Machine learning in trading: theory, models, practice and algo-trading - page 117
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
The market walks against its own statistics
Well, there is some truth in it, but only if by "statistics" we mean certain types of approximations figuratively speaking "obvious" types, because really the crowd must lose, but another question is to calculate how this crowd acts on average, what it uses as a reference point, there are infinite variations, you can find types of statistics with which the market in the future is positively correlated and easily find those that are negative. However, if you mainly mean the history of the current price series and its patterns, then yes, they are more likely to repeat the opposite than the same. But there are a lot of predictors, and the functions of the past prices are only a small part of them and not the most important ones. However if such a pattern is noticed, that the market moves against the naive statistics this knowledge may well become a new predictor! Let's invert the naive statistics and trade!)) It remains to formalize the types of naive statistics and check the statistical validity of this source of information.
3) And you try to refute ;) because what you say that my words are fiction is also fiction, but just your own, native ... do you agree?
Yes it is easy. I already have regression results posted on my blog on market time series based on "black box" learned patterns. On all 5 currency pairs on an out-of-sample period of 25 years, the regression gives a non-random prediction.
If the forecast were worse than the average value (zero), then I wouldn't even talk about building trading systems. And it is better, that's the point.
Then at least this picture: https://c.mql5.com/1/37/teaser2.JPG
The accuracy of predicting the sign of price increase. It is steadily above 50%, but falls a little short of the profit zone. But the dependencies are reproducible.
I have specific examples where the market follows the same pattern. I'm too lazy to lift the data and prepare it for you.) Try to feel it yourself. The phenomenon of reversion - have you heard? Take 10 years' worth of minutes, enough of the opening prices. Look at how neighboring price increases are expressed. You will see a reversion in all 10 years.
In short, stop making noise! )
1) Well, there is some truth in it, but only if by "statistics" we understand certain kinds of approximations figuratively speaking "obvious" their types....
2) But it's another question to calculate how this crowd acts on average, what it uses as a reference point, there are endless variations, you can find types of statistics with which the market is positively correlated in the future and easily find those that are negative. However, if you mainly have in mind only the history of the current price series and its patterns, then apparently yes, they are more likely to repeat the opposite than the same.
3) However, if such a pattern is noticed, that the market goes against the naive statistics this knowledge may well become a new predictor! Let's invert the naive statistics and trade!)) It remains to formalize the types of naive statistics and check the statistical validity of this source of information.
1) by statistics I meant all sorts of data values that a neural network will consider as regularities when training on this data
2) yes, while using the history of price series, the crowd is using this history ?
3) congratulations, you got it right the first time, but it's not so easy to make this predictor leading, and I don't know how, I have some ideas, but they are so dense that I don't even know how to register them
The advantage of this predictor will be that it will be stable in its readings over time - it won't have such an effect that you have trained the network and tomorrow the market will go against it.
In fact, this method resembles a critical form of network thinking, but in a very rough form of course
It looks like one net has learned the history and knows what will happen next, and then another net (critic) follows it to see if the forecasts given by the first net are really right, and draws conclusions and makes decisions instead of the first one
I recommend to start with target reversals by the type of a down-turn, up-turn, no reversal - those -1, 1 , 0
you may use any predictors, but not controversial - anything, but not indicators
If you want to experiment I will be glad to help you in any way I can
1) Accuracy of prediction of the price increase sign. It is steadily higher than 50%, but falls a little short of the profit zone. But the dependencies are reproducible.
2) I have specific examples where the market follows the same pattern. I'm too lazy to pull up the data and prepare it for you try to feel it yourself. The phenomenon of reversion - have you heard of it? Take 10 years of data, the opening price is enough. Look at how neighboring price gains are expressed. You'll see the reversion on all 10 years.
3) In short, stop making noise! )
1) I don't understand why these dependencies don't even pay off the commission, and can we even call them that... You're more likely to teach the model not to drain than to earn money.
2) No I haven't, I'm talking about more common things that everyone uses
3) When every 10 pages of the forum people write that they have trained the model, tested it on new data and everything seems fine, but the third sample or the real data dumped, and then they do the same thing again, and then they write the same thing again and it all repeats, repeats, repeats
wall - pea - bounce, wall - pea - bounce.... etc.
And I think really man does not even come to mind that it does not work because it is clear as day, but hell, he himself writes about it, maybe how else? And in general, why does not it work?
I just couldn't resist.....
i'm sorry i wanted to help you save several years of fruitless research ...
mytarmailS:
But the point is that your "theory" is not always confirmed by practice, but only when you change from trendiness to counter-trendiness and vice versa.
The blue chart will go against both small and large movements. If you look at a 200 candlestick chart you will go against the price, and if you look at 20 000 candlesticks the picture will be the same
Well, if the picture stays the same and stable, then congratulations!
You have found a "gold mine" (if it's not photoshopped of course).
There is only one thing left - monetizing figures.
...
predictors may be any, but not contradictory, anything but indicators.
...
Not inconsistent predictors or indicators that are anything but, what are they? Those will do:
Strike out what you don't need, write in what you do.
Not the only one. There are others confirmed by practice. And they are stationary... You have to look it up.
The last time this idea was expressed to me in such a discussion was by Matemat on 4.
It was given - tea, he is still looking for it, poor guy....
1) Well, if the picture does not change and is stable, then congratulations!
You have found a "gold mine" (if it is not photoshop of course).
Remains a mere trifle - to monetize figures.
2) Not inconsistent predictors or indicators that are anything, what are they? Those will do:
Delete the unnecessary, add the necessary.
1) Yuri, that's the point is that while it is not monetized, there is no leading properties of this blue line, it is not real to make money on it yet BUT there is an understanding of the process and it is not unimportant ... If you need it, I can tell you how I got it and you can try it yourself, I'm not sorry
2) For example, we have an indicator RSI, all the books write the indicator above 80 - sell, under 20 - buy
This is an example from life, everyone knows ...
an indicator above 80 - everyone looks, the price has fallen
on the second day, indicator above 80 - all look, the price has fallen
third day - indicator above 80 - all look, the price has fallen
on the fourth day - the indicator is over 80 - everybody sold, the indicator has been over 80 all day and the price went up, everyone was dead
so
Day five - the indicator is above 80 - what do you do?
I personally - delete the hell out of RSI
it may be 80 when price is falling and may be 80 when it is rising and we can't figure it out just like the neural network can't do it because what it learned in the past won't work in the future with this predictor...
And if we simply take even an ordinary price - say a series of 20 values, and the market is trending upwards - then the trend is up and there is no second one, everything is unambiguous and not contradictory, you know what I mean?
1) Yuri, yes that's the point is that while it is not monetized, there is no leading property in this blue line, earn on it is not real yet BUT there is an understanding of the process and it is not unimportant ... If you need it, I can tell you how I got it and you can try it yourself, I'm not sorry
2) For example, we have an indicator RSI, all the books write the indicator above 80 - sell, under 20 - buy
This is an example from life, everyone knows ...
an indicator above 80 - everyone looks, the price has fallen
on the second day, indicator above 80 - all eyes look, the price has fallen
third day - indicator above 80 - all look, the price has fallen
on the fourth day - the indicator is over 80 - everybody sold, the indicator has been over 80 all day and the price went up, everyone was dead
so
Day five - the indicator is above 80 - what do you do?
I personally - delete the hell out of RSI
it may be 80 when price is falling and may be 80 when it is rising and we can't figure it out just like the neural network can't do it because what it learned in the past won't work in the future with this predictor...
If we simply take an ordinary price - say, a series of 20 values, and the market is in an upward trend - then the trend is up and there is no second possibility, everything is unambiguous and not contradictory, you know what I mean?
2read: https://medium.com/@alexrachnog/neural-networks-for-algorithmic-trading-part-one-simple-time-series-forecasting-f992daa1045a#.qw3t34d4w
regression works best? from his conclusions