Hi Ton, I agree that the rule of thumb is subjective. So far, I've been willing to live with it.
It appears that you and I have different expectations of IS/OOS window sizes. I treat the calculation of walk forward window sizes as a second pass optimization, similar to a simple moving average (SMA) crossover system. - There are two variables (e.g. IS length/OOS length vs. fast SMA/slow SMA) - An optimal combination is desired - We use a fitness function to measure optimal (e.g. OOS:IS ratio vs. CAR/MDD) This is how I try to satisfy your Aronson quote "Each strategy will have its own best values for IS/OOS periods". Upon finding an optimal CAR/MDD using fast SMA/slow SMA, we should theoretically be able to trade that same optimal combination of fast SMA/slow SMA over different time periods and expect to get a somewhat stable CAR/MDD (subject to changing market conditions). I would not expect combinations of fast SMA/slow SMA to be stable relative to each other. Looking at a 3-D graph for this crossover system will reveal peaks and valleys. Taking a single slice of that graph (i.e. holding slow SMA constant and varying only fast SMA) will reveal a rising and falling wave. So, I would expect exactly the same in the IS/OOS experiment you describe. You are simply taking a slice of the 2 variable optimization graph (holding IS constant and varying OOS). I would expect a rising and falling wave representing the peaks and valleys that would appear on the full 3-D graph. If I optimize the ratio of OOS:IS using IS length/OOS length, then I expect to get a somewhat consistent OOS:IS ratio (subject to market changes) when using that same optimal IS length/OOS length over different data ranges. I don't expect to get a stable OOS:IS ratio using a fixed IS length and variable OOS length. Mike --- In [email protected], "Ton Sieverding" <ton.sieverd...@...> wrote: > > Thanks for your patience Mike -) > > 1. I know Pardo disagrees with Aronson. And yes I am also using Pardo's rule > of thumb. But a rule of thumb without a scientific explanation is still a > rule of thumb and therefore subjective. The result of this is when taking 1/8 > in stead of 1/3, I am getting a completely different results. That's what > Aronson tells me. So I do not understand why Pardo disagrees with Aronson ... > Of course I should ask him. And I will ... > > 2. Here you are telling me what Aronson says : "Each strategy will have its > own best values for IS/OOS periods". But trying to find the best values is > empirical and therefore without having a 'good theory' why your are getting > these values is highly subjective. Pardo is not giving me this good theory > and Aronson tells me this good theory does not exist ... > > 3. With regard to our topic, it's not so important which objective function > you are using for the WalkFoward. In general I use the CAR/MDD. But whatever > OF gives you the same random WalkForward results. Where of course by > definition you should use a return/risk related OF ... > > 4. The way I am analyzing the WalkForward result is simple. I am calculating > the differences between the IS and OOS results in percentages from OOS. Then > I am taking the average and standard deviation of all these percentages. This > gives me an idea about the average IS/OOS error as well as the spread around > this average. For the same AFL using the same Symbol you should do the > WalkFoward in the way I mentioned in my previous email and calculate the > above average/stdev relation. In order to get a stable WalkForward result > being independent of the IS/OOS ratio, the average/stdev relation should be > more or less stable. It's not. It's highly dependent on the IS/OOS ratio you > are using ... > > BTW ... To get things straight, I am not throwing WalkFoward out of the > window. I am just trying to believe in what I am using. And it's getting more > and more difficult for me ... > > Regards, Ton. > > > > > ----- Original Message ----- > From: Mike > To: [email protected] > Sent: Monday, October 05, 2009 11:09 AM > Subject: [amibroker] Re: Is the Walk forward study useful? > > > Ton, > > 1. Pardo disagrees with Aronson (and Bandy). Pardo suggests that a OOS to > IS ration of 25% - 35% is best, but that a good rule of thumb for empirical > testing is 1/8 to 1/3. > > 2. Yes, I suspect that each strategy will have its own best values for > IS/OOS and that other values will appear as useless. It is up to us to try > and find the best values. > > With respect to your comment: "I am getting results that show a random > pattern", my question remains; What are you measuring? In other words, what > values appear random - your fitness value? CAR? Something else? > > 3. I have done very much as you ask, except that I also varied my IS > period. I mostly kept my ratios within Pardo's suggested 1/8 to 1/3, but went > as low as 1/12 and as high as 1/2 just to be sure. > > For example IS=1 year, IS=2 years, IS=3 years giving > > IS1yr+OOS6mth, IS1yr+OOS3mth, IS1yr+OOS1mth > IS2yr+OOS12mth, IS2yr+OOS6mth, IS2yr+OOS3mth > IS3yr+OOS18mth, IS3yr+OOS12mth, IS3yr+OOS6mth > > IS2yr+OOS6mth produced the most consistent CAR, even though a weighted UPI > was used as the fitness function for the actual walk forward. > > I do not have a strong opinion as to whether or not there really is a > relationship between IS and OOS sizes. I found that Pardo's rule of thumb was > as good a starting place as any. I was happy that my values (25%) coincided > with what he advised. But, had my studies suggested a ratio outside of > Pardo's range, I would have still gone with what my results suggested, > despite Pardo's advice. > > Mike > > --- In [email protected], "Ton Sieverding" <ton.sieverding@> wrote: > > > > Hi Mike, > > > > What I am saying is : > > > > 1. That according to David Aronson "There is no theory that suggests what > fraction of the data should be assigned to training ( IS ) and testing ( OOS > )." and that "Results can be very sensitive to these choices ... ". I assume > that he knows where he is talking about ... > > > > 2. That when I am doing WalkFoward tests following the advice of Howard > Bandy, Robert Pardo AND Van Tharp, I am getting results that show a random > patron when changing the OOS en IS periods. So my conclusion is that > WalkFoward is a subjective test ... > > > > Therefore I have serious problems using WalkFoward tests. If you can help > me to get things done in an objective way then I will be delighted to know > how you want to do that. But for sure Van Tharp did not help me ... > > > > Please do a simple WF test with OOS=1year and IS=1month...12months. So > creating WF results for OOS1y+IS1m, OOS1y+IS2m etc. And see what you are > getting. This is purely random. The result says nothing to me ... > > > > Regards, Ton. > > > > > > > > ----- Original Message ----- > > From: Mike > > To: [email protected] > > Sent: Monday, October 05, 2009 9:29 AM > > Subject: [amibroker] Re: Is the Walk forward study useful? > > > > > > Ton, > > > > Are you saying that you have not found an IS/OOS pair that works well? > What measure are you using to judge "stability" of the walk forward process > (i.e. what measure are you using to judge the process as random)? > > > > After testing with multiple IS periods, and with multiple OOS periods, I > was able to identify "fixed" window lengths that proved more consistent than > the others tested. > > > > I reached this conclusion by charting a distribution curve of CAR for the > OOS results. My fitness function is currently based on UPI, and thus my walk > forward is driven by that value. However, ultimately my interest is in how > consistent CAR would be which is why I used that for evaluating the goodness > of fit for the IS/OOS period lengths. > > > > In my case, over a 13 year period, a 2 year IS and 6 month OOS (for a > total of 26 OOS data points) produced the most normal looking distribution of > CAR results (i.e. central peak, smallest standard deviation). Excluding the > results from all of 1999 and the first half of 2000 (during which results > were abnormally strong), the distribution curve looks even better. > > > > Also, have you tried working with different fitness functions? Perhaps > your fitness function doesn't adequately identify the "signal" and thus > misguides the walk forward, regardless of IS/OOS window lengths. > > > > I am in the process of running a new walk forward over the last 7.5 years > using Van Tharp's System Quality Number (SQN) as my fitness function. I have > kept the same 2 year IS/6 months OOS for a total of 15 OOS data points. My > system strives to generate a minimum average of 2 trades per day, so each IS > period generally has 1000 or more trades from which to calculate the fitness. > > > > It has not run to completion yet. But, for the periods that have produced > results, the results look promising (at least with respect to the SQN of the > OOS relative to the SQN of the IS, I have not yet created the distribution of > CAR for OOS). > > > > Assuming that the remainder of the results are equally strong, I will > walk forward further back in history to get the full 26 data points to > compare against the results produced using my UPI fitness. If the CAR > distribution is more normal using SQN as fitness, then I will officially > start using SQN for generating optimal values for my next live OOS. > > > > If you are willing to share, I would be curious to hear if SQN as a > fitness function was able to produce a more stable walk forward for you, and > what measure you are using to judge "stable". > > > > Mike > > > > --- In [email protected], "Ton Sieverding" <ton.sieverding@> > wrote: > > > > > > Hi Howard, > > > > > > I still am struggling with the following sentence from David Aronson : > "The decision about how to apportion the data between the IS and OOS subsets > is arbitrary. There is no theory that suggests what fraction of the data > should be assigned to training ( IS ) and testing ( OOS ). Results can be > very sensitive to these choices ... ". Because this is exactly what I am > seeing. WalkFoward results are more then sensitive to the IS/OOS relation and > in many cases a pure random story. I am getting more and more the feeling > that WalkForward is not the correct or better objective way to test trading > systems. With all respect to Robert Pardo's idea's about this topic and what > you are writing in QTS ... > > > > > > Regards, Ton. > > > > > > > > > ----- Original Message ----- > > > From: Howard B > > > To: [email protected] > > > Sent: Monday, October 05, 2009 12:48 AM > > > Subject: Re: [amibroker] Re: Is the Walk forward study useful? > > > > > > > > > Greetings all -- > > > > > > My point of view on the length of the in-sample and out-of-sample may > be a little different. > > > > > > The logic of the code has been designed to recognize some pattern or > characteristic of the data. The length of the in-sample period is however > long it takes to keep the model (the logic) in synchronization with the data. > There is no one answer to what that length is. When the pattern changes, the > model fits it less well. When the pattern changes significantly, the model > must be re-synchronized. The only person who can say whether the length is > correct or should be longer or shorter is the person running the tests. > > > > > > The length of the out-of-sample period is however long the model and > the data remain in sync. That must be some length of time beyond the > in-sample period in order to make profitable trades. It could be a long time, > in which case there is no need to modify the model at all during that period. > There is no general relationship between the length of the in-sample period > and the length of the out-of-sample period -- none. There is no general > relationship between the performance in-sample and the performance > out-of-sample. The greater the difference between the two, the better the > system has been fit to the data over the in-sample period. But that does not > necessarily mean that the out-of-sample results are less meaningful. > > > > > > You can perform some experiments to see what the best in-sample length > is. And then to see what the typical out-of-sample length is. Knowing these > two, set up a walk forward run using those lengths. After the run is over, > ignore the in-sample results. They have no value in estimating the future > performance of the system. It is the out-of-sample results that can give you > some idea of how the system might act when traded with real money. > > > > > > It is nice to have a lot of closed traded in the out-of-sample period, > but you can run statistics on as few as 5 or 6. Having fewer trades means > that it will be more difficult to achieve statistical significance. The > number 30 is not magic -- it is just conventional. > > > > > > I think it helps to distinguish between the in-sample and out-of-sample > periods this way -- in-sample is seeing how well the model can be made to fit > the older data, out-of-sample is seeing how well it might fit future data. > > > > > > Ignore the television ads where person after person exclaims > "backtesting!" as though that is the key to system development. It is not. > Backtesting by itself, without going on to walk forward testing, will give > the trading system developer the impression that the system is good. > In-sample results are always good. We do not stop fooling with the system > until they are good. But in-sample results have no value in predicting future > performance -- none. > > > > > > There are some general characteristics of trading systems that make > them easier to validate. Those begin with having a positive expectancy -- no > system can be profitable in the long term unless it has a positive > expectancy. Then going on to include trade frequently, hold a short time, > minimize losses. Of course, there have been profitable systems that trade > infrequently, hold a long time, and suffer deep drawdowns. It is much harder > to show that those were profitable because they were good rather than lucky. > > > > > > There is more information about in-sample, out-of-sample, walk forward > testing, statistical validation, objective functions, and so forth in my > book, "Quantitative Trading Systems." > > > http://www.quantitativetradingsystems.com/ > > > > > > Thanks for listening, > > > Howard > > > > > > > > > > > > On Sun, Oct 4, 2009 at 10:56 AM, Bisto <bistoman73@> wrote: > > > > > > > > > Yes, I believe that you should increase the IS period > > > > > > as general rule is not true "the shortest the best" trying to catch > every market change because it's possible that a too short IS period produces > a too low number of trades with no statistical robustness --> you will find > parameters that are more likely candidated to fail in OS > > > > > > try a longer IS period and let's see what will happen > > > > > > I read an interesting book on this issue: "The evaluation and > optimization of trading strategies" by Pardo. Maybe he repeated too much > times the same concepts nevertheless I liked it > > > > > > if anyone could suggest a better book about this issue it would be very > appreciated > > > > > > > > > > > > Bisto > > > > > > --- In [email protected], "Gonzaga" <gonzagags@> wrote: > > > > > > > > Oh, sorry, I am lost in translation ... ;-) > > > > Yes I meant trades of my IS period. > > > > I've got about 70 trades in my IS period, three months. > > > > BUT, I buy stocks in a multiposition way.This means, that my hole > capital divides among several stocks purchased simultaneously. > > > > So, in my statistics, I use to average my trades. When I use > maxopenpositions=7, I use to average my results every 7 trades. > > > > Considering that, my trades in three months are not 70, but less ( > not exactly 70/7, but less than 70) > > > > > > > > If I use maxopenposition=1, which is, invest all my capital every > trade, in three months I would have about 29 trades. > > > > So I suppose I have to increase the IS period.. isn`t it? > > > > > > > > > > > > --- In [email protected], "Bisto" <bistoman73@> wrote: > > > > > > > > > > What do you mean with "I don't have many buyings and sellings"? > > > > > > > > > > If you have less than 30 trades in an IS period, IMHO, you are > using a too short period due to not statistical robustness --> WFA is > misleading, try a longer IS period > > > > > > > > > > Bisto > > > > > > > > > > --- In [email protected], "Gonzaga" <gonzagags@> wrote: > > > > > > > > > > > > Thanks for the answers > > > > > > To Keith McCombs : > > > > > > > > > > > > I use 3 months IS test and 1 month step, this is, 1 month OS > test. My system is an end-of day-system, so I don't have many buyings and > sellings.. > > > > > > Perhaps I should make bigger the IS period? > > > > > > > > > > > > anyway, my parameter behaves well in any period. Of course it is > an optimized variable, but it doesn't fail in ten years, in none of those ten > years, over 500 stocks.. a very long period.. > > > > > > So, couldn't it be better, on the long run, than the parameters > optimized with the WF study? > > > > > > (In fact, I am using it now, the optimized variable) > > > > > > That's my real question.. > > > > > > > > > > > > To dloyer123: > > > > > > I haven't understood the meaning of the Walk Forward Efficency, > and seems interesting. > > > > > > can you explain it better, please..? > > > > > > > > > > > > > > > > > > > > > > > > --- In [email protected], "dloyer123" <dloyer123@> wrote: > > > > > > > > > > > > > > I have had similar experiences. I like to use WFT to estimate > what Pardo call's his "Walk Forward Efficency", or the ratio of the out of > sample WF profits to just optimizing over the entire time period. > > > > > > > > > > > > > > A good system should have as high a WFE as posible. Systems > with a poor WFE tend to do poorly in live trading. > > > > > > > > > > > > > > If you have a parm set that works well over a long period of > live trading, then you are doing well! > > > > > > > > > > > > > > > > > > > > > > > > > > > >
