I can see there have been a lot of discussions, mostly centred around the processes and/or the methodology of optimization. While these are good discussions, and with many points of views. I wonder if there are any quantitive research being done to verify the different points of view. I have consumed thousands of hours my own time plus even more computer processing time in system development, and have made a few interesting observations. However, these are not rigourous enough to quantify my view. However, these are interestingly enough observation to share.
1. Does sensitivity analysis provide similar consistency as OOS guidance walk forward analysis. I use that term now for the sake of continuity in the discussion. Sensitivity merely takes a different path to the same data set that you optimize on. But OOS takes a completely data set. My own observation is that I definitely need the guidance of OOS. They dont do the same thing. 2. How does the definition of fitness affect the result of the optimized system in terms of robustness, performance and stability through time and with different markets? As the old saying goes, the answer is only important if the right question is asked. Defining a fitness criteria is like framing a question. The answer comes in the form of result and will be different of course depending the fitness criteria. Even for a single goal inside one's mind, it can be expressed in term of different fitness criteria. How is that different intrepretation going to affect us in terms of equity curves comparsions. Eg CAR/MDD vs UPI vs some of the more complicated variety? 3. Related to above, How about instead of defining fitness as a single number, we define fitness as a time series. E.g, UPI as a series of UPI every six month. We now have a distribution of UPI. How does different distributions affect our In sample optimization. It is like doing rolling optimization, and collating the results. 4. How does artifical division of data set affect our results? For example, If I optimized a system based on all the stocks on the ASX exchange, including both current and delisted stocks, How would it perform if I run it on ASX100 (The top 100 Australian company as defined by S&P) 5. What is the real difference between Guiding the optimization with OOS vs merly validating our optimization through OOS in terms of results? Is it really better to skip Guidance and go straight to Validation, or in my case, skip Validation and stick with Guidance. Food for thought, I'll be trying to answer some of them myself. Cheers Paul. --- In amibroker@yahoogroups.com, "Howard B" <[EMAIL PROTECTED]> wrote: > > Hi Brian -- > > I tend to agree with Fred. I, personally, do not use the guidance data > set. If you want to use it, and you are looking for consistency between the > two data sets, that might be valuable. But another measure of that is > simply the equity curve and other performance stats over both periods. > > Another approach is to look at the robustness of the system by perturbing > each of the perturbable variables (not all of them are), computing the > scores for nearby points and rewarding "plateaus" in preference to "peaks." > > Thanks, > Howard > > > On Thu, May 8, 2008 at 7:38 PM, Fred Tonetti <[EMAIL PROTECTED]> wrote: > > > Personally I couldn't find any value in the guidance phase which I > > allowed for in IO for a couple of years and have since removed the > > capability. > > > > > > ------------------------------ > > > > *From:* amibroker@yahoogroups.com [mailto:[EMAIL PROTECTED] *On > > Behalf Of *brian_z111 > > *Sent:* Thursday, May 08, 2008 8:32 PM > > *To:* amibroker@yahoogroups.com > > > > *Subject:* [amibroker] Re: Fitness Criteria that incorporates Walk Forward > > Result > > > > > > > > Howard, > > > > Thanks for a very nice summary of the framework. > > > > I would say that, since the training search is exhaustive (therefore > > we must have identified all possible candidates for the strategy) the > > best we can hope for, in the guidance phase, is to change our choice > > of top model to one or another of the 'training top models', or > > abandon the strategy altogether. > > > > Also I wonder, if the training model/guidance model combination, that > > passes a minimum requirement in both phases, and shows less variance > > between the training and guidance results, is the most generic model > > of them all i.e. suited to a wider range of conditions but not > > necessarily returning the highest possible result in any particular > > market? > > > > brian_z > > > > --- In amibroker@yahoogroups.com <amibroker% 40yahoogroups.com>, "Howard B" > > <howardbandy@> wrote: > > > > > > Greetings all -- > > > > > > I am coming to this discussion a little late. I just returned from > > giving a > > > talk at the NAAIM conference in Irvine. Some of the discussions I > > had with > > > conference attendees was exactly the topic of this thread. > > > > > > If you are using some data and results to guide the selection of > > logic and > > > parameter values (as described in the earliest postings as OOS > > data), that > > > incorporates that data into the In-Sample data set. In this case, > > there > > > must be three data sets. They go by various names -- Training, > > Guiding, and > > > Validation will be adequate for now. > > > > > > Optimization, by itself, begins by generating a lot of alternatives. > > > Optimization with selection of the "best" alternatives means using > > an > > > objective function (or fitness function) to assign a score to each > > > alternative. > > > > > > The method of searching for good trading systems used in AmiBroker's > > > automated walk forward procedure uses a series of: search over an > > in-sample > > > period, select the best using the score, test over the out-of- sample > > > period. Use the concatenated results from the out-of-sample > > periods to > > > decide whether to trade the system or not. > > > > > > Another method of searching for good systems (that might be what > > some of the > > > posters to this thread were suggesting) is to perform extensive > > searches of > > > the data and manipulations of the logic using the Training data, > > then > > > evaluate using the Guiding data. Repeat this process as desired or > > required > > > as long as the results using the Guiding data continue to improve. > > When > > > they show signs of having peaked, roll back to the system that > > produced the > > > best result up to that point. Then make one evaluation using the > > Validation > > > data. Now, step forward in time and repeat the process. It is now > > the > > > concatenated results of the Validation data sets that are used to > > decide > > > whether to trade the system or not. > > > > > > Thanks, > > > Howard > > > > > > On Thu, May 8, 2008 at 9:24 AM, Edward Pottasch <empottasch@> > > > wrote: > > > > > > > thanks. Will have a look, > > > > > > > > Ed > > > > > > > > > > > > > > > > ----- Original Message ----- > > > > *From:* Fred <ftonetti@> > > > > *To:* amibroker@yahoogroups.com <amibroker%40yahoogroups.com> > > > > *Sent:* Thursday, May 08, 2008 5:42 PM > > > > *Subject:* [amibroker] Re: Fitness Criteria that incorporates > > Walk Forward > > > > Result > > > > > > > > There's a simple example of this in the UKB under Intelligent > > > > Optimization ... > > > > > > > > --- In amibroker@yahoogroups.com <amibroker% 40yahoogroups.com>, > > "Edward Pottasch" <empottasch@> > > > > wrote: > > > > > > > > > > hi, > > > > > > > > > > "While optimization can be employed to search for a good system > > via > > > > > methods utilizing automated rule creation, selection and > > > > combination > > > > > or generic pattern recognition" > > > > > > > > > > anyone care to explain how this works? Some kind of inversion > > > > technique? Here is what I want now give me the rules to get > > there :) > > > > > > > > > > thanks, > > > > > > > > > > Ed > > > > > > > > > > > > > > > > > > > > ----- Original Message ----- > > > > > From: Fred > > > > > To: amibroker@yahoogroups.com <amibroker% 40yahoogroups.com><amibroker% > > 40yahoogroups.com> > > > > > Sent: Thursday, May 08, 2008 2:37 PM > > > > > Subject: [amibroker] Re: Fitness Criteria that incorporates Walk > > > > Forward Result > > > > > > > > > > > > > > > While optimization can be employed to search for a good system > > > > via > > > > > methods utilizing automated rule creation, selection and > > > > combination > > > > > or generic pattern recognition most people typically use > > > > optimization > > > > > to search for a good set of parameter values. The success of the > > > > > latter of course assumes one has a good rule set i.e. system to > > > > begin > > > > > with. > > > > > > > > > > As far as your prediction is concerned ... I suspect there are > > > > lots > > > > > of people, some of who post here, who could demonstrate > > otherwise > > > > if > > > > > they chose to ... > > > > > > > > > > --- In amibroker@yahoogroups.com <amibroker% 40yahoogroups.com><amibroker% > > 40yahoogroups.com>, > > > > "brian_z111" <brian_z111@> > > > > wrote: > > > > > > > > > > > > "IS metrics are always good because we keep optimizing until > > > > they > > > > > > are" (or words to that effect by HB) which is true. > > > > > > > > > > > > It is not until we submit the system to an unknown sample, > > > > either > > > > > an > > > > > > OOS test, paper or live trading that we validate the system. > > > > > > > > > > > > Discussing your points: > > > > > > > > > > > > IMO we are talking about two different trading approaches, or > > > > > styles > > > > > > (there is no reason we can't understand both very well). > > > > > > > > > > > > One is the search for a good system, via optimization, with > > the > > > > > > attendant subsequent tuning of the system to match a changing > > > > > market. > > > > > > > > > > > > If I understand Howard correctly he is an exponent of this > > > > style. > > > > > > > > > > > > It is my prediction that where we are optimising, using > > > > lookback > > > > > > periods, that the max possible PA% return will be around 30, > > > > maybe > > > > > > 40, for EOD trading. > > > > > > > > > > > > Do we ever optimise anything other than indicators with > > > > lookback > > > > > > periods? > > > > > > If so that might be a different story. > > > > > > > > > > > > Bastardising Marshall McCluhans famous line I could say "the > > > > > > optimization is the method". > > > > > > > > > > > > It is also possible to conceptually optimize the system, > > before > > > > > > testing, to the point that little, or no, optimization is > > > > required > > > > > > (experienced traders with a certain disposition do this quite > > > > > > comfortably but it doesn't suit the inexperienced and/or those > > > > who > > > > > > don't have the temperament for it). > > > > > > > > > > > > So, if a system has a sound reason to exist, and it is not > > > > > optimized > > > > > > at all, and it has a statistically valid IS test then it his > > > > highly > > > > > > likely to be a robust system, especially if it is robust > > across > > > > a > > > > > > range of stocks/instruments. > > > > > > The chances that this is due to pure luck are probably longer > > > > than > > > > > > the chance that an optimized IS test, with a confirming OOS > > > > test, > > > > > is > > > > > > also a chance event. > > > > > > > > > > > > However, if I had plenty of data e.g. I was an intraday > > trader, > > > > > then > > > > > > I would go ahead and do an OOS test anyway (since the cost is > > > > > > negligible) > > > > > > > > > > > > Re testing on several stocks. > > > > > > > > > > > > If the system is 'good' on one symbol, (the sample size is > > > > valid) > > > > > and > > > > > > it is also good on a second symbol (also with a valid sample > > > > size) > > > > > is > > > > > > that any different from performing an IS and an OOS test? > > > > > > > > > > > > For stock trading, I call the relative performance, on a set > > of > > > > > > symbols, 'vertical' testing as compared to 'horizontal' > > testing > > > > > > (where horizontal testing is an equity curve). > > > > > > > > > > > > Yes, if an IS test, with no optimization, beat the buy & hold > > > > on > > > > > > every occasion (or a significant number of times) in a > > vertical > > > > > test > > > > > > and the sum of that test was statistically valid and the > > > > horizontal > > > > > > test (the combined equity curve) was 'good' it would give you > > > > > > something to think about for sure. > > > > > > If some of the symbols, in the vertical stack, had contrary > > > > > returns, > > > > > > compared to the bias of my system, I probably would start to > > > > get a > > > > > > little excited. > > > > > > > > > > > > (I think perhaps you were alluding to something along those > > > > lines). > > > > > > > > > > > > BTW did you know that the Singapore Slingers play in the > > > > Australian > > > > > > basketball league? > > > > > > > > > > > > Cheers, > > > > > > > > > > > > brian_z > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >
