% Accuracy
- Thread starter GIG
- Start date
Quote from GIG:
bubba7,
Yes, doing this stuff is really fun.
What I'm presenting in this post is just the 'tip of the iceburg' so to speak. I've taken into consideration other factors as well, such as
-percent return per transaction
-percent return (only when correct)
-percent return (only when incorrect)
-smart percent return (drops the stock if it dips below 1.5%)
-intraday volatility
-yearly results for the above
-the effect of the "next day's" opening and closing price
-price movement between the 'buy' and 'sell' day.
There's probably a few more that I haven't added. Anyway, as you could imagine if I wanted information about all of the above I would have a post 10 pages long
Looking ahead for the fall, I'm planning to focus on
Money management
Slippage Effects
Commision Effects
If there are crucial metrics I've left out, please let me know.
All of this research has been implemented as one big software system (I've been developing the application for quite some time now).
There's quite a few electronic resource available, including a supercomputing facility (which can speed up analysis quite a bit), access to electronic publications, other universities, etc...
Brandon
Thank you for responding. I appreciate it.
OddTrader,
Not to worry. The project is nearly complete, I'm just wrapping things up.
At the beginning of the semester my project supervisor and I looked through numerous papers - and like you mentioned in your post, you've read about a project similar to mine. This has been the case, close but no cigar.
Because of this, we made a decision at the beginning of the semester to move forward with the project, since there was a significant amount of work to do. The came to the conclusion that time required to find suitable papers as a basis for my project would not be of value if my system wasn't complete. So, I decided to see if I could get some information from 'real-world' traders as a supplementary discussion.
Thanks for caring though
Brandon
Not to worry. The project is nearly complete, I'm just wrapping things up.
At the beginning of the semester my project supervisor and I looked through numerous papers - and like you mentioned in your post, you've read about a project similar to mine. This has been the case, close but no cigar.
Because of this, we made a decision at the beginning of the semester to move forward with the project, since there was a significant amount of work to do. The came to the conclusion that time required to find suitable papers as a basis for my project would not be of value if my system wasn't complete. So, I decided to see if I could get some information from 'real-world' traders as a supplementary discussion.
Thanks for caring though
Brandon
Doing any field studies in the trading industry has been the harest job for academics, if not impossible. Because very few, if any, would like to openly disclose all information (trade secrets) to outsiders, whether non-disclosure agreement is signed.
Another key issue is whether the information collected through any newsgroups or forums over the Internet is reliable up to your expectations/requirements.
Nice attempt anyway!
http://www.wallstreetuncut.com/wsuArchive.htm
In regards to the linearity of markets and the potential success of a linear model listen to the interview with Anders Johansen. There are other interviews which discuss this issue as well, but I can't remember which. Anders also has a website with a number of research papers available in .pdf format for download (this is all free). Diddier Sorniet (sp?) worked with Anders and is an expert as well. There may be an interview with him too, I would have to check, but if there is give it a listen I am sure the linearity issue would be broached.
In regards to the linearity of markets and the potential success of a linear model listen to the interview with Anders Johansen. There are other interviews which discuss this issue as well, but I can't remember which. Anders also has a website with a number of research papers available in .pdf format for download (this is all free). Diddier Sorniet (sp?) worked with Anders and is an expert as well. There may be an interview with him too, I would have to check, but if there is give it a listen I am sure the linearity issue would be broached.
"How did you get past the data problems? Linear Time series analysis usually apply to continuous functions. And markets are anything but linear except for applying different kinds of series and synthesizing (or analyzing).
What range of data types and kinds are you allowowing yourself to use?"
I'm assuming the data problems you're talking about (with respect to continuous functions) is the weekend/holiday effect.
This was a tough decision to make. This also reminded me of one of the metrics that I forgot about. I looked at several ways of handling missing Saturday/Sunday data, I ultimatley decided to emulate a 5-day week instead of a 7-day week.
I agree that markets are anything but linear. This pretty much ties is with a search for 'pockets of predictability' in chaotic systems.
A nonlinear 'pure' random walk process with two states (up, down) should generate a probability of 50% for each state.
Just like flipping a coin. It's a completely random process, and the probability of getting a head or a tail is 0.5.
I'm investigating to see whether 'simpler' linear techniques can exploit obtuse/masked pattern in the market - which may be interpreted as temporal market inneficiencies. Hopefully this makes sense, I summed it up really, really quickly.
As for data types, are you referring to things such as open, close, volume or float, int, double?
Regards,
Brandon
What range of data types and kinds are you allowowing yourself to use?"
I'm assuming the data problems you're talking about (with respect to continuous functions) is the weekend/holiday effect.
This was a tough decision to make. This also reminded me of one of the metrics that I forgot about. I looked at several ways of handling missing Saturday/Sunday data, I ultimatley decided to emulate a 5-day week instead of a 7-day week.
I agree that markets are anything but linear. This pretty much ties is with a search for 'pockets of predictability' in chaotic systems.
A nonlinear 'pure' random walk process with two states (up, down) should generate a probability of 50% for each state.
Just like flipping a coin. It's a completely random process, and the probability of getting a head or a tail is 0.5.
I'm investigating to see whether 'simpler' linear techniques can exploit obtuse/masked pattern in the market - which may be interpreted as temporal market inneficiencies. Hopefully this makes sense, I summed it up really, really quickly.
As for data types, are you referring to things such as open, close, volume or float, int, double?
Regards,
Brandon
I was referring more to the daily gap after close and continuing until the open the next day. This interval through all thecivilized time zones around theworld often has a greater effect than the few hours the markets are open daily in the US.
from your comments I am guessing that you do not use market extremes like high or low for your data points. You probably have chosen an end value like open or close. Open has more items impacting it as we know so you probably chose the close.
The risk encountered from this daily gap is best characterized by the differing margin regirements for intraday traders versus interday traders.
All in all, you have the kind of model that is based on data that is not persuasive for handling important questions for making money. The very difficult challenge you created was choosing 500 stocks and making that sample a random one. There are no financial incentives for ever doing that any time.
I believe it is also very important to use market variables for analysis as well. I read your posts and it looks like you are using just two variables in your model: price and time. The price values also only occur at one instant of time per day. This is a very rarefied atmosphere in which to be operating. Too risky for me were I to draw any conclusions for making money.
Please do not regard this as a put down in any way. I am just making comparitive observations.
For optimizing making money all people get to micro models that are skewed as favorably as possible using every technique they can. The mass of data the market generates is definitely something to be avoided at all costs. It appears that it is now so big that it serves fairly well to generate randomness for mathematicians. I know you are in the practical realm of applied stuff using computers.
Once you move to the most exciting problem of all, namely making money, I'm sure you will move away from the macro and random stuff. That stuff they generate around 50% is such a classically poor vantage point from which to get into making money.
from your comments I am guessing that you do not use market extremes like high or low for your data points. You probably have chosen an end value like open or close. Open has more items impacting it as we know so you probably chose the close.
The risk encountered from this daily gap is best characterized by the differing margin regirements for intraday traders versus interday traders.
All in all, you have the kind of model that is based on data that is not persuasive for handling important questions for making money. The very difficult challenge you created was choosing 500 stocks and making that sample a random one. There are no financial incentives for ever doing that any time.
I believe it is also very important to use market variables for analysis as well. I read your posts and it looks like you are using just two variables in your model: price and time. The price values also only occur at one instant of time per day. This is a very rarefied atmosphere in which to be operating. Too risky for me were I to draw any conclusions for making money.
Please do not regard this as a put down in any way. I am just making comparitive observations.
For optimizing making money all people get to micro models that are skewed as favorably as possible using every technique they can. The mass of data the market generates is definitely something to be avoided at all costs. It appears that it is now so big that it serves fairly well to generate randomness for mathematicians. I know you are in the practical realm of applied stuff using computers.
Once you move to the most exciting problem of all, namely making money, I'm sure you will move away from the macro and random stuff. That stuff they generate around 50% is such a classically poor vantage point from which to get into making money.
bubba7,
Thanks for the feedback. As you can imagine, this is a pretty specific area of research. We don't have any profs. on campus who do this kind of work (one of the limitations of undergrad research - by your senior year, after you figure out what you're interested in, you may not have faculty doing similar work). It's nice to hear from people more experienced in the field.
Yes, the two primary factors used in analysis are price (close) and time (or the manipulation thereof). These are the only two parameters which can trigger a buy signal. I also log additional information, such as the next day's open (after a buy signal), the %change between the intraday open and low, and %change between the intraday open and close. I hope to draw some type of conclusions by observing this data.
"The very difficult challenge you created was choosing 500 stocks and making that sample a random one. There are no financial incentives for ever doing that any time."
That's a very good point. My primary reasoning for choosing companies at random was to avoid the potential for overfitting results, tailored to a particular sector, for example. Also, I didn't want to target companies who share a characteristic, not found in other groups, and get results (for better or worse) that could potentially innapropriately reflect the algorithm's worth as a predictive tool.
I understand your point, however, regarding the use of other factors as predictive measures. Using only price/time won't be the 'end all' of market analysis. I really enjoy doing this kind of stuff, so I imagine I will have many years left to try a variety of techniques. What I have done is built a core framework (from a software point of view), which can be extended to an n-degree of complexity as needed. Ultimatley, I wanted to start small and simple, understand the relationship (or lack thereof) of primitive/raw metrics, and from there continue to exploit market behaviour with more complex techniques. It's been a pretty good learning experience, which will be invaluable for future work. At this point, I wouldn't want to have a solution so complex (even if it worked well), as to not have a reasonable explanation for the solution's behaviour.
The micro analysis is definitley starting to formulate, (at least in my head right now). It's just a matter of time before I develop a more focused solution.
Quick question: If 50% is a classically poor vantage point for making money, what would you consider to be a good vantage point (in your own personal opininon)?
Regards,
Brandon
Thanks for the feedback. As you can imagine, this is a pretty specific area of research. We don't have any profs. on campus who do this kind of work (one of the limitations of undergrad research - by your senior year, after you figure out what you're interested in, you may not have faculty doing similar work). It's nice to hear from people more experienced in the field.
Yes, the two primary factors used in analysis are price (close) and time (or the manipulation thereof). These are the only two parameters which can trigger a buy signal. I also log additional information, such as the next day's open (after a buy signal), the %change between the intraday open and low, and %change between the intraday open and close. I hope to draw some type of conclusions by observing this data.
"The very difficult challenge you created was choosing 500 stocks and making that sample a random one. There are no financial incentives for ever doing that any time."
That's a very good point. My primary reasoning for choosing companies at random was to avoid the potential for overfitting results, tailored to a particular sector, for example. Also, I didn't want to target companies who share a characteristic, not found in other groups, and get results (for better or worse) that could potentially innapropriately reflect the algorithm's worth as a predictive tool.
I understand your point, however, regarding the use of other factors as predictive measures. Using only price/time won't be the 'end all' of market analysis. I really enjoy doing this kind of stuff, so I imagine I will have many years left to try a variety of techniques. What I have done is built a core framework (from a software point of view), which can be extended to an n-degree of complexity as needed. Ultimatley, I wanted to start small and simple, understand the relationship (or lack thereof) of primitive/raw metrics, and from there continue to exploit market behaviour with more complex techniques. It's been a pretty good learning experience, which will be invaluable for future work. At this point, I wouldn't want to have a solution so complex (even if it worked well), as to not have a reasonable explanation for the solution's behaviour.
The micro analysis is definitley starting to formulate, (at least in my head right now). It's just a matter of time before I develop a more focused solution.
Quick question: If 50% is a classically poor vantage point for making money, what would you consider to be a good vantage point (in your own personal opininon)?
Regards,
Brandon