm
Great system(s)(?) you got there. I thought i had a similiar system last week but it turned out to be a postdictive error in my code....: (

Anyways, I still consider myself a newbie so sorry about the noob questions haha.

What is a fitness algo? and you mentioned about how the MAR has a feature that almost ruins it...what do you mean? I mean MAR = COGR / Max DD.

From the win precentage I am guessing it is a TF system? Correct me if I am wrong...or is it a combination of multiple systems? You probably have a very good MM given that TF systems rarely perform that well.

Anyway, great work!

MG
  • #1,317
  • Mar 2, 2011 6:18pm Mar 2, 2011 6:18pm
Quoting wguan
From the win precentage I am guessing it is a TF system? Correct me if I am wrong...or is it a combination of multiple systems? You probably have a very good MM given that TF systems rarely perform that well.
TF system? I'm not sure what you mean. Yes, it's a combination of 21 systems (of about ~95000 candidates (which also are all profitable and have certain features why they are included at the pool)).

Fitness algo is the core of evolutionary optimization, I suggest reading the whole thread.. Or reading some related literature, some info also here http://en.wikipedia.org/wiki/Fitness_function
  • #1,318
  • Mar 3, 2011 4:18am Mar 3, 2011 4:18am
Quoting mikkom
I have been testing a new fitness algo for a few days and I have to say I'm a bit astounded..
Feel free to ask questions if you want to, there are wome things I might reveal (?). Maybe the first one should be: You should know what kind of results you are aiming at.
Hi Mikkom,

Congratulations on generating such nice out-sample results! I would be very interested to hear you expand on your fitness algo.

I've been considering a few options here; profit / drawdown is the simplest fitness criteria that I've found to be reasonably effective. However, it can easily be skewed by excellent performance during only a small part of the in-sample data.

Ultimately, what we want is an equity curve that continues in an upward direction throughout the entire in-sample period (which is what you have appeared to have achieved)

One option is to use a profit/drawdown parameter that remains consistently positive throughout the entire in-sample period. I haven't quite figured out how to turn this into a useful fitness parameter as of yet.

Another option is to calculate the slope of a linear regression line (a line of best fit through the equity curve), and divide by the standard deviation of the square of the error (i.e. the distance an equity point is away from the line of best fit). Have you tried this option?

Cheers,

Michael
  • #1,319
  • Mar 3, 2011 5:17am Mar 3, 2011 5:17am
Quoting PipMasterMik
Another option is to calculate the slope of a linear regression line (a line of best fit through the equity curve), and divide by the standard deviation of the square of the error (i.e. the distance an equity point is away from the line of best fit). Have you tried this option?
I have. The problem with that approach is that exponential looking curves give you the best results (assuming you only use downside deviation - I don't personally like using non-directional deviation, there is no point in penalizing good behaviour).

Just a small note on my curve, it is in log scale even though numbers (dd etc) are real exponential curve numbers. Visualizing with exponential curve doesn't make much sense.
  • #1,320
  • Mar 4, 2011 4:35am Mar 4, 2011 4:35am
Quoting mikkom
I have. The problem with that approach is that exponential looking curves give you the best results (assuming you only use downside deviation - I don't personally like using non-directional deviation, there is no point in penalizing good behaviour).
That sounds sensible.

I didn't realize that you were displaying a log scale on your equity curve; that makes these results even more impressive.

I guess another way of looking at it is by considering the equation:

Fitness = Good_Stuff / Bad_Stuff

Good_Stuff can be quite nicely defined as Profit (or P&L), while bad stuff is a little harder to define to ensure a constantly rising equity curve throughout the entire in-sample period (hence draw down on its own is not a good indicator).

Another idea is to slice the in-sample period up into smaller periods, and if the Profit/DD is not acceptable during any of these periods than you throw the whole optimization result away as being unsatisfactory. Alternatively, you could scale the total fitness result based on the % of periods that meet the minimum Profit/DD criteria. I might give these ideas are whirl at some point.

Does anyone have any suggestions for better definitions of Bad_Stuff or am I on the wrong track here?

Cheers,

Michael
0 traders viewing now