Market etiquette or good manners in a minefield - page 62
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
gpwr, you make a good point, of course. Well, firstly, let people teach the behemoths to fly properly. This will come in very handy when the area of application of the neural net is still clearly and distinctly defined.
An analogy is appropriate here: Coulomb (or who else), who researched the law of attraction of electric charges, also sort of taught hippos to fly. After all, at that time no one had a clear idea of the field of application of electricity. But they invented the generator - and that law came in handy.
By the way, for the time I sit in front of my computer studying the nets and having one eye on the market - I've noticeably improved my manual trading. I start to "feel" the market without any market indices. And I have nothing at my disposal, except a personal neuronet. Of course, it's premature to talk about flying... but there's already a run-up with hops.
Very clever words! To trade successfully you need to know the psychology of most market participants who continue to trade manually and use trading channels, i.e. trade in the direction of the channel until it breaks through. Once it is broken through, they close their positions and wait for confirmation of a new trend, after that they open new positions. Sometimes it is just interesting to watch the price moving away from the channel's border, as if some invisible wizard hands were pulling it out of there. Why is it so? Because millions of other traders have built this channel and are betting that the price will bounce. Their bets are precisely what moves the price back inside the channel. If most traders used neural networks in their trading, it would work too. But alas...
By the way, for the time I have been sitting in front of my computer, studying grids and looking at the market with one eye, I have become much more successful in manual trading. I am beginning to "feel" the market without any market indices. And I have nothing at my disposal, except a personal neuronet. Of course, it's premature to talk about flying... but there's already a run-up with hops.
I guess it's more emotion than a real sense of the market using the neural net. Try to be more careful, hopefully the fate of the hero depicted on your avatar should warn you of a fundamentally possible danger. Well, I won't even remind you that created astrolabes on perceptrons, especially single-layer ones, are typical adaptive filters, about 101% incapable of stable operation (in the sense of "target function"). Also I won't remind (it's usually in any textbook), that if you don't have significant correlation in series (I think you use x(n)-x(n-1), and it's not there), you have very small probability of getting what you want. Beware of the illusion that NS gives. But in any case I wish you and your teacher good luck :o)
PS: NS is not a miracle at all, if there is no "statistical phenomenon", it will not work.
Very clever words! To trade successfully you need to know the psychology of most market participants who continue to trade manually and use trading channels, i.e. trade in the direction of the channel until it breaks through. Once it is broken through, they close their positions and wait for confirmation of a new trend, after that they open new positions. Sometimes it is just interesting to watch the price moving away from the channel's border, as if some invisible wizard hands were pulling it out of there. Why is it so? Because millions of other traders have built this channel and are betting that the price will bounce. Their bets are just moving the price back inside the channel...
For some reason I am becoming increasingly dubious about the possibility of predicting price at a particular point in time. First, because the market is not discrete - there are highs and lows to the price, between which it jumps as it pleases. At 12:00 Moscow time it will be somewhere between its maximum and minimum, but what position it will be in relative to the open - who knows... And this determines the "colour of the candle". So, the colour of the candlestick can be quite random (except for the strong trend with small "pullbacks").
So, it makes sense to "find" the market movements that will move beyond the current minimum (maximum). Or really work "on the pullback". And determination of where the price will go in the next period of time is not very informative.
If most traders used neural networks in their trading, it would work too. But, alas...
There seems to be no consensus here either. Some say the more people use the system, the worse it works. Others seem to say the opposite :)
For some reason, I am becoming increasingly dubious about the possibility of predicting price at a particular point in time. Firstly, because the market is not discrete - there is a high and a low for the price, between which it jumps as it pleases. At 12:00 Moscow time it will be somewhere between its maximum and minimum, but what position it will be in relative to the open - who knows... And this determines the "colour of the candle". So, the colour of the candlestick can be quite random (except for the strong trend with small "pullbacks").
So, it makes sense to "find" the market movements that will move beyond the current minimum (maximum). Or really work "on the pullback". And determination of where the price will go in the next period of time is not very informative...
thanks, next))
after some time your doubts will change or fall and you will have clear ideas.
the market has a lot of states and intersections of these states...everyone sees the market differently
the most patient ones win
I honestly don't understand all this misery. There are so many already written C++ networking codes with ORO on the web. Here, for example, is a simple code with a good description of the theory. There is one bug in calculating MSE (read last post on this page)
http://www.codeproject.com/KB/recipes/BP.aspx
I spent only one day to understand the code and it worked reliably for me in forex (not in the sense of profit, but in the sense of learning the network on currency pairs).
I think you spend a lot of time adapting matcad to neural networks, and then you are going to drag and drop its formulas into C++ and debug the code there. Not efficient, gentlemen. You can spend years like this before you run something that works in forex.
I'm inquisitive by nature. I remember when I was a child, when I was given a toy, it was usually taken apart in the evening. Not broken, not smashed, but disassembled with my dad's tools (which he was angry about). So here too - I'm just curious to know what's inside that Black Box under the pretty "Neural Net" wrapper. Also, as you correctly point out, a third-party off-the-shelf product is generic in nature, it is possible that it contains bugs and is not adapted to the specific task at hand. Given that the predictability of market-type VR is extremely low, the NS needs to be retrained at every countdown. I think it is impossible to satisfy this requirement using a commercial "heavy" package. The entire code of a two-layer universal mesh written in MQL can fit in 35 lines! I think it is worth it.
In general, the task of trading, in my opinion, boils down to solving three major problems:
1. Rejection of the "efficient market" hypothesis.
2. Understanding that of the three ways to make money:
a) insider information,
b) analysis of macroeconomic news,
c) analysis of historical instrument price data,
the last two are available to us. Of them, the analysis of macroeconomic news does not imply the use of MTS (at least I do not see a way to actually implement it). This leaves us with the last point.
...A network using data from the same time series would be an autoregression; linear if you have one layer or non-linear if you have more than one layer with non-linear activation of neurons. Training such an autoregressive network is nothing else but an approximation of the series by a non-linear function. That is the difference of such description from the fitting of polynomial or Fourier series is very small. The network's prediction of future values is nothing more than an extrapolation of the fitted non-linear function into the future. The only advantage of a neural network is the universal ability to approximate any non-linear function. The question to ask here is: why do you think that the current price is a non-linear function of past prices?
I agree with you there, indeed the network is similar to a non-linear AR model, with one exception (as you correctly noted) - it does not require a priori knowledge of the process model, and that is its most important advantage given the non-stationarity of market BPs. Of course, we mean their quasi-stationarity (according to the first point), otherwise no NS can work. Thus, I don't care whether or not the condition is met: "the current price is a non-linear function of past prices", in general the non-linear NS will allocate the right area. Hence, there are no alternatives for NS!
As for the choice of a specific NS (multilayer perseptron, Cohenen maps, etc.), I think there is no choice here - multilayer perseptron - because of the non-stationarity of the initial BP (identification of significant patterns requires a lot of history).
Just because you were able to fit a non-linear function to past prices doesn't mean you found a market model. Therefore an autoregressive network will not give you any trading advantage.
This goes back to the issue of quasi-stationarity. That's why I pre-train NS at every BP count, and I don't use bars as BPs because they are virtually unpredictable due to very weak links between them. If this rule of thumb is not followed then the connection between the real and the expected results will be red, and the same rules apply for all other EAs.
...Why is that? Because there are millions of other traders besides you who have built this channel and are betting that price will reflect. Their bets are precisely what moves the price back inside the channel. If most traders used neural networks in their trading, it would work too. But alas...
Thanks, I'm aware that I'm not the only one trading here. However, about whose bets are moving the market - I'm beginning to doubt it's "bets by millions". It's more likely to be "unit bets". If most traders used neural networks in trading, I'd be learning how to build autoregressive channels, polynomials, etc.
Apparently you don't understand another basic aspect of the market (or any other game): most always lose. Learn on your own. The time has not yet come for you to learn and give advice.
Thus, I don't care if the condition is met or not: "the current price is a non-linear function of past prices", in general the non-linear NS will allocate the right area. Hence, there are no alternatives for NS!
This goes back to the issue of quasi-stationarity. That's why I pre-train NS at each BP count, but I don't use bars as BPs because they are virtually unpredictable due to very weak links between them. The drawings in this topic are for hour bars only as an illustration of some particular properties of NS and their features and are not used for real trading.
I do not understand the phrase about an alternative to NS. Why have you come to such a conclusion? And why the condition of determinism in the market is not important for you? And why is there no connection between the bars? What data do you take for the real trading if not the history from the bars?