Thanks. I get annoyed at having to defend my ideas sometimes and I now see where you are coming from. Also, sorry about the arrogant and non-informative comment aimed at you. I often forget that are many ways to build good systems and that my way is not the only way. That said, what I work towards is avoiding "fit" in any sense of the word. I avoid indicators/inputs most likely to my detriment. Most of my systems are purely price based. Also IMO, a very large sample size provides as much objectivity as possible, especially when observed over multiple market cycles. Also, I downloaded and ran the "Superbands" system you posted on WL and it fits my criteria for what a robust system should offer. What has been your experience with it thus far? Any live results to speak of? Mike
In reality, it takes an account well into the six figure range to work properly, and I believe disproves the EMH weak and strong. I'm rich for my age, but certainly not anywhere near a half mill that's probably optimal for this script. Real time I did do a stint with it on c2, but I was in surgery for a deviated septum on a day it would have increased over 17%. I had expected a large following to come from the WL community, but I just didn't see any point in continuing while paying $108/month for data from e-signal on just the NAZ100. I found results with eSignal's data looked just as robust as the backtest, and it's a real shame I couldn't do it. It took about 2 hours per night to setup in NT, and would probably require the same amount of devotion in TS. Would you mind posting your results on those arbitrarily selected symbols at 1% position sizing? I'm pretty sure it's as close to the holy grail as anyone can get from publicly available scripts.
I'm certainly one to utilize optimization, which I'm surprised some people don't use more often. It appears they simply believe that by "optimizing" with software they're curve fitting. This isn't the case when you do it properly, and finding the range of good values is actually where most of the quantitative analysis comes in. The values in SuperBands have been fixed for very long periods of time, but with regard to PTQQS the values have been the same since a pivotal nobel prize winning change to volatility based overbought, oversold indicators was discovered. Add standard deviations to some of your calculations creatively and you'll be very surprised with your results. I presented the orignal SuperBands to the Kentucky Math Association in April of 2006 during my final semester at Centre College. I believe most were surprised the results, and would be very happy if their fund manager could duplicate its performance, especially the improvement with Linear Regression Analysis. Someone could just look at the script and probably get a glimpse of the tricks I've used to write and build my systems. LinReg, standard deviations, these are essential in my opinion to be included in any script. Obviously, then, extremely advanced use of z-scores and critical values for PTQQS, which predicts fair market values. What most people don't understand is that the market mostly trades at fair value, and "sometimes" crosses above or below fair market value. When my predicted values are higher or lower depending on overbought or oversold conditions, only then will you trade, which is why I seriously doubt anyone has seen it before. Fair value is actually a very wide range that most likely spans 3-5% from the overbought or oversold levels. One of the best pieces of advice I ever got in terms of development was to only use non-extremum indicators. PTQQS truly utilizes it's own indicator that does not have extreme bounds, so in that sense you might see at as an indicator, but it has no upper or lower bounds. Developing these kinds of indicators is essential to long term success, and you can see how that's used with bollinger bands.
Interesting thread. I was wondering if you guys can suggest any books or links about quant modeling, algorithmic trading, etc. like a Quant for Dummies?
I hesitate to say this. (No, not really) there is something very Jack Herseyian about this entire thread.
You bring up a lot of good points. I use Walk Forward Optimization quite a bit to validate predictive value. On a day-to-day basis, I do tweak my mean reversion system parameters ever so slightly so as to get a good portfolio list for the next day, so I guess one could call this a form a optimization. The interesting thing about your SuperBands system is that it hasn't been properly tested. This is the result of your choice of tools. I can see several issues that when properly addressed will provide a more accurate measure of what this concept will trade like in real life. Foremost, the data from wealth Lab is of poor quality. Hence, using limit orders is a bit misleading. The LOD or HOD per Wealth Lab's data is often times wrong, so one needs to adjust expectations that limit fills will likely be exaggerated in ones favor. Second, you've done a good job in defining the bounds, now look at the actual execution. I've done a substantial amount of work with Market Profile and I traded MP concepts discretionarily for years. One of the many trades you'll find from market profile is what I call the value area completion trade. The value area is the price zone where, time-wise, 75% of the trading activity occurs. When prices move out and then back into the value area, the probability of the price trading back across the value area is very high. This is in a sense a mean reversion trade based on time spent away from value. The trick is that the price has to start moving back toward value after it has traded away from value. This concept is similar to your superbands and its what I use in the system I posted earlier. The trick is in defining where the point is, that when reached, signals a high probability of prices going back to value. Simulating the above requires custom tools due the the "buy Stop" type entry mechanism required. One essentially has to quantify the entry on a tick level to get it accurate. I suggest you build your own backtester to do this or you can use any number of off the shelf products. A cheap and easy one that comes to mind is Amibroker. In terms of indicators, you are right about non-extremum. The only set of indicators I ever use are true ranges. As a side note, are you familiar with volatility modeling? GARCH? If you have the inclination I suggest you read Engle's work on GARCH and see what you find. You've likely already seen it but: http://pages.stern.nyu.edu/~rengle/Garch101.doc My work as of late has been a refurbishing of this model using "Fuzzy" algorithms. The results are promising as I managed to create an adaptive standard model based on the fundamental concepts of GARCH. The key element is the inclusion of a Cauchy distribution function that allows for non-normal distributions (as I believe all volatility/price distributions are not normal). To the poster asking about good references for quantitative principles, the above paper is a good start. Also: http://press.princeton.edu/titles/8055.html http://www.amazon.com/Design-Testing-Optimization-Trading-Systems/dp/0471554464 Not sure of there is a "Quant for dummies" type text out there... you kind of have to have an advanced knowledge of stats to do this stuff. Regards, Mike
I actually don't think we're working off that different of a notion on trading.... I'll have a better reply soon. I'll have to look into GARCH a bit more. A case study in compounding for dummies I wrote here, recently, http://www.advisorworld.com/2009/01/17/the-power-of-compounding/ , and this was some of the allure of discussing the simplistic notion of beating the S&P by 10% annually for a long period of time.
On quant material, I really have to recommend busting ass through the CFA curriculum, then you'll see where it applies.
an inside story from IB is that a 40mm account went 10% USD 25%"short equities" 25% "short oil" 40% long 10yT 08-04-2008 the account didn't trade again in 2008.