What do you think of MrM's system

[jamiew]

Hi Jamie



205% is a nice number, but unrealistically high. It appears you took a normal average of the annualized results, which you can't do for returns. You have to take a geometrical average, by first normalizing the gains (+50% becomes 1.5, etc.), then you multiply all those results, take the (number of returns)th root from the result, and denormalize again. My result is 137%.

The second hair in the pudding is that starting the calculations with the annualized results may not be the best thing to do. To witness: when you average the results per stock and the holding period per stock, and annualize the averages, you get 50% annualized. I think we are approaching more realistic, if still inflated results here.

The third problem for the calculation is in the open positions. For reasons of caution, I always include them in my own calculations. My motto is, "let's trust it doesn't get any worse than this".

And my extremely cautious implementation of Followin MM is here, numbers and all: http://www.mrmarketishuge.com/showthread.php?t=455

Regards,

Karel

Ok. I agree my averaging is fairly unrealistic, it’s just that I’ve been comparing annular averages, although not realistic (since trades aren’t made on an every day business) Its just what I’ve been comparing so far. (I haven’t included openings, as I don’t know what MM has open)

How ever do you know what Mr Ms exit strategy is, it seem he stops at a gain of 16% generally, is intentional, what his indicator for ending a trade or is he stopped out? I need to know to be able to mimic.

[Karel]

Hi Jamie,

Officially there is a 15% sell stop. As $$$Mr.Market$$$ implements this as a mental stop (I use limit orders myself), the gain for a winner is 15% or more. You could implement this 15% with or without calculating the transaction costs. $$$Mr.Market$$$ ignores dividends (except for the gains calculation).

Regarding the losers, things are equally clear. $$$Mr.Market$$$ continues to hold them, because, based on his fundamental analysis, he is convinced that these are basically good companies, quite able to turn around when they get into a dip. When the drag of the open positions becomes too great, with the result of not enough turnaround, he may decide to sell some of the open positions and to adjust his streak. This is almost impossible to backtest. Other sell strategies for losers are possible (dropping revenue/earnings, f.i.), but fixed percentage stops seem to hurt the returns.

For a strategy you could look at:
- The universe of stocks from which $$$Mr.Market$$$ picks his winners?
- The sell strategy for a) winners (in my Marketocracy portfolio I try to let the winners run), b) losers.

And test!

Regards,

Karel

[skiracer]

Hi Jamie



205% is a nice number, but unrealistically high. It appears you took a normal average of the annualized results, which you can't do for returns. You have to take a geometrical average, by first normalizing the gains (+50% becomes 1.5, etc.), then you multiply all those results, take the (number of returns)th root from the result, and denormalize again. My result is 137%.

The second hair in the pudding is that starting the calculations with the annualized results may not be the best thing to do. To witness: when you average the results per stock and the holding period per stock, and annualize the averages, you get 50% annualized. I think we are approaching more realistic, if still inflated results here.

The third problem for the calculation is in the open positions. For reasons of caution, I always include them in my own calculations. My motto is, "let's trust it doesn't get any worse than this".

And my extremely cautious implementation of Followin MM is here, numbers and all: http://www.mrmarketishuge.com/showthread.php?t=455

Regards,

Karel

Thank you Karel for bringing some realistic numbers to the discussion. Lye, I would gladly sit in a room alone with Jamiew if that's what it took to acquire another flask of the "goji" but do you think he would come to Jersey to do that.
Just for my own info was I to harsh or out of line with my comments to Jamiew. I don't think so myself but another memebers perspective is always good to hear. In my own opinion I feel that one should have a much better grasp or understanding of the entire subject matter before considering taking on something like what he is trying to do. His numbers are out of line with what Ernie's true percentages are on a yearly basis much less for the entire period plus he has backed up on several occassions, see his last post, where he has had to explain why his numbers weren't working or correct as to what he is talking about. If anyone is going to post here and put their comments up for public scrutiny then they should be thick skinned enough to handle anyone's comments whether they agree or disagree with their points an arguement. And spend enough time on the subject matter to get it right and present it right if you are going to do something like this. I don't think he has done that to the degree that he should have done from the start.
Let me go on record and state that I'm not trying to disuade Jamiew from posting here or from running his experiement. I would like him to stay an explain his theory and what he is trying to do and become a strong participating member of this forum, but get it right and know the subject matter before someone like myself starts questioning what you are doing. And please don't try to explain it away as an attack on your person. It couldn't be farther from the truth.

[jamiew]

Skiracer certainly has allot to say about what people shouldn’t and should know, before they start talking about it, and then he starts talking about what is wrong or right in statistics, clearly something he must know an awful lot about

There are many different types of average

Root mean square, common in science
mean median and mode (maybe terms skiracer might more familiar with)
Geometric mean (as carl used)
Harmonic mean
Generalized mean
Interquartile mean
the list goes on

Still im sure he already know this, otherwise he wouldn’t have spoke about it and put himself on display making himself look a bit stupid thinking there was only one right answer, and telling everyone what people should and shouldn’t know before they speak about it... the irony

There are different ways at looking at results, I have chosen my way as result at either end of the distribution stand out, so make it easier for me to compare. Normalising would dampen this affect, but give a more realistic answer to the trader. Still skiracer, criticise me as you wish, as i have said in an earlier post, those who shout the loudest often have the least to say.

edit - ignore is indeed a handy function, this insult throwing contest with skiracer is pointless, but thank you for the entertainment, it has been most enlightening

[jamiew]

Officially there is a 15% sell stop.

Regarding the losers, things are equally clear. $$$Mr.Market$$$ continues to hold them, because, based on his fundamental analysis, he is convinced that these are basically good companies,

For a strategy you could look at:
- The universe of stocks from which $$$Mr.Market$$$ picks his winners?
- The sell strategy for a) winners (in my Marketocracy portfolio I try to let the winners run), b) losers.

Ok that’s great, exactly what i need to know, thanks karl

And so do you personally cut your losses short at say a 15% loss? I don’t think you finished b)

[skiracer]

Ok. I agree my averaging is fairly unrealistic, it’s just that I’ve been comparing annular averages, although not realistic (since trades aren’t made on an every day business) Its just what I’ve been comparing so far. (I haven’t included openings, as I don’t know what MM has open)

How ever do you know what Mr Ms exit strategy is, it seem he stops at a gain of 16% generally, is intentional, what his indicator for ending a trade or is he stopped out? I need to know to be able to mimic.

I'll let you go on yourself. That in itself will be enough to expose just how much you know about any of this. It's so typical when individuals like yourself revert to name calling an insinuations. You should spend that time studying the subject matter you want to discuss to at least have a decent handle on it. I won't have to open my mouth or say a thing. You'll be more than sufficient at providing that info for all of us.
You should put me on ignore, that way you won't have to learn anything that might be beneficial or enlightening.

[peanuts]

An AI system already in place

Jamie,

I found this neural system, maybe you want to check it out:

http://toplnk.net/trading-system/

I do not have any experience with it. I just came to it while doing a search on one of my companies. If you find anything useful from it, let us know.

[skiracer]

If Jamiew or anyone else wants to better understand and learn something about Neural Networks and how they would apply to trading they can go to this site: www.traders.com. Once there you will have to sign up for the free 30 day trial to their site. There is no cost, no obligation, and they don't ask for a credit card beforehand. Just fill out the application and you will get in almost immediately after they send out your password via email. Once in fill in "neural" in the search block an it will take you to a number of articles which will explain the basics of using neural networks and the code for writing a number of programed scans for neural networks. Aside from the neural stuff the site is great for all kinds of trading info to much to try to mention in this post. Traders.com is the home site of a magazine called "Technical Analysis of Stocks and Commodities" which costs $49.95 per year for a subscription and should be in every traders library. Great stuff plus the site has a wealth of FREE information available with the subscription. It'll make you a better trader, but then who needs that.
There are three very good articles on using neural networking in trading which will bring you up to speed on the subject. They are:
"Using Neural Networks for Financial Forcasting" by Lou Mendelsohn, "Neural Network" by Jason K. Hutson, and "Trading the S&P with a Neural Network", by Jeremy G. Konstenius. Learning and understanding the use of the code to program scans for the setups you are looking for is complicated for someone like myself, not having any training in that area, but these articles will get you started an up to speed if you want to spend the time with it.

[IIC]

Dave wrote some articles for that magazine...I get comped from them...Although they keep sending me bills that I never pay...But they keep coming...I happen to like TAofS&C...They have some interesting articles...For example...I happen to watch the short term A/D's...3,5,10,12 day and I do check the 30 day...But they had an article a year or two ago which was longer range on A/D's going back to the early 60's...Long range it does not seem to be indicative...Doug(IIC)

[skiracer]

Dave wrote some articles for that magazine...I get comped from them...Although they keep sending me bills that I never pay...But they keep coming...I happen to like TAofS&C...They have some interesting articles...For example...I happen to watch the short term A/D's...3,5,10,12 day and I do check the 30 day...But they had an article a year or two ago which was longer range on A/D's going back to the early 60's...Long range it does not seem to be indicative...Doug(IIC)

It's a good mag and site for traders. There was an article, I would have to go back and look for it, on neural networks that was quite interesting an except for the problem of understanding how to program the code, and that your platform also has to be set up for running that kind of code, was very interesting and was working out very profitably for the author of the article. He would program the code for the setups that he wanted and when specific points were met it would trigger buy and sell signals automatically. He didn't even have to be at his screen once the programmed code was in place. He was still refining it at the time of the article but even with that was doing quite well with the system. Like any scans you have to know exactly what you are looking for in any stocks cycle and then program the code to prompt the scan for those results. I've been studying the concept in a very loose fashion just because I find it interesting and feel that it does hold water. Just very complicated. To some it comes easy while I have to really work at it. I'll go back and find the authors name and the article.

[IIC]

I also believe that you have to be very flexible...Different parameters work in different types of market action...I say that's why most system traders fail...They think one set of criteria works in all types of markets...It just doesn't work that way IMO...Doug(IIC)

[skiracer]

I also believe that you have to be very flexible...Different parameters work in different types of market action...I say that's why most system traders fail...They think one set of criteria works in all types of markets...It just doesn't work that way IMO...Doug(IIC)

I agree with your first statement Doug an you're probably right to a large degree about most system traders failing but I don't think that as large a percentage of the top guys are failing. Most of us at Erinie's are just trading stocks. I think most traders using mechanical systems, like Neural Networks, are trading futures in money, commodities, bonds, an everything else except stocks. Speaking for myself and not having the experience at either trading anything other than stocks an options on stocks I would say the mechanical systems like Neural Networks work much better with those types of trading vehicles. I think Spike trades the e-mini's on the DOW and the Nasdaq and has a system in place which signals trend reversals and buy and sell points. Not sure but maybe he'll explain some of it an how it works without giving up his system. I think he has been quite successful with it.

[ParkTwain]

You might say that any "system" is by definition "mechanical." It could be just a matter of differentiating systems by degree of complexity. I think that the idea about NN stuff is that it is a technology that is inherently "adaptive" and works best when there isn't a fixed set of relationships (or even when it is unknown whether any such relationships exist) among the inputs or drivers of the process that the NN is addressing, and as long as there aren't TOO MANY of those inputs and drivers. The problem can be that NN in the markets might take TOO LONG to adapt a solution to the problem that is actionable in the markets.

[ParkTwain]

You might say that any "system" is by definition "mechanical." It could be just a matter of differentiating systems by degree of complexity. I think that the idea about NN stuff is that it is a technology that is inherently "adaptive" and works best when there isn't a fixed set of relationships among the inputs or drivers of the process that the NN is addressing, and as long as there aren't TOO MANY of those inputs and drivers. The problem can be that NN in the markets might take TOO LONG to find a way to adapt.

Well, well, I was just posting with ski at the "dealing with a recession" thread and I came across this VERY INTERESTING paper about a project (an academic exercise, basically) to find an automated way to find and act on ETF arbitrage opportunities in the DIA (diamond) ETF security. The paper says that the project developed a neural network application to address the situation.

http://www.kellogg.northwestern.edu/faculty/jagannathan/teaching/practicum/2004/files/Final-Reports/ETF%20Arbitrage%20.pdf

For the period Jan 1, 2002 to May 5, 2004, the team found that there were in fact 19 instances of at least a $0.35 spread, either positive or negative, between the NAV of the DIA and the underlying basket of shares. They had determined that this spread level provided the minimum attractive opportunity for arbitrage, based on the current prevailing cost structure for conducting ETF arbitrage activity. The team tried using a few conventional predictive techniques (linear regressions against each of several conventional trade activity parameters as independent variables) but they failed to usefully predict the occurrences. They next tried developing and training a four-layer neural network system. The system was able to accurately predict and identify a qualifying spread situation over the data set for the chosen time period.

The paper lost me when it described the mathematical functions programmed into certain of the NN nodes ("neurons"). The NN technology uses a feedback mechanism and "convergence to goal" rules to find the combination of (in this particular application, nonlinear) functions that achieve the system's goals (a desired degree of predictive accuracy of the occurrences in the dataset of the minimal acceptable ETF arbitrage criteria). Training this model required 100 to 200 iterations. The team had split the time series data (covering the identified time period mentioned earlier) and used the first half of that data to train the system, then used the second half of the data to test the system.

[IIC]

This is becoming a bit complicated for me anyway...I've gotta a better system...Buy stocks that are gonna go up...Sell or short stocks that are gonna go down...Works pretty good IMO...Doug(IIC)