Machine learning in trading: theory, models, practice and algo-trading - page 2923
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
like the fact that 1 in 8 is white :-)
There are a lot of strange things in general - the model is being tested, but parabolic and AMA are added to the robot at the same time. There are strong suspicions that if you remove R and other stuff, the result will not change fundamentally - parabolic and ama in a pair will give similar results
Parabolic is not used, it is not used at all, it is its own, there were thoughts, tails in the parameters were left, it is not used in the algorithm.
AMA is used for stops, and R is used for inputs. I have not managed to make an Expert Advisor completely on AMA, though I tried it long ago.
The parabolic is not used, it is in general its own, there were thoughts, left tails in the parameters, it is not used in the algorithm
AMA is used for stops, and R is used for inputs. I have not managed to make an Expert Advisor entirely on AMA, though I tried it long ago.
In general, the problem is not in AMA, but in the fact that it incorrectly predicts strong market movements - over 30 pips in one hour.
Now I have looked at 10 strong movements incorrectly, there were more, but there was no signal on any of them. This is for two months.
The TS is built on predicting the next bar on R and then using this prediction in an Expert Advisor on MKL4.
Model on R.
It works on H1, the teacher is trending (ZZ), predicts the next bar. OutOfSampe is not used as the model is recalculated on each bar.
Efficiency of predicting the next bar 78-80%
Positive prediction performance on one of the next 8 bars is over 95%.
The model is excellent.
BUT.
It is impossible to build a workable TS on this model. We get about 78% of profitable trades, but they are small. But the losses are much larger.
In the Expert Advisor itself we cut losses and let profits grow. This leads to the fact that the number of profitable trades decreases with simultaneous growth of a profitable trade and reduction of a losing one. But it still does not solve the problem.
Here is one of the many results of the exercise with TR and SL.
If we look at the chart with trades, we can see that the erroneous predictions fall on strong market movements, i.e. the model for some reason erroneously predicts the directions of strong market movements. The reasons are not clear, as the model itself is trained on the sample obtained by Sample.
Such things prompted me to change my approach. IMHO, we need algorithms that have not numbers but probability distributions (or functions derived from them) as outputs. An example would be MO models from survival theory. Based on the distribution, input-output points are logically determined - for example, you can see the need to "trim the tail" of the distribution, etc.
The TS is built on predicting the next bar on R and then using this prediction in an Expert Advisor on MKL4.
2 months of testing is not enough. I've been testing on 5 years lately. I also had trending versions. So they sit in drawdown for up to 2 years. But then, came out in the plus, when the strong movements began their 2-3 in 5 years. In fact, they are rare white swans
They are not suitable for trading either, because after a month of drawdown I will run out of patience and switch off such an Expert Advisor. Besides, it is on M5 and the average profit from 1 trade is 0.00005.
Test it on 5 years. Maybe it will work better.
2 months of testing isn't long enough. I've been testing at 5 years lately. I also had trending versions. So they sit in drawdown for up to 2 years. But then, came out in the plus, when the strong movements began their 2-3 in 5 years. In fact, they are rare white swans
They are not suitable for trading either, because after a month of drawdown I will run out of patience and switch off such an Expert Advisor. Besides, it is on M5 and the average profit from 1 trade is 0.00005.
Test it on 5 years. Maybe it will work better.
No one needs a TS that has been in drawdown for 2 years. You are talking about testing, but real trading will be even worse.
This kind of thing prompted me to change my approach. IMHO, we need algorithms that have not numbers but probability distributions (or functions derived from them) as outputs. An example would be MO models from survival theory. Based on the distribution, input-output points are logically determined - for example, you can see the need to "trim the tail" of the distribution, etc.
I mentioned the 8 bar. This is some variant of your approach. I plotted the prediction probabilities for each of the 8 bars. The probability gradually dropped. For the 8th bar, it was about 40%.
He never made a video on stochastics and Navier-Stokes, too bad
Well, maybe he will - who knows.
I was surprised by another thing - the function is very similar to the "Head and Shoulders" figure - and got curious:
1. Is it possible to find a formula-constant for any other patterns.
2. How to evaluate the plausibility of the bar formation and the formula.
The TS is built on predicting the next bar on R and then using this prediction in an Expert Advisor on MKL4.
Model on R.
It works on H1, the teacher is trending (ZZ), predicts the next bar. OutOfSampe is not used as the model is recalculated on each bar.
So does it predict the outcome of the bar or the ZZ vector?
If the latter, then you are actually estimating the probability of a trend change at the next bar, and it is obvious that the default probability of continuation is higher, hence the problems with strong movements - the ZZ vector change point is quickly reached. The vector change point is known to the model, the average price movement is also known (the same ATR).... Knowing these two parameters, what will be the accuracy of prediction?
Maybe it is more valuable to classify rare events - will there be a strong movement on the next bar, or in general for the day?
Anyway, it's working out fine.
but you have to understand the long-term trend. and it's always a 50/50 guessing game.
Looks good, if it's a full automatic.
Maybe it's more valuable to categorise rare events - will there be a strong move on the next bar, or in the day in general?
That's news trading - I don't do that.