I didn't want to bother the forum on this topic again, since there is no discussion forthcoming, but I have since found out that multidimensional matrices are real and a standard programming procedure.
As far as I can tell, they are mainly used for solving complex problems .... I believe computational time raises exponentially as more dimensions are added ... I think they would be overkill for AFL so I am now, more or less, satisfied that cutdown pseudo matrices would be fine for the job. This would amount to little more than a collection of 'callable' arrays, per system OR subsystem ... they would comprise a much smaller sample space than the data itself, as in most cases they would be relatively short vectors .... since array processing would still apply I assume that processing could still be very fast once they have been created.... so while there is a time penalty involved in their creation there is some time payback to justify the effort (the more often the users queries the matrices, to perform sample space analysis, the sooner the payback line is crossed). My only interest there now is to speed compare languages, and programming methods, for managing the 'matrix' arrays. I believe that in C++ tensors is one method used ... essentially this indexes a single array (a N == 1 multi-dimensional matrix) to produce the sub-matrices. Since this has a time penalty attached the way to go would be to allow users, who don't want to use signal ranking OR sample space analysis etc, to toggle off matrix indexing and simply collate the database trades as a single time series. If any laypeople are interested in getting their head around the subject here are some starter links (not great ones but they do provide some insight into what I am talking about and show how simple, and intuitive, the required AFl functions could be): http://en.wikipedia.org/wiki/Matrix_(mathematics) http://www.absoluteastronomy.com/topics/MATLAB In both the total trade series OR the subseries, the sequence of trades, as per the time of entry, would need to be preserved or restored (sort, within the matrices,by entry time) for those who want to replicate the 'real' trading environment. My questions now are mainly centred around fast ways to rip the trade series (either as a single array or referenced by symbol for matrix use). So far I have three areas to think about: a) programmatical - AB uses trade lists, which must have a computational cost... I assume this is the current optimal method ... looking around to see what other programmers and other languages are capable of ... gains here are likely to be relatively small except for the fact that I am getting prompted, by my subconscious, that something else is out there or waiting to be discovered. b) multicore processing - a quantum leap ... no self-respecting backtester would be without it c) follow DLoyers lead into GPU processing to do the ripping (Very neat DL ... you might be the one who finally turned me into a programmer after Tomasz kickstarted the job). While on the subject of MCP ... I came across this article ... doesn't have anything to do with the current topic but since I have previously posted my 'idea' of giving AB a software 'brain' to manage processing load distribution, for MCP machines (particularly considering that AB will ecounter a large variety of MC configs) I thought I would post it as some sort of vindication ... I always live in hope that I am not nuts afterall ;_) If you are a mad philospher it always helps to hang out with a group of crazies and then you won't be so conspicuous! http://www.mc.com/uploadedFiles/MCF-API-Conf-Paper.pdf --- In [email protected], "brian_z111" <brian_z...@...> wrote: > > > I have been experiencing repeated thougnts about 64 bit > >programming, > > using tricky methods to store trade data, with minimal memory use, > > and 3,4, or more, dimensional matrices > > Does anyone know where I can find something about minimizing memory > use with data handling and/or programming multidimensional matrices > Or are we still waiting for someone to invent them? > > Where do the crazy programmers hang out and talk about this stuff? > > Which language is the best for writing a backtester in? > > > --- In [email protected], "brian_z111" <brian_z111@> wrote: > > > > IMO backtesters that combine money management, scaling in and > signal > > ranking in one step are not using the best, or most efficient > method. > > > > On the contrary, this approach can muddy the waters and create > > confusion for the user, especially new comers. > > > > It is well known that if piano students do not start by learning > the > > correct, albeit funNOT techniques, they will almost certainly never > > learn to play at advanced levels.... the end justifies the means > (and > > the high attrition rate). > > > > Collecting all matched buy/sells first is the best approach (as > used > > by RalphVince and MSA Analyzer) .... sequential or random testing > of > > MM, scaling in, trade sample space analysis and ranking etc can be > > performed, as per user requirements, from the trade matrix, which > > afterall is a constant for any constant set of rules and data. > > > > For anyone who is thinking of attempting it privately: > > > > - the system rules i.e. buy/sell/prices/stops etc (not MM rules) > need > > to be saved with the trade matrices > > > > - when a system is reopened from the historical archive the rules > > need to be auto restored to the backtester (exactly as per the > > original test) > > > > - versioning would need to be introduced ... if the user adds some > > code to the the rules then they need to elect to run it as a new > sub- > > version (save the matrix/rules separately) OR as a new test that > > overwrites the archived trade matrix. > > > > - for optimization a user election would save all matrices and > rules > > as sub-versions > > > > - scaling in should be treated as a new and different sub- > system .... > > > > system 1 = conditional on the set of rules A > > scaling in = system 1.1 = conditional on A then apply rule B > > > > .... different combinations of MM rules can then be tested on > system > > 1 + system 1.1 versus system 1 OR system 1.1. to find where the > > optimum returns are. > > > > - ranking signals is a challenge ... my first thoughts are that the > > ranking indicator would need to be collected with the trade signals > > and saved with them. > > > > The careful reader might have noted that this discussion is an > > extension of the topic SYSTEM PERFORMANCE INDICATORS where I came > up > > against the (wall) limits imposed by computation and AFL ..... > > ripping the mimimum data (trades only) required for performance > > metrics i.e. core metrics, process only the essential metric > > (objective function) and doing it on the fly (cached) for a > reasonale > > number of symbols is the challenge. > > > > > > Currently obtaining all matched signals (trades) in AFL isn't the > > default and perhaps a little difficult to achieve. > > > > Finding a fast efficient way to do this is the key to the challenge. > > > > From what I can find out from the docs the AB method is to save > > trades in a trade list .... I assume this has a time penalty > attached > > but in programming alternatives seem few and far between ??? > > > > My interest now is the relative speed OR otherwise of different > > langauges/matrix/array processing when ripping the trade series > from > > the data followed by the subsequent speed when calculating the > > metrics (the bias is on getting the trade matrix though). > > > > I have been experiencing repeated thougnts about 64 bit > programming, > > using tricky methods to store trade data, with minimal memory use, > > and 3,4, or more, dimensional matrices ... how crazy is that ... I > > don't even own a 64 bit compputer and barely know the first thing > > about computers or programming? > > > > I don't find AFL a good language to learn programming from (for > first > > timers like myself) possibly because it has been simplified and > also > > because it doesn't have an intimate relationship with the computer, > > being a sort of second order C++. > > > > For programming research and education I am looking at other > > languages and open source projects ... so far R Math is #1 on my > > short list. It only has a small amount of code written for trading > > but there might be some potential there. > > > > Plus there is a handful of AB people interested in R and we have > the > > plugin (thanks to Patrick). > > At least it gives me a chance to learn something about wider > > programming techniques, including lists and matrices, (AFL is > > reasonably transparent and documented but R language is much richer > > and completely transparent and owned by the user which makes for a > > better learning environment). > > > > What I find out there I could apply to AB via a plugin but that > would > > mean I would end up only using AB as a data downloader and charting > > package (more craziness)! > > > > I am not likely to publish anything for a while, if at all, unless > I > > get real excited about it.... one can always rely on the 'good ol > > workaround'! > > > > Lucky for AB I don't pay a programmer to write the AB plugin, > > commercialise it and copyright the code ;-) > > > > > > --- In [email protected], "brian_z111" <brian_z111@> wrote: > > > > > > I forgot to mention that another prime objective, of the overall > > > method, is to make efficient use of our precious data. > > > > > > All trades are recording in Phase I, so that our elected N value > (a > > > statistically valid sample)is approached using a minimum amount > of > > > data...... for WF analysis N could be the cut off point for the > > > sample space used. > > > > > > BTW this one is for the programmers, and serious dudes, who for > the > > > most part, tolerate my biblings and scribblings (hope you like it > > > Fred).... Tomasz and this forum must have good Karma! > > > > > > Also for K and T who stuck up for me when I was pelted with > rotten > > > fruit all those years ago! > > > > > > --- In [email protected], "brian_z111" <brian_z111@> > wrote: > > > > > > > > Thinking a little further on this. > > > > > > > > Unless there is a need for advanced matrix operations 'psuedo > > > > matrices', using arrays with limited AFL matrix operations, > would > > > be > > > > sufficient for AB needs (arrays are faster and maintain > backward > > > > compatibility?). > > > > > > > > For advanced users the first phase would be the power version > and > > > the > > > > method used would only store the core metrics with speed in > mind > > > (2, > > > > 3 or maybe 4 datapoints, as persistent arrays, depending on the > > > > objectives) and only calculate one metric, on the fly, from > them > > > i.e. > > > > the user defined objective function. > > > > > > > > All performance metrics would be derived from the core metrics > > and > > > > therefore able to be calculated in AFL so they would be fully > > > > transparent to users. > > > > > > > > Default metrics could be built-in and rendered as tables etc, > at > > > > phase II, for the benefit of newcomers, but as their skill and > > > > interest develops they could then easily drill down to the code > > > > behind the metric. > > > > > > > > Obviously the trade matrices are a precursor to advanced > > Portfolio > > > > Management in Phase III (RalphVince, MSA Analyser, van Tharp, > > > > McDonnell, MonteCarlo Simulation or whatever takes the users > > fancy). > > > > > > > > For my own use I intend to go right past phase II and implement > > my > > > > own Portfolio Manager (based on a compilation of my own work > and > > > the > > > > work of people like RalphVince .... I am still working in the > > > > background on BiNomial Simulation and if I am happy with it I > > will > > > > use it in Phase III). > > > > > > > > Part of the rational, behind the idea, is that the time value > > > > (processing) in the backtesting is saved, while the time value > in > > > > producing performance metrics is disposable (because the latter > > has > > > a > > > > lesser timevalue, the metrics of interest are based on > subjective > > > > choices and they change dynamically as new data arrives). > > > > > > > > I think it is a good idea but I am not 100% certain about it, > not > > > > being a mathematician or programmer. > > > > > > > > I am becoming a little frustrated with AB (perhaps I am > > outgrowing > > > > it) and the Portfolio Backtester is a good example of one place > > > where > > > > that frustration is focussed. > > > > > > > > I have three ways I can go with it: > > > > > > > > - send it to AB as a suggestion (which is why I need feedback) > > > > - implement my own backtester within AB (use a DLL?) ... maybe > > keep > > > > it private or publish it with open source code (possibly I > would > > do > > > > that via a team effort) > > > > - go to an existing open source charting project with my ideas > > > > - intitiate my own open source charting project > > > > > > > > Since I am biased to open source, and therefore my ideas/code > > would > > > > be in the public domain anyway, I have no reason not to share > it > > > with > > > > AB and the AB community. > > > > > > > > If it is not a good idea then nothing is lost. > > > > > > > > > > > > --- In [email protected], "brian_z111" <brian_z111@> > > wrote: > > > > > > > > > > There was some recent discussion about the need to learn more > > > than > > > > one > > > > > language to use AB. > > > > > > > > > > Just wondering if C++ Matrices could be used to store and > > > reference > > > > > trades, instead of Object Orientated Progamming, to make > things > > > > easier. > > > > > > > > > > C++ includes containers that can be considered matrices (such > > as > > > > arrays > > > > > and, in the Standard Library, vectors, lists, and maps) OR > open > > > > source > > > > > C++ Matrix libraries are available. > > > > > > > > > > > > > > > Would it be possible to: > > > > > > > > > > - nominate buy/sell/buyprice etc in the first phase of > > backtesting > > > > > - run backtest over the sample space (limit space as defined > in > > > > setting) > > > > > - store trades P & L as % in binary files > > > > > - store time in trade (binary) > > > > > - store entry bar OR time (binary- only needed to display > > arrows > > > on > > > > > chart) > > > > > > > > > > Storing trades in binary form would be very fast (stored as a > > > trade > > > > > matrix). > > > > > All metrics can be calculated from % P & L plus time in trade > > as > > > > > required so Money Management only needs to be applied in the > > > second > > > > > phase, if required. > > > > > Only the trade matrices need to be saved, with the backtest > > > > settings > > > > > (also binary), for future reference, instead of the BT > reports. > > > > > > > > > > In the second phase performance metrics could be calculated > as > > > > required > > > > > by referencing the matrix entries, either in total or as sub > > > > matrices. > > > > > > > > > > Matrix functions could be included in AFL and the advanced > > > > backtester > > > > > would use them to reference the trades to create equity > curves > > > and > > > > > custom or default reports on demand (no need to re-run > > backtests > > > > for > > > > > custom reports OR when adding/deleting columns from a report). > > > > > > > > > > It might also be possible that trade P & L, and time in > trade, > > > > could be > > > > > stored in one 32 bit byte thereby save processing time (the > > > > separate > > > > > pieces of info could be unpacked into the trade matrix and > the > > > time > > > > in > > > > > trade matrix when required). > > > > > > > > > > Happily, the trade matrices could be rendered as tables, with > > > rows > > > > and > > > > > columns for visual reference/charting etc, which is much more > > > > intuitive > > > > > than referencing the Trade Object as we do at the moment. > > > > > > > > > > The matrices could use a full range of Matrix operations or > > just > > > > some > > > > > of the basics for the AFL version e.g. referencing individual > > > > trades by > > > > > matrix M * N notataion (I don't know much about matrix maths > so > > > it > > > > is > > > > > up to the maths people to comment on possible uses for > advanced > > > > matrix > > > > > maths functions). > > > > > > > > > > > > > > > When data is added to the DB, or the sample range extended, > > > > backtesting > > > > > could commence from the time of the last closed trade and > then > > > add > > > > new > > > > > trades to the matrix (this would save time by avoiding the > need > > > to > > > > > repeat backtests). > > > > > > > > > > It is also a very nice format for exporting to Excel or > > advanced > > > > stats > > > > > packages like R or Matlab :-) > > > > > > > > > > Does anyone esle have any thoughts on the subject? > > > > > > > > > > > > > > >
