you might have missed this thread a while back http://elitetrader.com/vb/showthread.php?s=&threadid=113044
You are using an out of date browser. It may not display this or other websites correctly.
You should upgrade or use an alternative browser.
You should upgrade or use an alternative browser.
My SPM + SCT child
- Thread starter frostengine
- Start date
<i>"When trying to code a strategy, it's amazing how difficult it is, but I find it makes one a aware of flaws in discretionary trading."</i>
... but more importantly, it makes you aware of the flaws in system trading which are only overcome in discretionary trade decisions.
The reason for poor performance on the short side of this system is simple. Trade entry language is written to execute upon the open of next bar after MACD (or whatever) crossover. Now I could be wrong, but doubt it in this case.
Upside price action has a strong tendency to print relatively small bars in methodical fashion. Downside price action has a tendency to print bars in surge = blowaway fashion. Blindly telling a bot to enter a trade upon open of next bar after signal confirmation is too often entering ahead of a pullback. The stop doesn't hold, sell side probably goes on to work in many-most cases but the trade failed because entry technique was flawed.
The short side results suffer not because entry signals are any different... it suffers because price action is decidedly different on shorts versus longs overall. Gotta know that fundamental difference AND be able to compensate for it in order to win.
*
A discretionary trader would view the crossover occurance as a confirmation to seek shorts. Then that human mind (of a learned, skilled trader) would measure price action, access the situation and enter at a much more strategic spot that blind-selling the open of a bar at random after two lines crossed on a chart.
Bot writing is far more difficult and laborious than learning to visually measure = read price action itself.
... but more importantly, it makes you aware of the flaws in system trading which are only overcome in discretionary trade decisions.
The reason for poor performance on the short side of this system is simple. Trade entry language is written to execute upon the open of next bar after MACD (or whatever) crossover. Now I could be wrong, but doubt it in this case.
Upside price action has a strong tendency to print relatively small bars in methodical fashion. Downside price action has a tendency to print bars in surge = blowaway fashion. Blindly telling a bot to enter a trade upon open of next bar after signal confirmation is too often entering ahead of a pullback. The stop doesn't hold, sell side probably goes on to work in many-most cases but the trade failed because entry technique was flawed.
The short side results suffer not because entry signals are any different... it suffers because price action is decidedly different on shorts versus longs overall. Gotta know that fundamental difference AND be able to compensate for it in order to win.
*
A discretionary trader would view the crossover occurance as a confirmation to seek shorts. Then that human mind (of a learned, skilled trader) would measure price action, access the situation and enter at a much more strategic spot that blind-selling the open of a bar at random after two lines crossed on a chart.
Bot writing is far more difficult and laborious than learning to visually measure = read price action itself.
My system, past two years, has 52% winners and profit factor of 4.5 across all stocks in the Australian All Ords.
The system design is not complete yet as I have a large mod to do which should improve this result even more.
The system is coded not by myself (I desgn it) but by a professional coder who codes trading systems, so in other words it's not a backyard coding job.
The system design is not complete yet as I have a large mod to do which should improve this result even more.
The system is coded not by myself (I desgn it) but by a professional coder who codes trading systems, so in other words it's not a backyard coding job.
Quote from monti1a:
You're just CURVE-FITTING, because you keep massaging the results until it gives you the desired outcome. I'll bet money your 'system' falls apart in real-time trading.
Although I may be curve fitting the system, I am not entirely certain. Generally systems I have curve fitted would fail miserably when I switched to a different instrument which is not the case here.
BTW it had initially failed on the forex data I fed it, but after careful inspection of the data it was horrible data. I grabbed some more reliable data and the system is profitable across that asset class as well.
Another test I did to check how curve fitted the system may be was to massage the start and stop times of the system. Results held up well here as well..
So granted, I can't be sure that its not curve fitted, but I have taken many precautions. This is not my first trip into the world of automated trading
Quote from austinp:
<i>"When trying to code a strategy, it's amazing how difficult it is, but I find it makes one a aware of flaws in discretionary trading."</i>
... but more importantly, it makes you aware of the flaws in system trading which are only overcome in discretionary trade decisions.
The reason for poor performance on the short side of this system is simple. Trade entry language is written to execute upon the open of next bar after MACD (or whatever) crossover. Now I could be wrong, but doubt it in this case.
Upside price action has a strong tendency to print relatively small bars in methodical fashion. Downside price action has a tendency to print bars in surge = blowaway fashion. Blindly telling a bot to enter a trade upon open of next bar after signal confirmation is too often entering ahead of a pullback. The stop doesn't hold, sell side probably goes on to work in many-most cases but the trade failed because entry technique was flawed.
The short side results suffer not because entry signals are any different... it suffers because price action is decidedly different on shorts versus longs overall. Gotta know that fundamental difference AND be able to compensate for it in order to win.
*
A discretionary trader would view the crossover occurance as a confirmation to seek shorts. Then that human mind (of a learned, skilled trader) would measure price action, access the situation and enter at a much more strategic spot that blind-selling the open of a bar at random after two lines crossed on a chart.
Bot writing is far more difficult and laborious than learning to visually measure = read price action itself.
Austin, neither the long nor the short side is blindly taking trades after the cross of the MACD.
I may be forced to remove a component of the strategy. I already had to remove this piece for testing on forex. Its the volume confirmation. Removal of this hurts the results some, still returns around $33,000 per contract for the long only without this confirmation... But affects PF and Sharp...
The reason I am having to remove it is I am trading through IB and planned to use ninjatrader to auto trade this. Ninjatrader uses the tick data from ib to build its minutely bars instead of the 5 second snapshot. Therefore volume totals will be off.
The reason I am having to remove it is I am trading through IB and planned to use ninjatrader to auto trade this. Ninjatrader uses the tick data from ib to build its minutely bars instead of the 5 second snapshot. Therefore volume totals will be off.
For those wondering, the results for the forex. I have only tested EUR/USD. And the only data I have is from 4/1/2001 through 12/1/2002
And then another chunk from 1/1/2007-1/1/2008
Granted my data is very sparse. But I am hoping to get more... the thing is i'm using Gaincapitals forex data as that is what i'll be using if I trade the forex side with this strategy. And I have to download week by week the data and then run it through my app to convert it to use in ninjatrader, so getting forex data is going slow.
However testing on that 2.5 years worth of data. I get a profit of $12,740 per standard $100,000 contract. Using 2 ticks of slippage.
PF: 1.55
Sharp: .21
Trades: 325
And then another chunk from 1/1/2007-1/1/2008
Granted my data is very sparse. But I am hoping to get more... the thing is i'm using Gaincapitals forex data as that is what i'll be using if I trade the forex side with this strategy. And I have to download week by week the data and then run it through my app to convert it to use in ninjatrader, so getting forex data is going slow.
However testing on that 2.5 years worth of data. I get a profit of $12,740 per standard $100,000 contract. Using 2 ticks of slippage.
PF: 1.55
Sharp: .21
Trades: 325
Quote from frostengine:
Although I may be curve fitting the system, I am not entirely certain. Generally systems I have curve fitted would fail miserably when I switched to a different instrument which is not the case here.
BTW it had initially failed on the forex data I fed it, but after careful inspection of the data it was horrible data. I grabbed some more reliable data and the system is profitable across that asset class as well.
Another test I did to check how curve fitted the system may be was to massage the start and stop times of the system. Results held up well here as well..
So granted, I can't be sure that its not curve fitted, but I have taken many precautions. This is not my first trip into the world of automated trading
I'm not a fan of highly optimized systems. I prefer semi-discretionary methods myself. However, I'll throw this out there.
'Lock your rules down", and then run your rules on a randomly chosen 'out-of-sample' test period of an adequate length (look for at least 1,000 trades). If the system still holds up, it only tells you that the system worked well in the past, and merits further interest. You can not infer the future from the past. Hell, I'll even say that trying to totally rely on all these fancy statistics are futile when used on a non-linear dynamic system such as the market.
However, you can build a little confidence that the system MAY hold up in the NEAR future, by performing a forward test - WITHOUT CHANGING YOUR RULES - for 2-3 months.
If comfortable, monitor the 4-6 sigma drawdown to determine whether the system is currently working real-time. Once, you've made more than the 4-6 sigma drawdown, the system has paid for itself, and who then cares if it blows up. But like Bill Williams said, "All mechanical systems ultimately die".