Hi,
I once attended a Wealth Lab training taught by Thomas Vittner. There we learned that you should never optimize the entry and the exit together.
However, recently I heard some opinions which stated it can make sense for a better overall performance to optimize them at the same time.
So, I would like to ask you for your opinions on that matter.
Thank you in advance
Werner
I once attended a Wealth Lab training taught by Thomas Vittner. There we learned that you should never optimize the entry and the exit together.
However, recently I heard some opinions which stated it can make sense for a better overall performance to optimize them at the same time.
So, I would like to ask you for your opinions on that matter.
Thank you in advance
Werner
Rename
This is not a multiple choice question. This is a graduate school essay question, and the professor will give full credit for either answer depending on how you argue.
Generally speaking, for solving a linear system of simultaneous equations, you want to solve them together. Each equation imposes a requirement (or bias) that you need to meet. But with stock trading, the picture is more murky.
1) How big is your N? Do you truly have enough trades to definitively solve all parameters of your system? You should have a minimum of 5 degrees of freedom (5 d.f) for random error. (Why is that? What did your statistics professor say about estimating the standard deviation of a sample?) So if your model has 6 d.f. (i.e. 6 parameters), then you need at least 5+6=11 trades to make this minimum. Anything less than that is junk. (I hope you are not betting money on junk.) Aside: A good rule of thumb is to design your production model with as few parameters as possible so you can meet the d.f requirement (minimum N) for random error.
If you have a really big N (many trades), then optimizing the entry points and exit points separately maybe safe; otherwise, it maybe a bad idea.
2) Does the entry points and exit points depend on the same criteria? Are they using the same indicators?
If not, then it's probably best to optimize the entry and exit points separately because they belong to a different linear system. That is, you have a mixed model. And that maybe a good thing in some cases.
In contrast, if you're using the same indicators to determine entry and exit points, then I would solve their parameters together (simultaneously) because they all belong to the same linear system.
Generally speaking, for solving a linear system of simultaneous equations, you want to solve them together. Each equation imposes a requirement (or bias) that you need to meet. But with stock trading, the picture is more murky.
1) How big is your N? Do you truly have enough trades to definitively solve all parameters of your system? You should have a minimum of 5 degrees of freedom (5 d.f) for random error. (Why is that? What did your statistics professor say about estimating the standard deviation of a sample?) So if your model has 6 d.f. (i.e. 6 parameters), then you need at least 5+6=11 trades to make this minimum. Anything less than that is junk. (I hope you are not betting money on junk.) Aside: A good rule of thumb is to design your production model with as few parameters as possible so you can meet the d.f requirement (minimum N) for random error.
If you have a really big N (many trades), then optimizing the entry points and exit points separately maybe safe; otherwise, it maybe a bad idea.
2) Does the entry points and exit points depend on the same criteria? Are they using the same indicators?
If not, then it's probably best to optimize the entry and exit points separately because they belong to a different linear system. That is, you have a mixed model. And that maybe a good thing in some cases.
In contrast, if you're using the same indicators to determine entry and exit points, then I would solve their parameters together (simultaneously) because they all belong to the same linear system.
I have a question for you. :-)
And how exactly do you do that with WealthLab when the WL ScoreCard metrics (such as %profit or APR) that drive the parameter optimizer are measuring the overall performance of the strategy with both entry and exit points operating together?
For example, how does one run an optimization on just the entry criteria if the exit point is disabled altogether? Do you have a "custom designed" ScoreCard metric (referenced by the optimizer during optimization) that only samples %profit or APR immediately after opening a position so the exit point behavior matters less?
Perhaps there's a special WL extension, which I don't know about, that allows independent optimization of entry and exit points. Are we talking about analyzing MAE and MFE as ScoreCard metrics in greater depth and feeding those results back into the parameter optimizer? Has anyone tried doing that before, and do you get convergence?
QUOTE:
we learned that you should never optimize the entry and the exit together.
And how exactly do you do that with WealthLab when the WL ScoreCard metrics (such as %profit or APR) that drive the parameter optimizer are measuring the overall performance of the strategy with both entry and exit points operating together?
For example, how does one run an optimization on just the entry criteria if the exit point is disabled altogether? Do you have a "custom designed" ScoreCard metric (referenced by the optimizer during optimization) that only samples %profit or APR immediately after opening a position so the exit point behavior matters less?
Perhaps there's a special WL extension, which I don't know about, that allows independent optimization of entry and exit points. Are we talking about analyzing MAE and MFE as ScoreCard metrics in greater depth and feeding those results back into the parameter optimizer? Has anyone tried doing that before, and do you get convergence?
Hello Superticker,
thank you for your answer1
It is usually just about developing the entry while using a standard exit - e.g. selling after 3 bars.
Once the entry is fine - the strategy shows already a nice profit with this standard entry - development of the exit starts with the entry being fixed.
When developing the entry we use a method against over optimization: Only try your strategies on the Dow Jones index. If it can achieve more than 10% APR on the Dow Jones, you could run it on S&P100, 500 or NASDAQ, and hopefully the APR will increase vastly on the broader dataset. But then you are not supposed to make any further changes other than position sizing and the setting if multiple positions of the same security can be traded or not. This is a method to avoid over optimization / curve fitting.
Kind regards
thank you for your answer1
It is usually just about developing the entry while using a standard exit - e.g. selling after 3 bars.
Once the entry is fine - the strategy shows already a nice profit with this standard entry - development of the exit starts with the entry being fixed.
When developing the entry we use a method against over optimization: Only try your strategies on the Dow Jones index. If it can achieve more than 10% APR on the Dow Jones, you could run it on S&P100, 500 or NASDAQ, and hopefully the APR will increase vastly on the broader dataset. But then you are not supposed to make any further changes other than position sizing and the setting if multiple positions of the same security can be traded or not. This is a method to avoid over optimization / curve fitting.
Kind regards
QUOTE:
... about developing the entry while using a standard exit - e.g. selling after 3 bars.
Thanks for your answer.
I think if you're trying to determine which indicator to use for entry; that is, which indicator is predictive, this will work. But that's solving an indicator selection problem, not a parameter optimization problem. We are talking about two separate problems here.
Personally, for entry criteria, I try to look for trends in money flow indicators (e.g. KVO Klinger, MoneyFlowOscillator, MFI) because they factor in both price and volume.
---
I'm just wondering if it would make sense to create a special ScoreCard metric that would focus on %profit four days after each open position? The only reason for doing this would be to send metric feedback to the parameter optimizer. But I think this operation is not a parameter optimization problem, but rather an indicator selection problem.
Hi Superticker,
thanks for your answer. Yes, in principle you are right. Of course to have the most predictive indicators is the most important factor. Thank you very much for your ideas.
However, with my current strategy I have an exit a bit more complex and to optimize the entry and exit variables together seems to be way too slow. So, that is why I addressed it. However, I could simply create two walk forward studies: One with separate optimization of entry & exit, the other one with a joint optimization - and then see the results myself.
I understand that for a strategy as a whole it makes sense to go for marginal trades as well. But as I mentioned before, I try to calculate before the trade its prospects. For my strategy it seems that the volume change at the last two bars should be in the range 0.85 to 1.35 (or much higher) and a low price which went down at least half a percent (on the signal bar) increases the chances for a successful trade. My strategy of course is buying also outside of these ranges, but then the trades on average are less profitable.
Best regards
Werner
thanks for your answer. Yes, in principle you are right. Of course to have the most predictive indicators is the most important factor. Thank you very much for your ideas.
However, with my current strategy I have an exit a bit more complex and to optimize the entry and exit variables together seems to be way too slow. So, that is why I addressed it. However, I could simply create two walk forward studies: One with separate optimization of entry & exit, the other one with a joint optimization - and then see the results myself.
I understand that for a strategy as a whole it makes sense to go for marginal trades as well. But as I mentioned before, I try to calculate before the trade its prospects. For my strategy it seems that the volume change at the last two bars should be in the range 0.85 to 1.35 (or much higher) and a low price which went down at least half a percent (on the signal bar) increases the chances for a successful trade. My strategy of course is buying also outside of these ranges, but then the trades on average are less profitable.
Best regards
Werner
QUOTE:
create two walk forward studies: One with separate optimization of entry & exit, the other one with a joint optimization
I like that idea. Please share how that goes under this topic.
Volume is important. More volume implies more interest. (Just understand, on a swing, you can have high volume without much price change.) That's why I follow money flow rather than price.
Hi,
yes you are right. It does not fit the topic, I should open another one. It brought only a slight improvement and might be curve fitting.
yes you are right. It does not fit the topic, I should open another one. It brought only a slight improvement and might be curve fitting.
Your Response
Post
Edit Post
Login is required