Last week I wrote about a very simple rotational system where one would rotate in long the top 20 stocks with the highest negative 5 day rate-of-change and hedge those by shorting the top 20 stocks with the highest positive 5 day rate-of-change.
The post generated some interest, and so I wanted to follow up by checking in on how those 40 positions performed. The spreadsheet with the results is below.
The system generated profits of $236.90 with an average trade of 0.24%. The win percentage was 47.5%.
This type of performance would generate a compound annual rate of 11.32%.
Unfortunately, commissions were not included. If we allowed .005/share for commissions, the system just barely breaks even.
Some Observations…
As David Varadi has already begun to demonstrate, there are many tweaks and variations to be made before one could begin to consider this a viable system. However, with only one week’s worth of data, we can make some observations that are likely to improve performance.
1. What is the market doing? If the market is in an uptrend, say above the 50 day average, perhaps we reduce our short exposure and increase our long exposure?
2. Using the Maximum Favorable Excursion (MFA) and Maximum Adverse Excursion (MAE) we can begin to think about using stops, profit targets, and dynamic exits to improve performance.
3. Since this system basically just ranks stocks, how do we refine our ranking characteristics/criteria to better select stocks in order to improve performance, and possibly reduce exposure and commissions?
In summary, if one has ever wondered how hedge funds approach the market, when they have hundreds of millions to deploy and need a strategy that scales, this method, while very basic and overly simplistic, with refinement should generate returns better than the benchmarks.
Below, purely for fun and chart porn, is the equity curve from this system, from 1/1/2000 to today. Of course commissions and slippage were not included. It shows a compound annual rate of 18.11% with a max drawdown of 16.77%.
I appreciate your dedication and (ability to do) to back testing and analysis, and am gald you are part of this website. I’m a newbie, so rather than try to offer some insight, I’ll leave it at thank you.
MX2, thanks. Feel free to ask questions. I love to talk about this stuff.
May I ask what you use to backtest?
_johny, I use AmiBroker with a data feed from Norgate, premiumdata.net
Both come highly recommended. That report was generated from AmiBroker and simply exported to excel, where I added the win % and profit and average trade.
Hey Wood,
I have a question for you independent of this post. Hope you do not mind.
I believe that you use an Excel Add-In to graph Optimization data? If so which one?
Also, when running optimization or before, what criteria do you use to define if two variables are related? In other words, do you think it is important for multiple variables to be tested in relation to others and/or independently?
Thank you, as always, for your input.
Redshark
Hello again Red.
The add-in I use was designed by the guy who built AmiBroker, I believe. I do not believe it works without AmiBroker.
Your question about variables is very important. I’m not sure that I have a good answer in terms of knowing if variables are related. I suppose there are some statistical tests that could determine if they are, but I have never taken the time to run them.
My approach to optimization is as follows:
1. I typically test entries and exits separately, against a time stop. What I want to see is if the entry beats a random entry, and for the exit, I want to see at what point a time exit would outperform my exit. If a time exit outperforms, for a short-term system, then I assume that my exit is not very good.
2. Once I understand the effect of the entries and exits, then I may begin looking at improving the system by i. reducing exposure but increasing the size of the average trade, ii. optimizing position sizing (which is related to reducing exposure) and iii. adjusting how the entries are ranked, which is to say, If I have 5 entries and 3 open positions, which 3 of the five will be better?
3. Once I have all of that tweaked, I might, almost entirely for fun, run optimizations of multiple variables.
I think it is very important to first run them independently, in order to be able to know clearly how that variable affects performance, before combining it with another variable.
Good to hear from you Red….
Just a thought but I am curious as to how the stocks in the top 20 and the bottom 20 are correlated to each other.
Hedging via this means would work only if the top 20 and the bottom 20 are inversely correlated but, you’d get really hurt if the whole market were to auto-correlate like we saw several times in the last 20 months.
Leaving this link for anyone who’s interested in a spreadsheet way of calculating covariances. I’m sure Wood has a better way of doing those calcs. http://www.stator-afm.com/stock-correlations.html
p.s. if you doubt that the entire market could auto-correlate, check the link and look at the last table. there’s not an inverse correlation in the lot.
Hello Cuervos, good to hear from you.
I disagree though. While intuitively, yes, it seems they would need to be inversely correlated, the data does not show that to be true. Just look at the equity curve, over the last 20 months.
I think there is something more to it than correlations. More than anything else, I think it is just that there has been a lack of follow through, across many time frames, over the last decade. It is simply reversion to the mean.
Here’s a (possible) head scratcher for you….Do not the ROCs falling at either extreme show an inverse correlation, or sorts?
By the way, thanks for the link!
OK, just off the top of my head here but, would one possible optimisation in my brain would be to look for inner correlations and minimise the risk.
instead of 40 purchase, there would be maybe 5 but, i wonder if they would revert to mean in an inverse manner.
20 and 20 is not optimum. I’ve already gone beyond that.
It is probably not too much of a surprise that this performs better with a 2:1.3ish ratio of longs to shorts. However, the ratio should/could be dynamic. I would imagine there are times when you want more or fewer shorts.
When would those times be, and how would we quantify them so we can test the assumptions?
by the way Cuervo, what you are saying is to rank higher the issues that are not correlated? And exclude those that are? What would be your thresholds?
All this can be done in AmiB, but my prog skillz aren’t there yet..
I seriously learned about a lot of this, but however, I still thought it had been valuable. Sweet task!