Correlation cont'd. Well it didn't turn out as I had expected. I applied the weightings as spit out by the program and did the same totals as was done earlier. In column E are all 3 models combined with the optimal weighting. In column F is the combination of models 1 & 2 and in column G is the combination of models 1 & 3. It's pretty easy to see that using 1 & 3 is superior to 1 & 2 mainly because the largest losing month for 1 & 2 is more than double the largest loss for 1 & 3. So if you were to trade at twice the size using 1 & 3 you'd have a lower max losing month and a higher average profit per-month than 1 & 2. However this shows that if you used 1,2, and 3 in the ratios provided it would kick some serious butt over using 1 & 3 by itself. Also 1 & 3 have 11 losing months as compared to just 7 for all 3 models combined. In digging through the data I noted that model 3 didn't have very consistent results so the 1.57 profit factor probably isn't representative of it's overall results. In fact model 3 lost money for all of 1999 (a year in which the combined results for 1 & 3 show 6 of the 11 losing months). This just shows once again that the concept is only as good as the underlying models it's built on.
Correlation cont'd. Well, here's the snapshot of model 3's annual results. As you can see the win % ranges from 35 - 60 and the profit factor ranges from .79 - 2.89. Obviously this was a bad choice of a model to include in the tests. If the models are inconsistent, then everything else that flows from them will be suspect. One thing I meant to point out was look at the largest losing month for all 3 models combined with the weightings. Notice it's now lower than any of the models individually. I love this stuff. I hope I was able to convey the general concept without getting too deep. I spent hours on this trying to figure out how to boil it down to these few posts. In reality this took more than one full notebook of work over a long period of time. If you have any questions I'll try and answer them tomorrow morning. I'm going out to dinner now and I'll be around then. Enjoy the weekend!
How does one go about finding markets for these systems? Is each system you develop market specific? Or, was a system design inspired by experience with one market or another? What individual market tendencies do you look for before selecting a system for trade in that market? How do you characterize a market?
Thank you very much for the ongoing sharing of your thoughts on systems development. It has been very helpful and is very much appreciated. I'm hoping you might have thoughts on the following question. I use automation to trade individual stocks, and use systems on many different stocks. At any given time, I might be paper trading 300+ stocks, and have 100+ stocks active for live trading. I use a 93% confidence level criteria over the last 120 paper trades to determine whether to activate a security for live trading. In this way, I activate strictly based upon the risk/return of a specific security's past profit performance. However, I really like the way you activate based upon the actual diversification that the new security (or system) adds to your overall profitability. The issue I face is liquidity. For individual stocks, I run into liquidity issues for the less active stocks (<1M shares per day) that I trade which can reduce or eliminate system profits as my trading size increases in that stock. As a result, my 'optimal' combination of securities (i.e. systems) may only be accurate to trade will a smaller level of capital due to deteriorating performance caused by liquidity at larger position sizes. Do you have any suggestions on how I might consider improving my security (or system) activation routine based upon balancing both the liquidity constraints and diversification value? Thanks for your thoughts. -Eric
acrary, can you post the algorithm you use to adjust the number of contracts traded to volatility? Is it similar to the one used in the turtle trader method?
acrary, thank you. your contributions to this site are invaluable. you are definitely honored. regarding the correlation of multiple instruments versus model correlations: after finding no correlation between instruments should we look at the instruments as separate models if we are using multiple instruments to one model in order to carry on the sharpe work? i'm not sure if that's what you were implying here in response to virgins post: so for instance, instead of the spread sheet comparing model1>model2>model3 to one instrument it could compare instrumentX>instrumentY>instrumentZ to one model, no?... as opposed to only measuring say the correlation of the 30 day price average between instruments. just to make sure i understand the big picture, this is all being done to increase frequency in the desired profitable timeperiod, correct? so if i'm understanding, sub-par models tested individually with low frequency can be morphed into an above-par model when combined with other non-correlated sub-par models (assuming they're not too sub-par), thus increasing frequency, consistency, and lowering the need for a higher profit factor? thank you for presenting the information the way you did, i would not have made that connection otherwise (assuming i'm on the right track). fascinating. onelot
for those of us with no stats books, or if you just don't want to look it up, here's a cool little site that will calculate your std dev for your required confidence level in a neat little animated graph: http://davidmlane.com/hyperstat/z_table.html onelot