Machine learning in trading: theory, models, practice and algo-trading - page 965

 
What predictors can you come up with with real trading volume?
 

I would like to present the results of preliminary testing. I decided to present them as a two-dimensional table without unnecessary reports, graphs, only pure profits for better clarity. At this stage standard technical indicators and a couple of interesting ratios based on iHigh(...,iHighest()) and iLow(...,iLowest()) were used as predictors (inputs). Tomorrow I want to try to use more exotic data types, for example, indicators of linear regression or channel. If anyone has any suggestions for indicators, I can check them.


Files:
Bigtest.zip  16 kb
 
Ilya Antipin:

I would like to present the results of preliminary testing. I decided to present them as a two-dimensional table without unnecessary reports, graphs, only pure profits for better clarity. At this stage standard technical indicators and a couple of interesting ratios based on iHigh(...,iHighest()) and iLow(...,iLowest()) were used as predictors (inputs). Tomorrow I want to try to use more exotic data types, for example, indicators of linear regression or channel. If anyone has any suggestions for indicators, I can check them.


The material you presented can be improved by:

1. Creating an input file over 5000 ( at least)

2. Dividing this file into two parts: 4000 и 1000.

3. Getting a result of 4000, which is also divided into parts of 70%-15%-15%. We teach it to 70%, preferably with crossvalidation.

The trained model should be used on file 1000.

5. Plot the model's performance on all four parts.

6. Plot the graph of error depending on the number of trees - you can judge about the maximum number of patterns (trees) on your predictors.


The most significant drawback of your material is the lack of considerations about the predictive ability of your predictors. There are graphs in rattle that allow you to judge this. It makes no sense to feed garbage to the model input that has no relevance to the target variable.


PS.

If you were to use rattle in your endeavor, it would EVERYTHING improve the results of your materials. At least you would be able to compare 6 models with each other.

 
Aleksey Vyazmikin:
What predictors can be built with the real trading volume?

If there is a window in your basic strategy, and if it is also dynamic. Alternatively, you can measure the difference in volume between the beginning of the window and the signal. The same thing with the delta. If there is no window as such, I recommend to use Accumulation/Distribution and take its difference, but for how much data, I won't even admit.

In general, having information on the Volume and the delta of 11 symbols, I have managed to invent about 177 predictors. Stochastic based on real volume and delta shows itself very well. One or another stochastic is almost always involved in every model.... IMHO!!!!

 

In general, I believe that manipulation with the data should be minimal. The first thing you can take is the difference, preferably a dynamic window, when each signal has its own number of bars to analyze. Thus, each signal becomes unique. Naturally, it is better to take the difference of the indicators based on the real volume and delta.

In general, the indicators that are applicable to the real data of volume and delta, I use AD because it associates the volume with the price and we consider all of the bars in the taken window when obtaining the difference between the bars. In contrast to the simple difference between the first bar and the final bar of the window. In this case, the bars between the two bars are not taken into account. Although I have such a predictor and it can be important in some places.

Naturally, I build the AD and CumDelta stochastic function.

As well as the well-proven standard deviation of the delta and volume. StDev, so called.

That's basically all. Then I arrange the data in a circle, roughly as it is done in rattle. Well, and in general happy. I've got the latest model working for one day in a row and out of 23 trades only 6 are losing and the losses are not too big, besides the chart speaks for itself. So I think you should listen to my advice ....

And this is all OOC from 24.05.2018


I would like to add, if the basic strategy has no concept of the window, then you can take the difference between the first and second bar (the current trend) and, say, between the first and tenth (it's a global trend) then the number of predictors you have immediately doubled, because for one signal you take two differences. You can take three or four and the difference will be between a fixed number of bars, which is not as cool as the difference between the floating window for each signal.

Again, in the absence of a dynamic window in the basic strategy you can bound to any indicator, as an example. Stochastic/10. At the appearance of the signal, we look at stochastics, divide its value by ten, and take the integer part, we obtain a window range from 1 to 10 bars (stochastics between 10 and 100). But each signal will have its own number of bars.

Naturally, all these calculations should be done not only when saving data for training, but also when the NS works in real trading, because it will be trained for it.

I hope everyone has understood that the MO is such a delicate tool that a tiny error can play a HUGE role in the effectiveness of the TS.

I for example have been working in this area for 15 years and at least 3-4 months since I managed to get results worthy of attention and only a week ago I found inaccuracy in data preparation, after correction of which I got such a picture on the TOE. Above.... how in the future will behave TC as a whole is not known. The only thing that I now concluded that Elmnn-networks from Doc, showed the best results. Judging by the screenshot, the R-model overtook JPrediction by a head above and got far away from it. Let's see how they will work further......

 
Mihail Marchukajtes:

If there is a window in your basic strategy, and if it is also dynamic. Alternatively, you can measure the difference in volume between the beginning of the window and the signal. The same thing with the delta. If there is no window as such, I recommend to use Accumulation/Distribution and take its difference, but for how much data, I won't even admit.

In general, having information on the volume and delta of 11 symbols, I have managed to invent about 177 predictors. Stochastic based on real volume and delta shows itself very well. One or another stochastic is almost always involved in every model.... IMHO!!!!

Thanks for the reply!

In my basic strategy I don't use volumes at all, because they don't do anything on forex, I have forgotten about them for many years, but after I switched to exchange trading I decided to come back to ideas related to volumes again.

I have all sorts of windows. I was thinking to use the approach of accumulating data in a window, plus the MA MA on the window (I do not know yet which one, it needs to increase its window of averaging as the main window increases) and thus watch the delta between the MA and the increase, if the histogram will make abrupt changes, then somehow react to it. In other words, if the usual growth of 5%-10% is one situation, and if the bump at the last bar was more than 30% - it is another situation, and, respectively, take into account what was the window - whether the bump was large or not - that is another predictor.

I am interested in more variants with the volume, including the open interest - are there any ideas? If I visually observe the OM chart, I see some regularities, but they are hard to be tied to a numeric value - every day I should reset accounting and build approximate levels, at which the indicator will be set.

 

I wrote and wrote, then the forum glitched, so read from the picture, it's good that at least I made it...

 
Aleksey Vyazmikin:

Thanks for the reply!

I do not use volumes in basic strategy, because they do not bear anything on forex, so I even forgot about them for many years, but when I switched to stock trading I decided to come back to the ideas about volumes.

I have all sorts of windows. I was thinking to use the approach of accumulating data in a window, plus the MA MA on the window (I do not know yet which one, it needs to increase its window of averaging as the main window increases) and thus watch the delta between the MA and the increase, if the histogram will make abrupt changes, then somehow react to it. In other words, if the usual growth of 5%-10% is one situation, and if the bump at the last bar was more than 30% - it is another situation, and, respectively, take into account what was the window - whether the bump was large or not - that is another predictor.

I am interested in more variants with the volume, including the open interest - are there any ideas? Visually observing the OM chart, I see the patterns, but it's hard to tie them to a numeric value - every day I need to reset accounting and build approximate levels at which the indicator will be set.

Where do you get your OI data from????

I have an account in my opener and I see OI there, but it needs to be collected all the time, so I stopped FORTS for now. In general you just need to record OI and measure its change in the window. If the OI has fallen in the window and the volume has fallen, and the accumulative delta has increased, it can already tell us a lot. I think Delta+Volum+OpenInterest is the information that will be quite enough for the TS to make the right decision, but there is no OI on SME yet. Guys are making a glass, let's see if we can get OM out of there.....

 
Mihail Marchukajtes:

I wrote, then the forum glitched, so read from the picture, it's good that at least I made it...

Well, personally, I think there are more global patterns - that's what I'm looking for. I have no real data on the change in the nature of price movements from futures to futures - I need a specific metric. So far, you can see that with the reduction of the Central Bank rate the Si trend slowed, and then it reversed, which of course was expected... But it did so when many people were reluctant to expect it...


Mihail Marchukajtes:

Where do you get the data on OI????

I used this indicator for observation.

If memory serves right it is possible to pull out OI.

 
Aleksey Vyazmikin:

Well, personally, I think there are more global patterns - that's what I'm looking for. I have no real data on changes in the nature of price movements from futures to futures - I need some specific measuring indicator. So far, you can see that with the reduction of the Central Bank rate the Si trend slowed, and then it reversed, which of course was expected... But it has changed when many people gave up expecting it...


For observation I used this indicator.

It may be pulled out from Kvik (if memory serves me correctly).

I've made an Expert Advisor for MT5 that writes files with OI for 15 symbols in real time mode. And yes, I also chose Si because it has very good intraday moves. It has a very good intraday moving around. The OM is loaded into indicator and then all this kitchen is used in NS. For this reason I'm waiting for additional deposit in my account in opener and then I will use МТ5 in Si too. I'm sure that the use of OM will improve the already perfect system judging by the chart. Let's see, after the change of futures, if the work of TS remains at the proper level of stability, the transition to FORTS won't keep me waiting long.... Just need to save some OM for a couple of weeks and it will be possible to start.... Let's see....