Dear Keith, Your proposal for self-adaptation is certainly interesting. However, this is a very high goal.. Me personally, I am happy if I have a strategy which works most of the time for me in most of the circumstances. After all, even with self-adaptation you will have the problem it needs time. (It must look at a trace of trades and only after some time correction is possible.) It might easily be that at this point the situation has already (again) changed..
Klaus On 6 Jun., 01:04, Keith <[email protected]> wrote: > Dear Klaus, > > Many thanks for the links, I will check them up. > > I assume when you say it is supervised, it is the need to optimize by > back testing and forward testing with a good sample size. > > Is it possible to create an unsupervised, self optimizing system that > is able to self-diagnose itself and know when it needs to re-optimize > itself? > > Just thinking out loud. In real time, one thread would be trading the > current optimized version of the strategy. Another thread will be > monitoring the performance. A third thread could be using a shorter > window (say 10 minutes) for backward and forward testing. A fourth > thread would be making decision on whether it should stop the current > version and switch to the freshly optimized version. > > Many thanks for your generous and patient coaching of a newbie. > > Keith > > On Jun 5, 12:53 pm, Klaus <[email protected]> wrote: > > > > > Dear Keith, > > > by the way: a good quality indicator is the heat map (optimization > > map). > > If there is only a very small red area, well surrounded by blue. You > > immediately > > know that the approach did most accidentally struck gold. This will > > not generalize to other cases.. > > > Klaus > > > On 5 Jun., 21:09, Klaus <[email protected]> wrote: > > > > Dear Keith, > > > > which kind of machine learning it is, is actually a more complex > > > question than one would expect. > > > Because the typical approach which is used by JBT (and similar tools) > > > falls a bit through the cracks.. > > > It is roughly a time series prediction problem. - Albeit we do not > > > want to predict the next value, we only > > > want to predict the kind of trade to make. Thus, it can also be seen > > > as a classification problem. > > > If you look at categorizations like in en.wikipedia.org/wiki/ > > > Machine_learning > > > it is supervised learning in a way which is most closely to decision > > > trees (though the tree nodes are > > > parameter values - the ones your optimizer give you). > > > The difficulty are the indicators.. They can be seen as features which > > > are used as the basis for classification > > > by the decision trees. > > > > There are tons of ML approaches, but you / we will probably not > > > implement others, as they fit not well to > > > the kind of JBT-approaches. > > > However, the lessons from testing apply here. You may find it > > > worthwile to look > > > athttp://en.wikipedia.org/wiki/Cross-validation_(statistics) > > > and other literature in the same direction. > > > > Cheers > > > Klaus > > > > On 5 Jun., 20:32, Keith <[email protected]> wrote: > > > > > Dear Klaus, > > > > > I am very grateful to you for your indulgence in my ramblings and > > > > thanks again for the rules, I will keep them in mind as I explore JBT. > > > > > Drawdowns are a part of the reality of trading, so I have no problem > > > > with that. I remember witnessing 2 equity curves for the top 2 > > > > competitors in a forex contest. The first trader has small but very > > > > consistent winning trades - very little drawndowns. The second trader > > > > has much fewer trades but has great drawdowns and also greater > > > > profits. > > > > > My personal preference is to emulate the first trader, consistent but > > > > smaller profits, with little drawdowns, with that slight edge - that > > > > is the system I want to build. > > > > > What category of machine learning would you classify JBT's method so I > > > > can google them up. > > > > > Thanks. > > > > > Keith > > > > > On Jun 5, 11:09 am, Klaus <[email protected]> wrote: > > > > > > Dear Keith, > > > > > > regarding the overoptimization problem: > > > > > - rule one: don't have to many parameters. (the more parameters, the > > > > > closer you can fit, but the > > > > > less general your rules will become) - it is a similar difference > > > > > between understanding and > > > > > learning word by word. You want the system to generalize > > > > > fundamentals not to memorize the > > > > > training set.. > > > > > - rule two: in order to check: separate between training set and > > > > > test. > > > > > Be sure that both cover very different market situations. > > > > > (e.g., things like the flash crash recently, of course there one > > > > > only, thus it is better to have it in the > > > > > test set) > > > > > => see cross-validation if you want to go more in depth > > > > > - rule three: before you start try to understand what an acceptable > > > > > system might be for you: > > > > > if you dream up the system that has no drawdown, only wins. There > > > > > might be no such system, then > > > > > stop wasting your time :) > > > > > > Regarding machine learning: there is tons of knowledge out there. > > > > > Beyond the primitive rules, given above. > > > > > I am not a guru on it, so I cannot condense it shortly. If you want to > > > > > better understand what you are doing, > > > > > there is a lot of it free on the internet, but also good textbooks. > > > > > A special branch focus on time series prediction (this is what applies > > > > > here), and also many reports > > > > > use some sort of a market data for this.. > > > > > > Cheers > > > > > Klaus > > > > > > On 5 Jun., 19:09, Keith <[email protected]> wrote: > > > > > > > Dear Klaus, > > > > > > > Thank you so much for sharing your experience. > > > > > > > I agree completely with your sentiments especially on money > > > > > > management > > > > > > and control - black swan events do exist and in these very uncertain > > > > > > and bipolar markets, the danger is ever more present. > > > > > > > Based on a short exploration of the code, each strategy is a > > > > > > variation > > > > > > of volality, velocity and the volume impact of the bid and ask > > > > > > prices > > > > > > and the relation to the crossover of a fast and slow indicator. > > > > > > These > > > > > > parameters are optimized through back and forward testing over some > > > > > > time period, which is another parameter which may need to be > > > > > > optimized. Many gurus warn about over-optimization - so how do we > > > > > > know if we have the right optimized parameters? It was mentioned > > > > > > that > > > > > > we may need to optimise the parameters after a time - how and when > > > > > > shall we know that? If this is the wrong thread to discuss this, let > > > > > > me know where to take this because this is very interesting to me. > > > > > > > I appreciate that having a stop loss in the back tested strategies > > > > > > reduces its performance. Were the stop loss points fixed or also > > > > > > optimized. If a market is very volatile, a closer fixed exit will > > > > > > have > > > > > > a higher chance of being hit be it a profit target or a stop loss. > > > > > > Most gurus recommend a closer profit target to a stop loss for > > > > > > futures. > > > > > > > Warren Buffet said once in an interview that he does not trade > > > > > > anything he does not understand. Since you are familiar with machine > > > > > > learning, could you share with us your views on how these strategies > > > > > > function and how do they adapt to changing market behaviours? > > > > > > > Thanks. > > > > > > > Keith > > > > > > > On Jun 5, 7:46 am, Klaus <[email protected]> wrote: > > > > > > > > Dear Keith, > > > > > > > > as you asked me, also my 2c (though nonlinear has MUCH more > > > > > > > experience > > > > > > > on this, so if in doubt take his answers). > > > > > > > > My status is as this: for some time I am following here and played > > > > > > > around.. > > > > > > > For a few weeks now, I am collecting data for my own experiments. > > > > > > > And only recently I became serious about designing strategies, > > > > > > > though > > > > > > > this is > > > > > > > purely a hobby. I have some background on machine learning, > > > > > > > though. > > > > > > > Thus, > > > > > > > all this is not foreign to me. > > > > > > > > At this point: most of the time the system is in forward testing > > > > > > > mode, > > > > > > > for pure recording > > > > > > > (I still want to extend my data basis). Last week, I traded it > > > > > > > live > > > > > > > only on friday, as I had > > > > > > > only then time to monitor the system. Currently I do not yet have > > > > > > > too > > > > > > > much trust > > > > > > > in the strategies (because you can never be sure how much your > > > > > > > tested > > > > > > > strategies > > > > > > > extrapolate to unseen data). > > > > > > > This said, I leave the computer also alone for half an hour or > > > > > > > so, but > > > > > > > at this stage not much longer. > > > > > > > In the long term, I hope to increase the time frame, but I am not > > > > > > > sure, whether I will eventually > > > > > > > (can) fully automatic. Besides the quality of the strategies, > > > > > > > there is > > > > > > > also the question regarding > > > > > > > stability of the platform (and connection, etc.). For example, > > > > > > > last > > > > > > > week while I was forward testing, > > > > > > > I suddenly found TWS dead. In such a situation you are stuck with > > > > > > > whatever position you had before, as > > > > > > > JTB can no longer change it. > > > > > > > If you want to go fully automatic, you must be able to accept this > > > > > > > financially... > > > > > > > > Regarding stop-losses, profit targets. I use them when I trade > > > > > > > manually - why? Because this is a way I > > > > > > > can get basic automatization from TWS. > > > > > > > However, any experiments in autotrading are exactly in order to > > > > > > > get > > > > > > > smarter than this! > > > > > > > My personal take on these are: they are very simple approaches to > > > > > > > ensure to get out of the market. > > > > > > > However, any good strategy (in my eyes) defines exactly when to > > > > > > > enter > > > > > > > AND when to leave the market. > > > > > > > Because, why should I be very sophisticated wrt whent to enter the > > > > > > > market and extremely simplistic > > > > > > > regarding leaving it? Both contribute in equal manner to the total > > > > > > > system performance. > > > > > > > If you approach strategy design in this way, that you carefully > > > > > > > design > > ... > > Erfahren Sie mehr » -- You received this message because you are subscribed to the Google Groups "JBookTrader" group. To post to this group, send email to [email protected]. To unsubscribe from this group, send email to [email protected]. For more options, visit this group at http://groups.google.com/group/jbooktrader?hl=en.
